What you need to know
Hiring in AI has been a candidate's market for specialist skills and a founder's market for everything else — capital is volatile, model costs are falling but inference bills are not, and the gap between a company's pitch deck and its actual numbers has rarely mattered more to the person about to sign an offer. Candidate-side diligence — treating the interview as a two-way negotiation where you are also allowed to ask hard questions — is no longer an eccentric extra step. It is the same rigour an investor applies before writing a cheque, just pointed the other way.
- Runway is a date, not a vibe. Ask for the cash-zero month specifically; a founder who has done the maths will have it memorised.
- Most "AI products" are still thin wrappers. Ask for the company's own evals, not a demo, and test whether a public chatbot could reproduce 80% of the output.
- Data provenance is now your personal exposure too. DPDP in India and UK GDPR both carry obligations that land on whoever is close to the pipeline — including engineers.
- Inference is quietly the biggest new cost line in AI products. Ask what share of revenue it eats, and whether that share is falling.
- Who leads a round, and whether it was priced or a bridge, tells you more than the headline amount raised.
- How founders answer these questions matters as much as what they say. Deflection, irritation and shifting answers are signal in themselves.
Ask a co-founder or the finance lead directly, not the recruiter — a recruiter usually doesn't know the real numbers and can only relay a sanitised version. Frame it as understanding the company's trajectory rather than as a challenge: "Can you walk me through your current runway and what the plan is if the next round takes longer than expected?" is a request any well-run company can answer without flinching.
Runway and the cash-zero date: the maths founders should already have cold
Runway is simple arithmetic — cash on hand divided by net monthly burn — but the number that actually matters is what it produces: a specific calendar month by which the company runs out of money if nothing changes. Ask for that date, not the number of months. A founder who has genuinely internalised their own financial position will answer in seconds; one who has to think about it, or answers in vague ranges ("a year, year and a half, something like that"), is telling you they haven't been looking closely, or would rather you didn't know precisely.
The bar has moved. Per multiple 2026 startup-finance trackers (Puzzle's founder guidance, Culta's stage benchmarks and First Round Capital's State of Startups research), the loose 12-to-18-month cushions common when capital was cheap have given way to a firmer 18-24 month expectation post-raise, with many VCs treating anything below that as a company that will be back fundraising sooner than its story suggests. First Round's data reportedly found that companies holding 12+ months of runway have roughly 3.5 times better survival odds than those under six months — not a subtle difference, and one worth asking a founder whether they track at all.
Increasingly, investors also look at the burn multiple — net burn divided by net new annual recurring revenue — rather than burn rate alone, because it tells you whether the company is buying growth efficiently or simply spending faster. A company burning heavily but growing revenue proportionally faster is a very different proposition to one burning the same amount for flat or declining growth. Ask which number the founders track internally; "we don't really look at it that way" is itself an answer.
| Metric | Company A (hypothetical, Bengaluru seed) | Company B (hypothetical, London Series A) |
|---|---|---|
| Cash on hand | ₹9 crore | £2.4M |
| Net monthly burn | ₹75 lakh | £160,000 |
| Runway (months) | 12 months | 15 months |
| Cash-zero date, from today | July 2027 | October 2027 |
| Reading against the 18-24 month bar | Short — expect a raise inside 6-9 months, ask about progress on it now | Borderline — ask what milestone the next round is tied to |
Entirely illustrative figures for two hypothetical companies, not real cap-table data — the method is what matters, not these numbers. Substitute the actual figures a founder gives you and do the same division.
A founder who answers the runway question but visibly bristles at a natural follow-up — "and what's the plan if the raise takes longer than expected?" — is worth noting. Confidence in the number and comfort discussing the contingency are two different things, and the second is arguably more predictive of what your first year there will actually feel like.
Real moat or thin wrapper: pressure-testing the product
By 2026, calling foundation-model access a moat is close to meaningless — every credible competitor can call the same API. The practical question for a candidate is whether the product you'd be building has anything that survives a determined engineer trying to replicate it over a weekend with the same underlying model.
The 80% test
A blunt but genuinely useful heuristic, echoed across several 2026 write-ups on AI product defensibility: if a technical user pasted the product's core system prompt into a public chatbot such as Claude or ChatGPT, could they reproduce roughly 80% of the output? If yes, the product is a thin wrapper — a system prompt and an interface — and its defensibility rests entirely on distribution and speed to market, not the technology itself. That's not automatically fatal to a company, but it changes what you should be evaluating: go-to-market execution and retention, not "the model".
The evals and data-flywheel test
Ask to see the product's own evaluation suite — not a benchmark screenshot of the underlying foundation model, but the company's internal tests against its own task. Has the score improved month over month? Is there a dataset of real corrections and outcomes — user edits, accepted or rejected suggestions, resolved tickets — that only exists because the company is in the workflow, and that a new entrant could not simply buy or scrape? That accumulating, hard-to-copy signal is one of the few durable sources of defensibility left; a company months into shipping that still can't show you an evals trend, or has never structured its own outcome data, likely doesn't have one yet.
The complementary question is about switching cost: if a customer wanted to leave tomorrow, would they lose only an API key, or would they have to migrate data, retrain a team and rebuild integrations? A product that has become a genuine system of record inside a customer's workflow is a different bet to one that is a single call away from being swapped out.
Ask this single framing question directly in the interview: "If OpenAI, Anthropic or Google shipped this exact capability as a default feature in their next release, would your customers cancel?" A founder with a real moat has an answer ready and it isn't just "no, because we're faster to ship."
Do not let a polished demo stand in for evidence. A demo shows the happy path on a chosen example; it tells you nothing about failure rates, evaluation trends, or whether the "AI" layer is doing meaningfully more than formatting a single prompt call.
Data rights and compliance: GDPR and DPDP exposure that becomes your problem too
Data provenance questions used to be an investor's problem. They are increasingly yours as well, because as the engineer closest to the data pipeline you are often the one who inherits a shaky consent story, and because both India's and the UK's regimes now carry real, individually enforceable teeth.
In India, the Digital Personal Data Protection Act carries no minimum turnover threshold, no employee-count exemption and no SME carve-out — any organisation that digitally processes the personal data of Indian residents is in scope regardless of size. Per coverage of the DPDP Rules, most operational obligations phase in over roughly 18 months with full effect by mid-2027, which makes 2026 the practical build year for any startup that touches user data — including data folded into model training. Penalties are not trivial: failure to maintain reasonable security safeguards can attract fines running into the hundreds of crores of rupees, with separate, similarly sized exposure for breach-notification failures and children's-data violations. Worth asking directly: is there a documented lawful basis and consent trail for any personal data used in training or fine-tuning, and does the company maintain a data-flow inventory showing what goes to which model provider?
In the UK, the Information Commissioner's Office has published dedicated guidance on generative AI that expects organisations to assess training-data provenance, document risk in their outputs, and keep Data Protection Impact Assessments current for generative-AI-specific use — not treat an older DPIA as still valid by default. Layered on top, the Data (Use and Access) Act 2025, which received Royal Assent in mid-2025, reshapes several mechanics UK AI companies rely on, including a statutory definition of "scientific research" that can lower the bar for using personal data in research and development contexts. The ICO's enforcement posture has also hardened — its fine totals rose sharply year-on-year through 2025 — so "we haven't been fined yet" is a weak reassurance rather than evidence of good practice. Ask whether a DPIA has been updated since the company started using generative AI, and whether anyone can name the lawful basis for the data feeding the product.
Our companion piece on DPDP Phase II's six-month compliance playbook goes deeper on the India side if you want the fuller regulatory picture before an interview.
If nobody in the interview loop can describe, even roughly, where training or fine-tuning data comes from and what consent or lawful basis covers it, treat that as a genuine gap rather than an oversight you'll fix once you join. Retrofitting data governance onto a live product is materially harder — and more political — than building it in from day one, and you may be the one asked to do it.
Compute economics: who is really paying for the tokens
Inference is the AI-era equivalent of cloud hosting costs, except it scales with usage in a much less forgiving way — every additional user query runs the model again, so as a product grows, its inference bill grows roughly in step, while training costs stay comparatively fixed. Per ICONIQ Capital's January 2026 State of AI report (based on a survey of roughly 300 software executives building AI products), AI companies now run cost-of-goods-sold of around 40-50%, with inference alone accounting for close to 23% of revenue at scaling-stage companies — leaving gross margins around 50-60%, against the 75-90% traditional SaaS has historically enjoyed at scale. The same report found overall AI gross margins had improved to roughly 52% by January 2026, up from about 41% in 2024, but still well short of software-era norms.
What this means for a candidate: ask directly what share of revenue goes to inference, and — more importantly — whether that share is falling as the company scales, or holding flat. A falling share suggests real engineering leverage: caching, model routing, distillation, or negotiated compute deals. A flat or rising share, even alongside revenue growth, suggests the business is on a treadmill where growth doesn't translate into improving unit economics — and that gap eventually shows up as pressure on hiring budgets, on-call load, or your own compensation review.
A useful, non-confrontational way to ask: "How has your blended cost per request changed over the last two quarters, and has it fallen faster or slower than your price to customers?" A founder who tracks this will answer with a number and a trend; one who doesn't will answer with a feeling.
Team stability and funding-round quality: reading the humans and the cap table
Reading team signals
Key-person risk is more acute at a startup than almost anywhere else — a founding engineer or a specific domain expert often carries knowledge, customer relationships or model-tuning judgement that isn't written down anywhere. Ask plainly who has left the company in the last twelve months, at what level, and why, and watch how directly the answer comes. A quick, specific answer ("our second engineering hire left in March for personal reasons, amicably, here's who backfilled it") is a very different signal to a vague one. A light pass over LinkedIn before the interview — looking for a cluster of departures updated within the same month, or several senior people quietly marked "former" — is a reasonable and increasingly normal part of preparing for a startup interview.
Reading the funding round
Not all money is the same money. A round led by a firm with a genuine track record of follow-on support carries a different signal than one where the "lead" is a collection of smaller cheques with no clear anchor — tier-one investors tend to exert a gravitational pull on future rounds precisely because their diligence and continued backing are themselves a quality signal to the market. Ask, plainly, who led the last round and whether they're expected to participate in the next one.
Also ask whether the last raise was a priced round or a bridge — and don't accept "extension" as automatically reassuring; per 2026 market coverage, bridge financings made up a striking share of Series A and B activity in the first quarter of the year, and founders don't always volunteer the word "bridge" even when that's what happened. Signs worth probing for: a valuation that was flat or fell against the prior round, options that were quietly repriced, convertible notes stacking up without a clear next priced round in sight, or a founder who answers "how did the last round go?" with enthusiasm about the relationship but no actual numbers.
"The interview question that told me the most wasn't about the model at all — it was 'who left in the last year, and why.' The founder answered instantly, by name, with what each person was doing now. At my previous company, that same question got a long pause and a change of subject. I should have listened to the pause."
— Meera, Verified Builder · London, UKThe scored checklist — and what red-flag answers sound like
Bring this into the interview itself, whether over a call or as your own private notes afterwards. None of these questions is unusual or aggressive to ask — a well-run company, whether an early Bengaluru team or a scaling London Series B, should be able to answer most of them plainly.
| # | Ask this | Strong signal | Red flag |
|---|---|---|---|
| 1 | What's your cash-zero date? | A specific month, given without hesitation | A vague range, or visible irritation at the question |
| 2 | Was the last round priced or a bridge? | Clear, specific answer including who led it | "Extension" language with no detail on terms |
| 3 | What would a public chatbot reproduce of your core output? | Founder names a specific proprietary data or workflow advantage | Defaults to "we move faster" with nothing else |
| 4 | Can I see your own evals, not a foundation-model benchmark? | A real eval suite exists and trend is discussed openly | No internal evals, or only a demo is offered |
| 5 | What share of revenue is inference, and is it rising or falling? | A tracked number with a clear trend | "We haven't really broken that out" |
| 6 | Who left in the last 12 months, and why? | Specific, named, unembarrassed answer | Deflection, minimising, or a long pause |
| 7 | Is your data-flow inventory / DPIA current for AI use? | Named person owns it and can describe it | Nobody in the room knows what a DPIA is |
| 8 | What's the plan if the next raise takes longer than expected? | A concrete contingency — cost cuts, bridge options, milestones | "It won't" with nothing behind it |
Score two points for a strong signal, one for a partial or hedged answer, zero for a red flag or refusal to answer. Six or more strong signals out of eight is a genuinely healthy company by these measures; below four, treat the offer as higher-risk regardless of the headline compensation.
The pattern across nearly all of these red flags is less about the content of the answer and more about its shape: specific, numeric, unembarrassed answers are a good sign almost regardless of what the number actually is, while vague, generic, or defensive answers are a bad sign even when no number is technically wrong. Investors have learned to read founders this way during their own diligence; there's no reason a candidate shouldn't apply the same lens. If equity forms a meaningful part of what's on the table, read it alongside our guide to vesting, exercise windows and dilution — the healthiest company can still hand you a badly structured grant, and the two questions are worth separating clearly. And if you're still weighing the decision at a higher level — not just this one company, but the shape of your career — our comparison of startup vs Big Tech vs AI lab paths and our India-UK pay benchmarks cover the cash side of the same conversation.
None of this changes the basic asymmetry of a job interview: the company is also evaluating you, and in a market where, per recent coverage of India and UK AI hiring, demand for specialist AI engineers continues to outstrip supply, a candidate who arrives with sharp, well-informed questions is more memorable, not less. The same logic runs in the other direction, too. While you're doing this diligence, the people across the table are quietly diligencing you — your GitHub, your write-ups, whether you can be found and verified at all. A Verified Builder profile is the artefact that does that work for you before the first call: proof of what you've actually shipped, not just a resume claiming it. Hiring teams across India and the UK are already browsing profiles like these to shortlist who they reach out to — and the earliest profiles carry a Founding Builder badge that later cohorts won't get. It costs nothing and takes about two minutes to set up, which is a lower bar than any of the eight questions above.