In August and September 2026, we sat down with 30 recruiting and engineering leaders in New York and San Francisco to learn how AI is changing hiring. To everyone who spent time with us: thank you. This is the report we promised. For context, we interviewed recruiting teams (sourcers to heads of talent) and engineering teams (tech leads to CTOs) at: 1. a16z 2. Anthropic 3. Bain Capital 4. Clay 5. Cognition 6. Dragonfly 7. Flint 8. GitHub 9. Google 10. HackerOne 11. Imbue 12. Index Ventures 13. Nash 14. Natural 15. Neural Concept 16. Notion 17. Rackhouse VC 18. Ramp 19. Sierra 20. Stripe 21. Structify 22. Vanta 23. Warp We also occasionally interviewed the agencies supporting these teams, such as Sundog Talent, Hiring Cubed, Jack & Jill, and more. Here are the report sections together with some of the highlights: - **Inbound** - Several of the most sought-after teams have given up on inbound, yet it closes better. - In remote roles, fraud makes up 25% to 80% of applications: buy a fraud detection tool, they work. - No team trusts AI to rank applicants; one outsources reading applications to contractors while keeping sourcing in-house. - **Sourcing** - Agentic sourcing outputs noise after the first 2-3 days unless someone keeps reviewing and steering it. - Some see success in dropping AI and "funnels of doom" to cultivate targeted relationships instead. - Splitting "candidate fitness" from "likelihood of looking for a job" makes it easier to get value from exotic datasets. - **Referrals** - Asking precise questions face to face works best for most. - Semi-automated processes seem even better: export connections, enrich and filter with AI, then discuss. - One scale-up's founders refuse to pay referral bonuses, to kill vicious incentives. - **Evaluating** - Cheap tests like Leetcode are dying. - New formats that work are system design, building with AI, reviewing AI-written code, using AI to debug a system the candidate didn't build, and pure product interviews. - All formats that work tend to consume much more resources and require heavier calibration. - **Winning** - Teams are outdoing each other in finding creative ways to sway candidates. - Getting a budget of founders' or VPs' time to take first calls with top candidates works. - Candidates' anxiety peaks with the offer; the best teams plan for it, from dinner with the candidate's partner to giving their best offer upfront. Start with Evaluating if you're a hiring manager, and Inbound or Sourcing if you're a recruiter. If you asked us a question during our conversation, you'll find it answered in [Bonus: interviewee's questions](#bonus-interviewees-questions). We hope you find this useful! > About the authors: we're [Greg](https://gkk.dev/) and [Corentin](https://corentin.trebaol.com/), two engineers turned founders. Greg was an early AI tech lead at Stripe. Corentin co-founded GitStart (YC S19), which hired 400+ engineers through its platform. > > This trip started while we were building evals for an AI prototype of our own and realized we had no idea how we'd interview anyone for that skill. We set out with three guesses: that teams would hire AI engineers as a distinct role, that they would have special treatment for hiring AI-native people, and that AI would automate the repetitive work of recruiting. We grade them in the [Conclusion](#conclusion). # I. Inbound > "I remember one video. This guy was in a completely pitch black room and you can only see his eyes. I mean, I don't even know what that is." - Head of Engineering, early-stage startup, on the video answers they started requiring with applications Several of the most sought-after teams we met have given up on inbound. Yet all four teams that process inbound and ran the numbers found that inbound candidates close at a higher rate than sourced ones. Inbound got much harder to run. Giving up on it is the popular answer, but not always the right one. Inbound can be thought of as a machine that converts brand into hires: post a job description, receive applications, filter them. It works as long as flow is a decent proxy for interest and the filter is manageable. AI broke both at once. Inbound is flooded with low-effort mass applications from bots and, in some cases, outright fraud, so the ratio of people worth talking to versus the overall number of applications plummeted. At the same time, resumes are more polished and role-tailored, so the noise is harder to separate from high-quality applications. Noise used to be manageable and weeded out at later stages like onsites, but at today's volume that is simply too expensive. A consulting recruiter who works across many startups told us that "companies no longer will even bother reviewing applicants because the volume is so high that they're not able to effectively filter.” ## The Flood ### Fraud is Everywhere, and in Remote Roles it Can Be a Quarter of the Pile or More Many of the people we talked to mentioned fraud. Three teams, all hiring for primarily remote roles, gave us numbers: | **Who** | **Share of applications** | **How they know** | **Remote roles?** | | --- | --- | --- | --- | | Recruiter working with Series A/B startups | "upwards of like 80-ish percent" | Marked for fraud by the Applicant Tracking System | Primarily remote | | Engineering recruiter, AI developer tools | "30 to 40 percent" | Flagged by a dedicated fraud tool | Primarily remote | | Recruiting lead, logistics startup | "at least maybe 25%" | Own estimate | Yes | One caveat on the highest number: fraud detection sometimes wrongly flags auto-apply tools that create a new email address for each application, so the 80% likely includes some of those. We don't have numbers from teams hiring on-site, so we can't say how much lower fraud is there. We define fraud as heavily fabricating work history and/or impersonating someone else. While many fraudsters are people who are desperate to get a job (or who want to stack multiple jobs at the same time!), some cases are much scarier. One team suspected a sophisticated North Korean hacker group of trying to land a remote engineering position. This kind of fraud is extremely difficult to catch without an onsite interview. And likely devastating if it succeeds. That's also why remote roles are the natural target for the professional end of the fraud spectrum: nobody ever has to show up in person. How teams spot it is in Making inbound work, below. ### The Best-looking Resumes Are Often Fake People are tired of consistently selecting fake resumes: - “Every time I find a resume I like, the person is fake.” - Talent Partner - "A lot of those fraud candidates are, you know, obviously writing their resume to meet our criteria." - Recruiting lead, early-stage startup - "I only want to be notified if the score [in my ATS] is like 95, right? But… I had situations where they passed the test and I got on the phone and the person was not… [who the resume said]." - Head of Engineering, early-stage startup - "The resumes look great. If they were real candidates, I would love to speak with these people, but… it's just noise." - Engineering recruiter, growth-stage company The perverse result: a resume that's too put-together became an anti-signal. ### Homemade Defenses Get Beaten or Backfire Some teams built their own defenses, and they either got beaten or backfired. The Head of Engineering quoted at the top went as far as requiring video answers with applications, and got back "Different faces… clearly AI generated, people seemingly in front of green screens doing weird stuff." Others added friction by removing the "Apply" button and listing an email instead; some removed the email altogether. These tactics may stop the cheapest bots, but not professional fraudsters, and they hit ideal candidates too. ### People Who Use Fraud Detection Software Love it Purpose-built fraud tools are another matter. One recruiter trusts his fraud detection software so much that "if it does flag a resume as fake, I just won't speak to the candidate." Another team says Ashby's fake detection "actually does work." Tofu is "the best product in the market for fraudulent applicants," as one recruiting agency's founder put it. ### Even with Fraud Handled, no Team Found a Good way to Prioritize the Rest Ranking the rest is a different story. Scoring tools save some time, but nobody we met trusts them to pick, because a score mostly measures how well a resume was tailored to the job, and fraudsters and auto-apply tools tailor best. One recruiter was nostalgic about the days when you could spend some time calibrating a tool like Eightfold to the specific role, auto-run the whole inbound pile through it and trust the surfaced 5% to be really high quality and worth working with: "Nothing exists like that anymore.” ### Some Resort to Externalizing Candidate Filtering The best-resourced answer to prioritization we saw is to pay people to read. One growth-stage AI company still makes about one in five of its hires from inbound, and its full-time recruiters don't read a single application: "It's not a good use of my full-time team to be reviewing the inbound applications. I have contractors doing that." Contractors and AI-assisted review work the pile, and a mandatory in-person onsite keeps fraud out. ## When Inbound Works ### Inbound Needs Some Brand Awareness, but Mainstream Fame Floods it Inbound needs enough brand awareness to bring in volume. A VC talent partner advises companies to focus on sourcing at the seed stage, inbound at Series A, and referrals past Series B, because before Series A there usually aren't enough applicants for inbound to make sense. Too much attention breaks it again. A different growth-stage AI company told us inbound worked well while it was relatively unknown and only people in the know applied. Once its funding announcements hit mainstream publications, it landed on mass-apply radars. Inbound now accounts for 1-2% of its hires, and tens of thousands of applications sit unreviewed. When that moment comes, you probably don't yet have the machinery to deal with the flood and the noise. ### Seniors or Specialists Rarely Apply, unless You Build a Special Attraction for Them "We obviously can't get any security engineers from an inbound, it's like incredibly rare," a Head of Engineering told us. Senior hires through inbound happen "kind of surprisingly." Inbound is "good enough for kind of generic back end product engineers," and the more specialized the role, the less you can depend on it. That's why he recently pushed his team to do a lot more outbound. But we met one company that built a door for specialists and found inbound working well. It's a security company, and its take-home, built on its own open-source datasets, doubled as a public challenge, similar in spirit to Stripe's famous Capture the Flag challenges from the early 2010s. People would get attracted to an interesting technical challenge, wonder "who is behind this", learn about the company and apply. Most of those who came through it were data and ML engineers with an interest in security, exactly the specialists a security company needs. Few of its regular inbound applicants turned into hires because of the noise, but the applicants who came through the challenge were good enough to make running inbound worth it. The company hires remotely, so the challenge reaches far. ### Filtering Inbound is a Pain, but It’s Easier to close > "If someone applies inbound and they're actually a good candidate, it's gonna be an easy close if they can make it all the way through." - Talent Partner > We met both ends of the spectrum: recruiters who marked inbound processing as dead, and hiring managers who said inbound closes best. One executive hiring manager went further: every hiring manager, not just recruiters, should look at inbound daily, or at least weekly. What gives? We think people are talking about two different aspects of inbound. Inbound is positively selected for job-switching intent and interest in the company. People who want to leave their current position and are interested in your company will apply. This is one thing outbound doesn't give you: your sourced Stanford CS person may be a perfect fit and not looking. That's why inbound closes better. Bots and fraud buried these people under noise, but among genuine applicants, nothing AI did changed this. However, inbound is negatively selected for quality, and it happens through two mechanisms. One: since applying went free with bots, the pool is dominated by mass applies, and mass applying is disproportionately what you do when you aren't getting traction elsewhere, or when plain fraud is the goal. Two: the strongest candidates rarely need to apply anywhere, because recruiters reach them first. > Note > > Nerdy aside: economists call this adverse selection, and it won a Nobel. In 1970, George Akerlof described the "market for lemons." Used car buyers can't tell good cars from bad ones, so they'll only pay an average price. Owners of good cars won't sell at that price and leave, the average car on the market gets worse, buyers lower their offers again, and the market unravels. > > Inbound is going through the same spiral. Recruiters can't tell a strong resume from a polished fake, so they stop reading inbound. Akerlof's model predicts the next step: strong candidates learn that applying is a black hole and stop applying, the pile gets worse, and reading it gets even less worth it. Akerlof shared the 2001 prize with Michael Spence, whose work, as we'll see at the end, is about how to break exactly this kind of spiral. ## Making Inbound Work ### Treat Job Descriptions as Products We've talked to companies that had a unique engineering culture, but for the life of us we couldn't tell from reading their job descriptions! This made us notice that no team had mentioned anything about investing in writing better job descriptions. We think it's an easy win for attracting the people you want, though it does nothing about the flood on its own. Most job descriptions are abysmally bad because they are written for hiring teams instead of for candidates. A laundry list of corporate speak like "Strong technical problem-solving skills, especially in fast-changing, ambiguous environments" doesn't add any value for candidates. In our experience, job descriptions are a product, and good products are easier to sell. Good examples of this approach are lines like: - You're the no. 1 priority hire because xyz. - You will get budget and headcount to scale successful experiments. - Here are the profiles and videos of the people you’ll be working with. A sharper job description also gives tailoring bots more to tailor to; that's what the filters below are for. ### Post where the Right People Look Not all sources are equal. Some niche job boards still work for people we talked to. They mentioned the YC job board working fairly well, and the same for VCs' job boards and curated talent pools, even if these tend to be low volume. ### Question upon Application: Ask for a Link, not an Paragraph (Untested iDea) Almost nobody we met asks qualification questions on the application. The one team we heard from that tried got exactly what you'd fear: "fake responses, fake everything." But that question was generic, and a generic question is just more text for a model to write when bots are applying. The other reason teams don't try is fear of friction: every extra step loses candidates. Yet the same teams are drowning. "I am drowning," one recruiter told us, with thousands of applications in review, while another observed that "a lot of recruiting teams aren't even checking applications anymore." If you're drowning in applications, losing a few isn't your problem. Not reading the rest is. Friction also works: one startup put a coding platform at the very first touch and lost 40% of applicants, but of the top performers it selected, 75% passed the recruiter screen, against the usual 30%. So add good friction. There's one idea that we would like to see tested in the wild: instead of a generic screening question, ask candidates for just a link to a favorite thing they read about e.g. AI in the last 6 months. Our suspicion is that LLMs would keep posting links from the same small pool, which would be easy to catch. On the other hand, genuine answers would make excellent conversation starters. This catches bots and low-effort mass applies, not professional fraud: a human fraudster can paste a real link. That's what the fraud tool and the onsite are for. > Note: This question worked beautifully for screening people attending an in-person AI event we’ve organized in the past. Interestingly, we successfully applied AI to evaluating answers to this question so this can be largely automated (Goodhart’s law still applies, though). ### Add Escape Hatches to Catch Non-traditional Candidates One team's rejection email included a paragraph on a take-home programming challenge that, if solved, guaranteed an interview. The challenge was cleverly selected to be difficult, in the leetcode sense, and beyond what frontier models could solve at the time. The idea was: maybe someone in Iowa has a non-traditional background but is an excellent programmer with very desirable foundational skills and doesn't have a recognizable school or employer name. This optional take-home got them some of their best non-standard hires. ### Check What's Hard to Fake The tell can be small. One recruiter got on a call with a candidate whose resume listed blue-chip names, including the very company he was interviewing with, and ten years in New York. "Where in Manhattan did you work?" "I don't know." When asked to accept a LinkedIn connection request, the candidate hung up. The practical ways people deal with fraud all come down to checking what's hard to fake: - being suspicious of LinkedIn accounts without a profile picture or with no content - fraudsters love to impersonate those profiles, - probing the claimed background with questions anyone who actually lived it could answer: - if someone claims to call in from e.g. Manhattan, asking which neighborhood specifically, - if they claim to have graduated from a certain university, asking for their favorite restaurant near campus. Aside, and not a fraud case: we also met someone who had hired the infamous Soham, who went viral on X in mid-2025 for secretly holding several jobs at the same time. The interesting bit: he was in fact an excellent engineer with some life goals calibration issues. ### Let a Fraud Tool and a Human Screen Do the Triage Buy a dedicated fraud tool and trust it: the teams that did skip calls with flagged candidates instead of second-guessing them. Keep a human recruiter screen after it, and if you still worry about fraudulent candidates getting to offer stage, require an in-person onsite. Reading the inbound pile doesn't have to be your recruiters' job either: one company we mentioned before hands it to contractors with AI-assisted review. And have hiring managers, not just recruiters, look at a sample of inbound every week. They're the ones who know what good looks like and how to calibrate recruiters and their tooling. ## Our Take: Rescue Inbound with Filters that Cost the Wrong Candidates More In recruiting and job seeking, Michael Spence's work on signaling and information asymmetry, which earned him a share of the 2001 Nobel Prize in economics alongside Akerlof, is a particularly helpful lens. The setup is pretty simple: candidates have hidden qualities like skills or experience that they want to convey to employers. The challenge for employers is to design a process that captures that signal reliably. Spence found a practical framing for that process: differential or asymmetric pricing. It must be harder for a low-quality candidate to fake the signal than for a high-quality one to emit it. If you ask candidates on the first call how they're using AI in their lives beyond chatbots like ChatGPT, you're asking a Spence-shaped question: candidates who actually use AI creatively or uniquely will be delighted to talk about their inventions, and those who are not AI-pilled will struggle, especially once you ask a follow-up or two. In an application form, the same question is worthless: a model writes a plausible answer in seconds. Other examples of Spence-shaped tactics are: - hard take-homes that frontier models can't solve, like the escape hatch - a strong programmer already has the skills and enjoys the problem, while a model can't fake the answer - a link to an AI article, paper, blog post or project a candidate liked and recommends - they've already done the reading, so pointing to it is easy, while we suspect bots would keep recommending the same few links - specialist challenges like the security company's public challenge or Stripe's Capture the Flag, described above - fun and horizon-expanding for a specialist, out of reach for everyone else This is also why a general set of leetcode-style problems is dead in interviews, yet a carefully selected subset works as an escape hatch: the same kind of problem is Spence-shaped only when a model can't solve it. The tactics that are not Spence-shaped: - removing the email address for job applications - it may stop the cheapest bots, but not a professional fraudster, and a good candidate pays the same price as a bad one - generic screening questions or tests that prompted AI can solve  *We suspect that there are many more asymmetric signals to be discovered in the “The target” box.* A well-built challenge or recruiting step aligned with Spence's theory should feel like a gift; a badly designed one feels like a hoop. In the AI era, asking about something candidates have already done (a project with AI, something good they read about AI) or something that exploits their hard-won skills (a hard technical challenge) is a good signal shape. It also coincides with the most important selling principle: it's always about the other person, not you and your filters. --- # II. Sourcing AKA Cold Outbound Virtually every company mentioned they got better results the more they focused on sourcing. This was even true for world-famous companies who are drowning in candidates and referrals. At the same time, increased competition is making sourcing harder for everyone. One company saw their response rate decrease by 2/3, even as they went from virtually unknown to becoming a famous startup. On the candidate side, we're getting more and more anecdotes of top engineers shutting down both their LinkedIn and emails. Most talent teams are reacting by scaling their sourcing efforts with AI. A few are taking the opposite approach of narrowing the funnel and doing things that don’t scale. Regardless of the approach, many are looking for alpha in assembling proprietary datasets. ## Scaling with Agentic Sourcing > “Agentic sourcing is accretive. It doesn’t replace human sourcing, but it’s impossible to go back.” - Head of Talent > To make the buzzword concrete, “agentic sourcing” refers to creating AI bots that navigate talent databases, score profiles, and often automatically send personalized messages. The most common tools were: - Juicebox: very commonly used across companies of all sizes. - Metaview: sometimes used in parallel. - Ringo: less commonly used. A recruiter mentioned it “has great semantics but is limited besides that”. - Jooba: an agentic sourcing solution with human oversight operating as a hiring agency. It was mentioned only once but by a happy Head of Engineering client who said: “the Slack approach works really well me, and they found some great people for us.” Here’s what we learned about how to get the most out of them: ### Beware of the Honeymoon Phase > “Past day 2 or 3 of getting about 10 profiles a day, you start to get slop.” - Senior Recruiter > As of September 2026, agents excel at finding great candidates for the first few days. Afterwards, quality tends to drop quickly. This is because most are currently built to be seamless black boxes, making it impossible for humans to tune their gears: > “An interesting problem with all the sourcing tools right now is that they surface candidates but don't give you any sense of how large the market is. With Boolean search on LinkedIn, you can gradually narrow down the list to the 100 best possible people in the world, and then you know that if you remove one criterion, you would have this many more. When you use agentic sourcing tools, you have no idea. All you see is the quality going down, but you have no control over what signals to add, what credit to add, what to toggle, and in which order. […] So you get slop.” - Head of Talent > In other words, the biggest mistake you can make with agents is to stop reviewing their work after you're satisfied with the first few batches. ### Agents Are Often More Creative than Humans > “AI is very creative. For example it’s really good at inferring skills from titles and surfacing candidates who have completely different backgrounds yet who are a good fit, and who I would never have found on my own!” - Hiring Agency Technical Recruiter > This makes agentic sourcing ideal for roles that don’t have a single ideal candidate profile, such as Chief of Staff or Forward Deployed Engineer. It also means that, even if you want to source manually, using agents early in the process can help generate more ideas. ### Use Dedicated Professional Tools and Challenge Them Often Apart from those who chose the extremely targeted approach, everyone was bullish on agentic sourcing except two groups: - Those using generic tools like ChatGPT and Claude to research candidates on the web and with LinkedIn Recruiter through computer use , - And those who said they tried everything on the market at least 6 months ago, but nothing really worked. This suggests you’re more likely to find success with dedicated professional tools, and to refresh your opinions often, since new generations of models upgrade these tools’ capabilities. The only exception here was one team that's on the frontier of AI and was successfully dogfooding with their own agents, but they still used off-the-shelf professional tools on top of those. ### Go Manual for Complex Senior Roles > “Sourcing with agents is good enough for average candidates but not star candidates, it’s nowhere close to as good as human level [sourcing]!” - Recruiting Agency’s Founder > Everyone we met who hires for very senior roles agreed with this quote. So did all the hiring agencies we met, although our data was likely biased as we only met agencies hiring for senior roles at the famous scale-ups. ## Focusing on Building Human Relations with Fewer People > I like to Zig when everyone Zags. […] Everyone is all in on scaling sourcing with AI right now, there’s so much alpha in going super targeted. - Head of Talent > More people are criticizing the "funnel of doom" approach, which is believing that gradually selecting the best people at every step of the funnel leads to making great hires. But what great candidate would ever reply to an generic email? No amount of "we only hire the top 1% of applicants" will fix a bad candidate pool, leading people to narrow their focus and bet on cultivating long-term relationships. ### Intensely Narrow the Focus A lot of people are now using executive search methods on non-executive roles. We've even seen someone limit their search for a “standard” product designer to 20 people only! Instead of boiling the ocean, they simply listed the world’s best products resembling their own and hunted their makers. But narrowing the scope creates two challenges: 1. People who work in the companies building the world's best product in their space are in high demand 2. Identifying the products that demonstrate the most craftsmanship in any space requires taste, and is especially difficult for technical products when talent teams are not technical An elegant way to fix these two is to set up a Slack channel where engineers share projects demonstrating elite craftsmanship. It obviously solves the latter, but also the former, because most of the projects people share will be personal side projects, which inevitably surface elite candidates who are not at top companies. ### Cultivate Relationships > “There’s way too much urgency in the market. People are rushing through processes and losing candidates when it’s much more efficient to start slow, like by asking for advice on how to build a similar thing at your company.” - Head of Talent at a VC firm > The flood of AI VC money is creating so much urgency that a lot of hiring is becoming very transactional. The resulting mercenarism is alienating not only for hiring teams but also for candidates who are therefore more receptive to human touches. Here are approaches that worked for people we met: - Establish the first contact - Have the CTO compliment a technical blog post and reach out for advice - Use snail mail and attach presents like coffee beans or chocolate - Gradually grow the relationship - “We try to be very intentional about always offering great lunch options so people enjoy coming over to have lunch with our team.” - Head of Talent - Organize game nights and hackathons. - Build ATS extensions to one-click invite people for above events. - Show you care during the process - Build an ATS extension to automatically create an expense card and a gym membership for candidates who fly in for onsite interviews. ## Investing in Exotic Datasets No matter the approach they choose, teams face the same challenge of having to compete with every other company for the same set of profiles, given both agentic sourcing solutions and manual sourcing tend to rely on the exact same dataset: LinkedIn. ### A Note on LinkedIn Data During this September's first dinner of the a16z Talent Engineering cohort, this question heavily divided the audience: “Will LinkedIn still be the main platform for hiring in 5 years?” While virtually everyone agreed LinkedIn fell behind by getting complacent, more than half believe LinkedIn’s network effect induced lead was too big for others to catch up. It’s by far the largest and most up to date talent database, and their ability to track user data gives them an unfair edge in understanding who is starting to look for a job. Now, while all hiring agencies we met swear by LinkedIn, a growing number of teams are looking elsewhere. We even met a newly hired Head of Talent who is experimenting with not setting up LinkedIn Recruiter! ### Example of Exotic Datasets Out of the people we talked to, the most commonly dataset rebuilt internally was that of published AI researchers. Since every paper has an author, talent engineering teams are able to list every published researcher and thus find hundreds of people who are not on LinkedIn. Some enrich this further by scouring YouTube for technical conference talks. Another common one was to download Hacker News comments and de-anonymize their authors. ### Merging Datasets into an Actionable Tool - Case Study One team was particularly impressive in how they packaged multiple datasources into an actionable tool. First, they had the obvious yet underused insight to split their candidate scoring system into two numbers: - Candidate quality / fitness for the role - Likelihood of the candidate leaving their job Most talent teams use their intuition for the latter, but having a dedicated score enables tools to do work that’s too heavy for humans, such as mapping people's managers and increase the score when managers leave the org, and scan SEC Form 4 filings for insiders selling company stock. Second, this team surfaced this data with a Tinder-style app that would showing candidates to the teams. Multiple iterations surfaced two learnings: - While they started by showing profiles to talent teams, they found more success with hiring managers instead, since those have more finesse in reviewing candidates and can typically customize outreach messages better. - Adoption is an issue even if the data is great, so they encourage daily usage by surfacing 10 new profiles a day, and by having these profiles disappear and go to the next one in line if they were not processed. This internal tool, built with AI, felt more advanced than most sourcing tools on the market, which prophesies an exciting future for domain experts who enjoy building with AI. ## What Holds whatever the Approach ### For AI Related Roles, Most False Positives Comes from Over-indexing on the Company and Under-indexing on the Team In demand skills, such as building a harness or iterating over evals, tend to be limited to very specific teams, which means there are legions of engineers working at top AI companies who have never got to grow them. In other words: > “It’s not enough to look at companies. Some [AI startup name] folks are great but many really bad at AI. […] It really depends on the team.” - Recruiting Agency’s Founder > ### Invest in Creating a Role Taxonomy A lot of companies call the same role differently, and this is especially true for frontier AI roles like AI engineer, harness engineer, etc. This is interestingly less of a hurdle for teams that are heavily leveraging agentic sourcing, since those titles tend to be closely related in the models’ vector space. But as humans get involved, a common language is required to communicate efficiently. An interesting example of taxonomy is that of Superlink (recently acquired by Anthropic) — it standardizes titles across the AI supply chain, linking together title, role definition, archetype profiles, and a sourcing list.  [superlink.ai/talent-map](https://www.superlink.ai/talent-map) ### Especially for Newer Roles like FDE and AI, Consider Targeting Candidates Who Aren’t in Demand and Training Them Many talent teams don't even consider this approach because training falls outside of their remit. But consciously compromising on some skills can allows for much easier sourcing, provided hiring managers agree to step up to fill the gaps. Note that implementing training is getting easier as a more companies are selling boot camps, certifications, and peer communities, which can also serve as inspiration and replicated internally. For example, we met a Founder for whom hiring FDEs was impossible until they realized that AI had made most of their technical requirements obsolete, such as building integrations with manufacturing tools, which could be taught to anyone with basic AI training. This allowed them to expand the candidate pool and raise the hiring bar for product skills and business sense. # III. Referrals AKA Warm Outbound > “If less than a third of hires come from referrals, something is wrong.” - Senior Technical Recruiter at several early stage companies > Virtually every team we met was pushing hard to get more referrals. This is an interesting reversal from a couple years ago when a growing share of companies were intentionally shutting down referrals in an effort to increase diversity, some going as far as shutting down their “amazing and very effective” custom built referral software. Still, teams differed in when to prioritize referrals: some use them to speed-run the early team growth: > “For us, referrals accounted for 75% of hires until we reached 75 people!” - Head of Talent > While others focus on referrals once the teams’ networks have grown: > “Referrals work best once you hit 100 people, before that I tend to advise people to focus on outbound at the seed stage and inbound at the series A stage”. - Head of Talent at a VC firm > ## Getting More Referrals If not everyone agrees on when to focus on getting referrals, every single person we met, including the most AI-pilled ones, agreed that heavily involving recruiters in the referral process worked best. Here are learnings from how teams do it: ### Full Face to Face Most early stage teams’ experiments reverted to the simplest possible process of having recruiters find face to face time with new hires to discuss referrals. Some advice at this stage includes: - The classic “ask for precise questions”: “who are the best engineers you know” doesn’t work nearly as well as a series of more precise questions such as “who was the best product engineer who was also good at design when you were at [companyName]?” - Wait until the new hire has shadowed a couple hiring processes, which helps with calibration. - Open LinkedIn together — one person mentioned this made the process a bit more awkward for some but yielded significantly better results ### Dream Shopping List One of the big labs’ experiments led to the “dream shopping list”, which explicitly expands its scope to anyone who could be a fit, regardless of whether the referrer has worked with them before. This yields otherwise unlikely suggestions such as authors of arcane technical blogs. Other features of this format are to ask for 10 names, which tends to be more than what most teams request, and to ask for all these names during the onboarding, since the urgency to hire is so strong. ### The Best (Semi-automated) pRocess Here’s the best process we encountered: 1. Whenever someone joins the company, a LinkedIn network data export is requested. 2. The exported CSV is dumped into a tool that automatically enriches all the profiles with databases like Exa. 3. Agents analyze the enriched data to flag what profiles could be a good fit. 4. Top profiles are sent to the referrer for them to confirm who are the strongest. 5. Recruiters receive the output shortlist, prioritize it based on their needs, and schedule face to face time to gather more quality information. This process is unique in that it saves time not only for recruiters but also for new hires. ## Converting Referrals Once teams know who they want an introduction to, the next challenge is to make that introduction successful. ### Get the Referrer to Reach out Two teams’ independent A/B tests unsurprisingly confirmed that getting the referrer to reach out yields much better results. ### Bet it All on the First touch Some of the best teams we met mentioned recruiters and referrers often sit down to write the first intro messages together, because: > “The first contact is by far the most important. If you let people ignore you once, it’ll be much harder to get to them later.” - Head of Talent > To stand out, some even experiment with snail mail bundled with coffee beans or other delicacies. ### Don’t Lead with Hiring Referrals convert better when referrers get to reignite the relationships first, whether it is by asking for advice on something they’re building, or by inviting the target to an event like a game night. Events like the later are a good place for talent teams to add value: new hires swamped with getting up to speed are unlikely to take on the burden of organizing any event, opening the door for talent teams to help, and even amortize the organizational burden across several candidates. ## Unusual Approaches ### Removing Referral Fees One scale-up’s founders was vocal about maintaining a no referral bonus policy, citing the vicious incentive bonuses create for employees to coach their protégés, which directly attacks the meritocracy of the process. ### Postponing Referrals until 6 Months in Another team was adamant about postponing the referrals until hires had been in the company for 6 months, citing two reasons: - When asked too early, the overwhelming majority of referred candidates were not a fit because they couldn’t relate to the mission. One explanation is the company develops niche B2B systems that most people don’t naturally relate with. - Referrals are much better once a hire has had the time to calibrate, defined in this case as having shadowed at least one entire hiring process end to end. # IV. Evaluating > Our coding interview is totally a scam. It doesn't give me any signal. It's completely ignored. I feel like we have this coding interview just because [if] we don't have it, it's a bad signal [to candidates]." - Head Of Engineering, growth-stage company > The cheap engineering tests stopped working. A model solves a coding puzzle, and a short take-home comes back mostly the same from every candidate. The teams furthest along in adapting to the AI era replaced them with something bespoke, conversational and elaborate: watch candidates build with AI, hand them a system they didn't write, and make them defend every decision out loud. The best version we saw takes five to six hours of a candidate's day and at least two engineers per candidate. It’s expensive to run, but other teams are copying it anyway. ## The Cheap Tests Stopped Working ### Leetcode is Dead, just not Everywhere yet Of the 7 teams whose interview loops we documented in detail: - 2 have dropped the classic coding round, - 1 keeps it while ignoring the result, - 2 still run it while trying to get rid of it, and - 1 never had one. The only team that still trusts its coding test lets candidates use an LLM during it. The engineering leader quoted above checked his own data and found zero correlation between coding scores and offers. He keeps the round so candidates don't think the company is unserious. Another team tried a softer, less leetcode-y version first, then ripped it out entirely. Leetcode was always a proxy. Nobody's job was to invert a binary tree on a whiteboard; the puzzle stood in for a bundle of skills (breaking a problem down, fluency in code, correctness under pressure) that correlated well enough with the job, and it was cheap to standardize. Even before AI, one hiring lead watched PhD researchers and math competition winners fail these interviews. AI broke them in two ways. In remote interviews, a model solves the puzzle, so you're grading whatever runs on the candidate's second screen. In person, the puzzle measures a skill that drifted further from the job once AI started writing most of the code. Leetcode survives on inertia. A recruiter at a company that still bans AI in every round told us: "I think we're stuck in the past." William Gibson's line fits: the future is already here, it's just not evenly distributed. We also heard, informally and outside our interviews, that the big AI labs still run leetcode-style rounds. Our guess, not a finding: labs hire for exceptional computer science fundamentals, and their people build the evals that teach models to solve exactly these problems, so for them leetcode may still be the least-bad proxy. Also, leetcode-style problems can still work when they’re too hard for models to solve, which is why they were used for the escape hatch described in the Inbound section. ### High-context Take-homes, Defended Live, Still Work Teams disagreed about take-homes because they had different ones in mind. An early-stage startup dropped its two-hour take-home: everyone prompted their way to roughly the same result and there wasn’t really any signal to differentiate candidates on. Another team rejected a one-to-two-day project as too much to ask. But an engineering executive at a growth-stage company called hers the highest-signal step in the loop. Candidates worked with real open-source datasets, then presented the project live, and she watched for critical thinking and how well they presented. The difference is material. A two-hour generic task has nothing in it the model doesn't already have. Real, messy datasets have no single right answer, and the live presentation makes the candidate own each choice. The take-home is an onramp to the conversation which is the real test. ## What You Test Now: Judgment and Taste What teams want to measure now is what the model can't supply, and the same two words kept coming up in our conversations: judgment and product taste. ### Judgment: Steering AI through Work it Can't One-shot Two engineering leaders told us their most AI-pilled hires disappointed. One: "they were just producing code and not analyzing it… the thing that kept coming out to me in conversations was good judgment." The other found her most AI-pilled engineers were the weakest at modeling processes, and what they built was chaotic. She hired strong process and system thinkers instead and leveled them up on AI. It's easier to teach AI skills (it's a green field, after all) than software engineering or the discipline of thinking through a workflow. Asking about AI is a good filter for curiosity. The best question we heard: "Outside of work, how do you personally use AI?" "ChatGPT" is a bad answer; a side project or a whimsical automation, say a betting agent for fantasy football, is a good one. Watch how fast they answer: quick, detailed answers mean they've already done the thinking, and they separate people who build with these tools from people who only use them. But it's still a story about a hobby project. It tells you the candidate builds things with AI, not whether their judgment holds up on messy, real work. The cleanest test of judgment we saw is code review: hand candidates code AI wrote and see whether they catch the things that look right but aren't. An early-stage startup uses a large, slightly sloppy AI-written pull request with about fifteen issues and a few hidden footguns. It's their highest-signal round, and it shows whether someone has been burned before by maintaining their own mistakes. A growth-stage company is considering a screen built on deliberately AI-sloppy code with AI switched off, on the argument that code review is most of what engineers do today, and prompting sits on top of it. Another team offers review as one of two options in its AI-assisted interview. What nobody has yet is a rubric: "When do you trust it? When do you not trust it? But how do you create a rubric for that?" And nobody we met tests "AI engineering" as its own discipline. The company behind our reference loop cares that candidates use AI in their development process and teaches the model-specific parts after hiring. Our read: judgment only shows when the work is too big or too messy for the model to one-shot. That's why the formats that work are long builds, AI-written code with problems planted in it, and bugs the agents can't find. ### Taste: Product Sense is now an Engineering Skill A growth-stage company ends its loop with a product interview that has nothing technical in it, and a founder sits in on every one. Its recruiter: "they crush our technical rounds, just murder them. And then they get to the product round and they fall flat." That's after he spends 20 to 30 minutes prepping each candidate for it. They talk to about 68 engineers per hire. You can't read taste off a resume. The same recruiter started out looking for engineers who had shipped something to real users, then met excellent product thinkers who never had. Another engineering leader, whose loop has no product round, thinks he needs one. An early-stage founder told us someone ultra sharp technically who can't talk to a customer is no use to them. Taste also leaks out of the interview. A recruiter told us about a PM candidate who aced the onsite, then sent the hiring manager a thank-you note "so clearly written with ChatGPT." When building gets cheap, deciding what to build becomes the expensive part. Engineers who spent years implementing someone else's specs may never have had to flex it. ## The Modern Loop, step by step We got to the most complete answer through a referral: a VC talent partner told us his team copies one company's evaluation process end to end. It's built for a company that can afford it: roughly a hundred engineers, in office, a strong brand and five to ten engineering hires a month. We use it as the reference and contrast each step with what other teams do. | Step | What happens | Time | Pass rate | | --- | --- | --- | --- | | Recruiter screen | What did you own and ship end to end? | | about 40% (was about 60%) | | System design | Virtual, "very traditional" | | 40 to 50% | | Onsite: build with AI | 30 min scoping, 2 h solo build with any AI tools, 30 min review | 3 h | a quarter to a third pass the whole onsite | | Onsite: debugging | A sandbox codebase, AI allowed | 45 min | | | Onsite: behavioral | With the hiring manager | 45 min | | | References | | | most get offers; some don't | ### Screen for Ownership, not Titles When the company rebuilt its loop, it recalibrated the recruiter screen first. Recruiters stopped screening on companies and titles and started asking what the candidate owned, what they shipped end to end, and whether they actually led the project. The pass rate fell from about 60% to about 40%, after what the recruiting lead called "a very painful process": recruiters want their candidates to go through, especially the ones they sourced. Simple pattern matching on keywords is what break here. Having built with RAG isn't the same as having shipped a product to production and struggle with improving AI products quality. One recruiter we met skips technical questions and just has candidates walk her through what they've built. ### Gate on System Design System design replaced the coding round as the first technical step, and it's the only technical round done remotely. When AI writes the code, the engineer's job shifts toward keeping a clear view of the system, so that's the skill to test. And because it's a conversation, a memorized answer falls apart a couple of follow-ups in, even though the questions leak to Reddit, Glassdoor and Blind. Failing it means no onsite, because the onsite is expensive. The recruiting lead isn't comfortable with that: "I'm torn by this. I know we're passing on some candidates that are not doing hugely well in the systems design that we end up missing out on." So they make exceptions for candidates with an exceptional background or a back-channel vouching for them. That's the reverse of the escape hatch in Inbound. There, a hard test rescues people without the pedigree; here, the pedigree rescues people from a hard test, and the candidates nobody vouches for stay rejected. Another engineering leader who leans on the same round names the same weakness: it catches true positives, but he can't see the false negatives. ### Watch Them Build with AI This is the centerpiece. The candidate brings one or two ideas they'd be excited to build from zero to one. Two interviewers use the scoping half hour to spec it up or down, so nobody walks in with something prebuilt. Then the candidate builds alone with any AI tools, and the review covers trade-offs, product decisions and what they'd do with five more hours. The house rule: "you're the author of the code, but you can use the coding agents to be the ghostwriter." Other teams run cheaper versions. An early-stage startup pairs the candidate with one of its engineers on a mocked-up problem from its own domain and asks them to bring their own AI IDE; the session replaced its coding test. A growth-stage company is considering five or six hours alone in a room with its product. Another team is trying to squeeze this whole format into 60 minutes. The candidate's own idea is material the model doesn't have, and the scoping makes sure it wasn't built in advance. As with take-homes, the signal is in the review. ### Debug a System They Didn't Build The build tests zero to one; the 45-minute debugging round tests one to n. Candidates get a sandbox codebase made for the interview, AI allowed, and they're graded on how they get to the bug that AI can’t find right away. The recruiting lead calls it "not perfect." Others are heading the same way. One growth-stage company wants candidates to navigate, debug and extend existing code, and floated debugging real incidents that include e.g. logs. Another engineering leader loved a format he'd heard about: an old cut of a company's own codebase with "a pretty tricky bug that at least right now the coding agents can't solve directly," every coding tool installed, in person. Note the "at least right now." Stripe ran a debugging round when one of us worked there: a short bug report and an open-source library to fix it in. It was one of the most fun and highest-signal interviews in the loop. Today a frontier model would likely fix many of those bugs on its own. What keeps a debugging round useful is how much it looks like real debugging: a system misbehaving in ways that are hard to pin down, tests that don't cover it, and logs to dig through on a hunch. Engineers role in such scenario is to use AI as high leverage, high efficiency tool and point it at right directions as rapidly as possible. ### Finish with Behavior and References The onsite ends with a 45-minute behavioral interview with the hiring manager, then references, which occasionally block an offer. Other teams lean on this end of the loop harder. One runs 90-minute career deep dives backed by references and back-channels. Another has candidates name their past managers, calls them, and asks who the best people they ever managed in that role were; about 80% of candidates pass. A head of engineering at an early-stage startup skips technical rounds altogether. He asks for the hardest bug the candidate ever debugged and trusts his "spidey senses," then runs a paid work trial, a month by default, where he can see whether someone overuses or underuses AI. ### The Same Shape at an Early-stage Startup An early-stage startup runs the same shape without a bench of interviewers: a 30-minute intro call that goes three levels deep on past work (about 30% pass), a 45-minute screen combining system design with the sloppy pull request review (about 50%), and an onsite implementing something real in its own product, on a shared laptop with coding tools and its own Claude subscription (about 50%). They also look at the candidate's prompt history. The material is their own product, so nothing had to be built for the interview. ## What it Costs to Run ### It's Subjective, so Calibrate Your Interviewers The recruiting lead, on his own loop: "I think that the biggest fallacy of this interview is that it's highly subjective." With no answer key, consistency has to come from three places: - **Who interviews.** Only primary interviewers lead the build; secondaries observe, and new interviewers shadow until they're promoted. Only about 20% of the company's engineers will ever qualify as primaries. - **What they ask.** A shared question bank with a rubric is what makes two interviews comparable. One engineering leader saw it go wrong at a previous employer: interviewers didn't trust the bank, so each improvised, and candidates in the same loop got interviews that couldn't be compared. - **How they write it up.** After a subjective interview, the interviewer's scorecard is the only evidence the debrief sees. A free-text box produces write-ups nobody else can rely on; a structured scorecard with around a hundred questions doesn't get filled in. Keep it structured, with fewer questions than you think. One upside: formats graded on process need less maintenance. The reference company doesn't have to keep refreshing its questions, because there's no answer to leak, only a thought process to watch. On a fun note, an interviewee shared that, at a scale-up, “We we did a bunch of analysis on what was the highest signal. […] The highest signal was just what the CEO thought.” ### It Puts AI in the Room, so Decide the Logistics up front > If you're not gonna let me use my tools, I don't want to interview with you. - a candidate interviewing at one of the companies we talked to > Of the six recorded teams with a technical round, two ban AI, two allow it in some rounds and two expect it, and every change we heard about points toward allowing it. One Series B company went from banning AI in interviews to requiring it in about a year. Most teams haven't settled the practical questions: - **Whose laptop?** The reference company offers both. The early-stage startup above keeps a dedicated machine with its own subscription. - **Whose subscription?** A candidate's only AI subscription may come through their current employer, and nobody wants them interviewing on it, or wants to ask them to pay for one. - **How much AI, and when?** Some teams allow a coding platform's built-in agent only in designated parts of the interview. - **What do you look at afterwards?** The prompt history is free signal. ### It's New, so Check it against Who Succeeded The reference loop shipped less than a hundred days before we met, so nobody has performance data on it yet. Teams that checked their old loops found surprises. One growth-stage company compared interview scores with who left or underperformed within about a year. Candidates strong on the behavioral interview and weak on one technical round did better than technical stars who did badly on the behavioral. An engineering leader ran the same analysis at a previous employer and found the single best predictor was the CEO's opinion. When the CEO expected a hire to fail and the team pushed it through anyway, the hire failed, even when the two never worked together. His theory: "a company is just a reflection of the founder." A VC talent partner sits in on engineering performance calibration to ask whether hiring missed something. If you're smaller, have AI compute pass-through rates for each stage of your funnel, then reorder the stages or raise the bar where the numbers say to. ## Our Take: a Test Measures a Person only if it Contains Something the Model Can't Supply Every format that held up had something in it the model couldn't supply. We see three ways to get there: - **Real material.** A codebase, a dataset or a product the model has never seen, or the candidate's own idea. - **Observed judgment.** Watching decisions get made and making the candidate defend them live. The artifact is a byproduct; the conversation is the test. - **Problems beyond the model.** A bug the agents can't find, a puzzle they can't solve. A test with none of these measures the model, and you could have tested that yourself for $20 a month. This is Spence again, from the Inbound section: a good test is cheap for the right candidate to pass and hard for anyone else to fake. A strong engineer enjoys defending their own design. Someone who prompted their way through and delegated most understanding to AI, can't. Two things follow. First, cost caps how close to the job you can get. Moving from a test to a conversation to watching someone build to the job itself costs more engineer hours at every step, which is why the full loop belongs to companies at later stages. Small teams can still get most of the way by using their own product as the material instead of building a playground. Second, model progress moves the line, as the Stripe debugging round shows. Tests built on "the model can't" have a shelf life. Tests built on "the model doesn't know this" (your codebase, the candidate's idea) last longer, and tests built on conversation last longest. And measure the person with the tool. A loop that bans AI is measuring someone you won't actually be employing. # V. Winning AKA Closing > “How do others win candidates over the competition… and over the big labs?!” - *Almost everyone* > Making offers has turned into a bidding war for a home: people invest weeks projecting themselves with a candidate only for the person to get snatched away by an unknown highest bidder at the last minute. According to a Head of Talent who coached hundreds of early stage founders, the most common hurdle to winning candidates is that teams start selling way too late in the process, so here are the learnings we gleaned starting from the top of the funnel. ## Top of Funnel - Getting Candidates’ Interest ### Prioritize Inbound All four teams that ran the numbers reported that inbound candidates closed at a significantly higher rate than sourced ones. Get an AI agent to confirm these numbers for your specific situation and see whether inbound’s increase in win rate is worth its associated additional work. ### Catch People before They’re Searching for Jobs, such as by Organizing Events > We thought spending $5k on our fundraising party was way too much… but it brought us a candidate we hired 6 months later. […] It was the cheapest hire we made all year! - Head of Talent > As organizing events is getting fashionable, a lot of people try it only to be disappointed with the lack of direct results and stop everything. Teams for whom this channel works mention expectations should be to see no results for at least 6 months, but that the social flywheel can bring very cost effective hires given one can organize several great events with the agency hiring fee of a single candidate. ## Middle of Funnel - Interviewing Stage ### Involve Core People at the Very Start of the Process Having core people like CTOs or Heads of Engineering lead screening calls for top candidates is getting more common. Recruiters typically get a budget that varies up and down based on feedback. When we asked a head of engineering whether that was working well, he answered that it was worth it: 2 of the best hires they made were candidates he had been the first person to talk to. ### Integrate Work Trial Candidates into Daily Meetings Teams found that it increased the win rates because it allowed candidates to better project themselves into the team they would be working with. One team also mentioned it was especially important for big tech profiles, since those are typically worried about the craftsmanship bar being too low in earlier-stage startups, and that seeing future colleagues working as opposed to getting tested by them helped alleviate that concern. ### Create Content to Help Balance Interviewing and Selling A Head of Talent at a VC fund mentioned his first advice to founders and recruiters was to create a one-pager about the company that could be shared with candidates ahead of interviews. The reason is that finding balance between selling and interviewing is very difficult. Sending content allows interviewers to interview better because - it allows them to spend less time selling and therefore more time interviewing - it allows them to sell better by asking what got the candidate’s attention and personalizing the exchange accordingly What is less common, but in our view probably even more impactful, are videos of the future colleagues of that person introducing themselves and saying why they would love to work with that candidate. ### Gather More Data so You Can Go above and beyond Later A fun example of this principle is a company asking for candidates’ favorite food as early as the application form, which allows for personalized food orders during onsite interviews. The inspiration came from the book “Unreasonable Hospitality” which relates the story of Eleven Madison Park’s rise to fame as they became increasingly obsessed about and creative with customer experience.  ### Hone the Office Visit Experience During our trip, we offered to treat everyone at a nice restaurant, yet most people preferred inviting us to their office. For most people, it was probably out of convenience, but for some of the best recruiters, it is because they are always selling! Funnily enough, one head of talent was pretty open about the fact they wanted us to experience visiting their office because they put a lot of care into the whole experience, and because we may be looking for a job in the future. Another took the opportunity to ask who were some of the best engineers we knew in a particular field we worked in. When it comes to the office experience, some finish the visit with a stop by the cafe, where the “Head of Coffee makes amazing brews”. Others mention they are very intentional about offering outstanding lunch options for candidates, and have the office manager welcome candidates and write the candidate's name on the lunchbox. ## Bottom of Funnel - Offer Stage ### Shape the Process around Candidates’ Anxiety Peaking with the Offer For recruiters, making an offer is typically fun, but for candidates, receiving one typically induces anxiety, because making a life-altering decision is stressful, and so is negotiating salary and benefits for many. Very few teams take this into account. Some of the best recruiting teams we met were very intentional about this, with strategies like: - maximizing face-to-face time, such as by inviting the candidate and their partner to dinner - removing the negotiation pressure, such as by mentioning they give their best offer upfront, or that they use a company-wide salary grid, which means they are open to raising the person's salary, but that will would entail raising every other person on that same level. ### Overwhelm Candidates with Love Bombs Surprisingly few teams use the full extent of the “love bombing” arsenal. Listing it would be too long, so here are highlights on who to involve: - **Investors and role models** - one team's standard process is to ask who the candidates’ role models are during interviews, which allows them to hunt those people for a recommendation to the candidate to join the company, whether it’s through their network, their investor’s network, or through platforms like cameo.com - **Founders** - founders’ involvement is a great heuristic for how much companies care about talent. Virtually every single one of the most talent-dense teams we met had founders invest significant face to face time with candidates, even as their companies scaled in the thousands of employees. - **Power laws apply here** - one of the world’s most famous founders insists on spending face to face time with candidates, but focuses disproportionally more energy on the first hires for a given role or field of expertise. The idea is that a stellar hire will make hiring other top profiles significantly easier, because people will want to join them, and because hiring them grants access to their network. - **Anti-selling works** - Interestingly, we got two testimonies from engineers who got swayed by founders mentioning all the reasons why they should not join the company. Intellectual honesty works! - **Peers who get it** - since many people identify with their work, another way to touch their heart is to compliment it, or simply to show genuine interest and the ability to thoughtfully discuss it. - **Referrer -** finally, back-channeling with whoever made the referral is one of the easiest ways to better understand which of the above might work best, yet not every company is actively doing it. Teams are also getting more creative with how those people are expressing their love for candidates, with things like: - Custom Loom videos from the team - Invitations for dinner to the candidate and their partner - Making personalized gifts, such as adding tickets to a baseball game with the candidate’s favorite team to the signing bonus - We even heard of a team personalizing job offers to themes like David Goggins and Broadway musicals! ### Offer Personalized Benefits We were surprised to not hear anyone mention personalized benefits. In our experience, candidates are so used to being offered the same things, like gym memberships, that they don’t notice them anymore. This make personalized benefits an easy way to stand out. For example, if a sales candidate's dream is to get better at programming, offering an hour a week of pairing with a senior engineer from the team doesn't cost much, yet it has a big emotional impact on the candidate. Another good example is offering an hour a week of house cleaning to a sleep-deprived young parent. # Conclusion We arrived with three guesses at the start of our AI hiring field trip, and got most of them wrong. We thought AI engineering — building the evals and harnesses inside AI products — would be a distinct role teams hire for. Instead, building with AI turned out to be a skill every engineer is expected to have, while building AI products remains real work that almost nobody hires for as a separate discipline. We thought teams would have special treatment for hiring AI-native people. We found none: AI shows up as one step in existing loops, and two leaders told us their most AI-enthusiastic hires disappointed. And we thought AI would automate the repetitive work of recruiting, which was half right: it is being tried wherever the work is search, and rejected wherever the work is judgment about a person. That last distinction turned out to be the small version of a larger pattern. Across thirty conversations, the story was not that AI automated hiring. It was that AI made it nearly free to produce the raw material of hiring — applications, resumes, outreach, candidate lists, plausible work samples — while leaving the cost of deciding whether any of it is real, good, or worth acting on exactly where it was. The bottleneck in hiring has moved from finding people to trusting the evidence about them. That one move explains most of what surprised us. It is why the channels built on trust — referrals, genuine inbound interest, people someone on the team already knows — regained value just as the channels built on volume lost it. It is why interviews stopped grading what a candidate produced and started watching how they decide. And it is why judgment appears on both sides of the table: it is the thing recruiting teams will not delegate to AI, and the thing they now most want to hire, for the same underlying reason. AI didn't make hiring cheaper. It made cheap hiring worthless: most signals that still carry information cost more than the ones they replaced. We think that is a phase, not a resting place. Expensive verification — longer onsites, more human reading, more founder time — is what you fall back on when the cheap proxies break. The best of today's formats already cost the wrong candidate more than the right one; what they don't do is stay cheap to run. The advantage will go to whoever builds the next proxy: a signal that is cheap for the right candidate, expensive for the wrong one, and cheap for the company. The escape hatch from Inbound is an early example: a problem too hard for models, sent with every rejection email, costs a strong programmer an afternoon and the company almost nothing until someone solves it. A few teams are experimenting in that direction; none has made it work beyond a niche yet. Two cautions before our final bets. First, this is a snapshot of practice, loosely tied to outcomes: measurement across the teams we met is weak, several of the processes described are months old, and almost nobody has yet checked whether their new signals actually predict who succeeds. Second, our sample leans one way: well-funded startups in San Francisco and New York, and agencies that hire senior people for famous scale-ups. Treat what precedes this page as a map of what good practitioners are doing, not proof that it works or everything should be copied. Our read on AI-native hiring: in hindsight, history predicted what we found. AI is a new source of leverage, and being AI-native means using it well. Software has always had 10x variance within the same job, and hiring absorbed it through a higher bar and leveling: there's no 10x software engineer process. Using AI well bundles two things: fluency with the technology, which is becoming a baseline like being internet-native, and the leverage on top, which is the old 10x spread in judgment. Our bet: as models improve, the gap will widen, but along judgment. Hiring will test judgment harder and price the spread into levels, rather than invent an AI-native track. AI engineering is a different question, because a job splits off only when it demands a bundle of skills that existing roles don't cover. AI engineering may be one, but today the skill can't be defined well enough yet to interview for, and the people who have it are too few to build a hiring track around. The open question is whether it becomes a role or dissolves into every role. The DevOps path: the work has real depth, and recruiting is a lagging indicator — you cannot build a hiring track for a job until enough people have held it. The prompt-engineer path: a title appears, then folds back into software engineering as teams teach the skill after hiring. Today's evidence leans toward the second; the two of us disagree on which way it breaks. The sign to watch is whether a well-known company publishes an AI-engineering interview loop that others copy. If, a few years from now, teams are still hiring on the proxy of "shipped an AI product," the role will have dissolved. If the loop exists, it will have arrived. We set out to learn how AI is changing hiring. We did not expect the answer to be that it is sending everyone back into the room with each other — but that is where the evidence kept pointing. # Bonus: Interviewee's Questions At the end of every interview, we asked people what questions they would like us to research on their behalf. While most are already answered in the report, we thought it would be useful to bring them together in a Q&A format. So here are, as a bonus, the 20 most common questions, alongside AI-generated summaries drawn from call recordings and our manual notes. These questions are organized in five categories: - AI-era interview loops - Product thinking and AI product engineers - Screen accuracy - Inbound, screening and fraud - Sourcing and engagement ## AI-era Interview Loops ### 1. What Have Companies Put in place of the Traditional No-tools Loop of Coding, Algorithms and System Design? - One company has replaced the coding screen with system design. The onsite is a planning session, a two-hour build of the candidate's own idea with any AI tools, and a review of the trade-offs. A debugging round in an existing codebase is in pilot. - An early-stage company moved a pairing session, in the candidate's own AI editor, to the front of the loop and dropped the separate coding test. - Another early-stage company runs system design and code review over video, then an onsite feature build on a shared laptop with AI tools. - One company ends with a paid two-day trial on a real project, in the office when the role is onsite. ### 2. How Many Teams Still Run Technical Interviews with no AI at All, and how Are the Rest Bringing AI In? - In a small sample, about five teams allow AI somewhere in the technical loop. That includes an online coding test at one company, and only some of the coding rounds at another. - About three teams still run technical interviews with no AI. A former employee of a large company said that company was in this group when they left. - A recruiter described a Series B client that went from banning AI to requiring it within a year. - The sample is too small to treat as a market share. ### 3. Should You Add a Separate AI-skills Interview, or Convert the Existing Rounds to Allow AI? - One company allows AI in the technical work rather than adding a separate interview. - At one company, only the team that builds models is asked extra machine-learning questions. For everyone else, what matters is that AI is already part of how they work. - Another growth-stage company plans to retrain interviewers and switch every coding round at once, so it never compares an AI round with a non-AI round. - An engineering executive offered a caution from hiring for this. The people most aggressive about using AI produced work that had to be redone. Hires who learned AI more slowly, and who were already strong at thinking a process through, did better. AI skill should not outrank engineering judgment. ### 4. If Candidates no Longer Write Code by Hand, what Core Skill Should the Interview Test? - Most answers point to reviewing and debugging existing code. One engineering leader called that the main technical skill now. - One company is piloting a debugging round in an unfamiliar codebase, with AI allowed. Another team got more signal from debugging a past incident than from a coding exercise. - An early-stage company gets its best signal from a review of a large, sloppy, AI-written pull request with bugs buried in it. - An engineering recruiter at a growth-stage company disagrees. He worries that the ability to read code will fade if nobody is ever tested without tools. ### 5. What Have Teams Tried that Failed? - One team tried a softer version of the old coding questions. Interviewers and candidates both disliked it. - A head of engineering at a growth-stage company compared coding-interview results with offers and found no relationship. He is also unhappy with the AI coding pad the team uses. - An early-stage company dropped a two-hour take-home. Candidates prompted an answer and did not read it. - Another early-stage company asked for video answers on applications and got fakes. A growth-stage company has tried AI interview platforms and has not found them useful. ### 6. How long Should an AI-allowed Interview Run, and in what Environment? - One published version is a planning conversation, about two hours of building, and a review. - A growth-stage company is considering a build that takes most of a working day. - Secondhand: a session of roughly two hours, on a laptop with the usual coding tools installed, built around an old bug that current agents cannot solve. - Nobody described a version that fits in an hour. ### 7. In Practice, how Do You Let Candidates Use AI: Their Own Tools and Subscriptions, or an Environment You Provide? - Where a company has published its onsite, candidates use the AI tools they want and are told in advance how they will be judged. - One company lets candidates bring their own laptop or borrow one. - One early-stage company has candidates use their own AI editor. Another hands over a fresh laptop with the company's own model subscription and reads the prompt history afterward. - A growth-stage company's vendor pad ships with bundled models, and they are not the ones its engineers want. Engineers at another growth-stage company, testing interview platforms, kept asking for their own terminal. - Not answered: API keys, paying for candidates' subscriptions, fairness between candidates. ### 8. When Candidates Can Use Any Tool, how Do You Tell Their Work from the AI's? - Most teams have stopped trying. One engineer said there is no reliable way to tell. - One team grades the reasoning and the trade-offs, not the finished output. The candidate is supposed to be the author, and the agent drafts. - An early-stage company reads the prompt history. Another uses a paid trial of about a month, which shows whether someone leans on AI too much or too little. ### 9. How Do You Get Reliable Signal when Many of Your Interviewers Are Inexperienced? - One company keeps a fixed list of interviewers who lead every round. Everyone else only observes. They think only a minority of engineers will ever be ready to lead. - One company forbids interview feedback written by AI. - A growth-stage company is starting to compare interview transcripts with its hiring bar. An early-stage company calibrates interviewers every month. ## Product Thinking and AI Product Engineers ### 10. What Does a Product-intuition Interview for Engineers Look Like, and Who Should Run It? - At one growth-stage company the founder runs the last round, and wants to hand it off. Candidates bring problems they see in agent-driven development and walk through the user's journey. None of it is technical. - One company does not have a separate round. Two engineers judge product decisions inside the build review. - At an early-stage company the last step is an in-person sales role-play, with the founders playing the client. - A head of engineering, speaking about a previous employer, said the chief executive's judgment was the highest signal in the loop. That argues for keeping a founder in the round. ### 11. How Do You Evaluate Engineers Who Build AI Products (Evals, nOn-deterministic UX, pRototyping)? - One company mostly does not test for it. Only the team that builds models gets extra machine-learning questions. Everyone else learns how models behave, and how to evaluate them, on the job. - A recruiter described a seed-stage client that turned down a research-lab machine-learning candidate who had not shipped AI products to production. - In a previous role, one of the people who asked this had a candidate describe an applied-AI system they had shipped, then probed its scale and its quality. Beyond that, the question is open. ## Screen Accuracy ### 12. What Are System-design Screens Missing in False Negatives? - One recruiting lead said the design screen turns away people the company would have wanted. A strong outside signal can waive it. A standard resume cannot, because the later rounds take a lot of interviewer time. - One company lets a rejected candidate earn an interview by solving a very hard take-home. - A former employee of a larger company saw candidates drop out sharply, by as much as about two in five, after an automated coding screen was added. - Nobody measures the misses. ### 13. Which First-call Questions Separate Good Candidates from Great Ones? - A head of engineering at an early-stage company asks one question: the hardest thing the candidate has ever debugged. - One company asks what the candidate owned and shipped, and whether they actually led. Tightening that screen meant fewer people got through. - An engineering recruiter asks how candidates use agents, what frustrates them about agents, and how they would decide what to build. - Nobody has shown that these answers predict later stages. ## Inbound, Screening and Fraud ### 14. Is Inbound Still Worth Keeping, given Fraud and AI-written Applications? - Answers conflict. One growth-stage company still makes about one hire in five from inbound. Contractors read the applications, with help from the ATS, and the full-time recruiters do not. An early-stage company keeps inbound and scores candidates against criteria on well-known startup job boards. - Others have stepped back. An early-stage company turned its application form off. A growth-stage company has thousands of applications sitting unread and sources instead. Another growth-stage company has mostly given up, and inbound is now a small fraction of hires. - In between: a growth-stage company finds inbound fine for general backend roles and not for specialized ones. ### 15. How Do You Screen the top of the Funnel at Scale when Agents Apply on Candidates' Behalf? - What works is narrow. One early-stage company scores applications against criteria its engineers wrote. One growth-stage company makes the application a small project and grades it with AI. - Generic AI filters got poor reviews from recruiters we spoke with. One said a popular ATS's screening was useless. Another said it is only as good as the criteria you write. At an early-stage company, the ATS score let a fake candidate through. - A head of recruiting will not use an AI interviewer. The candidates they want will not get on a call with a bot. ### 16. Which Fraud Tools Work, and how Do You Spot Fake Candidates? - Tofu is the tool people praised most. A growth-stage company drops anyone it flags, so its false-positive rate is unknown. A recruiter ranked it first. - A growth-stage company trusts the fraud check in its ATS. A recruiter said that kind of check flags a large share of remote applicants, and also flags real candidates who apply through an agent with a new email address. - Recruiters described the same manual tells: a missing or empty LinkedIn, a resume that looks too perfect, a voice that does not match the claimed background, audio lag, and failing simple questions about the place the candidate claims to be. - Companies that bring final rounds in person said fraud there is rare. ## Sourcing and Engagement ### 17. Do AI Sourcing Agents Actually Work, and how Do You Verify what They Find? - A growth-stage company built its own agent on a web-search API, with a profile it rewrites as recruiters give feedback. They think it works. - One company's own agent finds nothing a bought tool would miss. Two vendor products add a little. - A recruiter disagrees that this can be fixed with a better model. Agents do not learn taste, and sorting what they return takes as long as searching by hand. A growth-stage company watched a vendor agent turn to junk after a few days. - Verification is unanswered. Everyone still reads the results. Nobody has automated that. ### 18. How Do You Build a Candidate List that Isn't the Same as Every other Company's? - One recruiting lead said there is no trick: taste, making the outreach feel human, and the hours spent. - One team starts from products it admires and asks who built them. - A former operator watches for a candidate's manager leaving, or for insiders selling stock. - An early-stage company filters for a long stretch on one specialized team at a large company. ### 19. How Do You Get Candidates' Attention when Every Recruiter's Message Sounds the Same? - A growth-stage company leads with a mutual connection. - One company writes the outreach in the hiring manager's name, and asks new hires to name people worth sourcing. - At two companies, the head of engineering takes the first call. - A recruiter said outreach personalized by AI reads as trying too hard. ### 20. What Unusual Candidate-experience Moves Are Worth Copying? - One company runs recruiting on the idea in *Unreasonable Hospitality*. Its application asks for a favorite junk food, so the team can do something specific for that candidate. The point is to make the process feel human. - A former operator gives each hiring manager a daily batch of profiles that expire, and a one-step way to invite the candidate to dinner.