Cap Table Decoded
Cap Table DecodedFourteen Due-Diligence Questions Every ML Engineer Must Ask a Series A AI Startup Before Signing
Equity NormsLong read

Fourteen Due-Diligence Questions Every ML Engineer Must Ask a Series A AI Startup Before Signing

Fourteen questions that expose hollow AI startups before you sign.

· 10 min read

An ML engineering offer from a Series A AI startup is worth pursuing when the company can prove its technical foundation, afford the role, and define the work. These fourteen due-diligence questions separate high-impact opportunities from career setbacks by testing those claims before you sign. A pitch deck is not evidence, and attractive equity cannot rescue a product nobody understands.

Why the diligence burden has shifted to the ML engineer at Series A

Historically, the startup evaluated the candidate. In 2026, the evaluation is genuinely two-sided because joining a technically hollow company can cost an ML engineer years. Series A AI companies compete with frontier labs and large tech companies for engineers with demonstrable 0-to-1 shipping evidence. That scarcity gives candidates leverage before they sign. Use it to ask about data ownership, inference economics, and model architecture, which a polished demo tells you almost nothing about.

How to evaluate an ML engineering offer using the fourteen questions in sequence

Treat the questions as filters. Questions 1–6 test the technical foundation, including data and IP integrity, and expose failures that can end the process. Questions 7–8 ask whether the company’s finances and equity terms give you a real chance to gain from the offer. Questions 9–11 test whether the role is clear and the company supports the work, while questions 12–14 cover working conditions, team health, and what you are expected to deliver. Follow that order so a persuasive founder or generous-looking grant does not distract from an earlier disqualifier.

Whether the AI is foundational or cosmetic — and why this is a kill criterion, not a nuance

Ask what specific workflow step depends on the model. The CTO or founding engineer should explain the architecture without substituting a third-party provider’s marketing for an answer. Then ask to replace a recorded walkthrough with a product run in a sandboxed environment. The live product should resemble the pitch.

A strong answer connects an indispensable model capability to in-house ML talent and a working product. A weak one resolves into scheduled API calls and a UI layer, with no clear account of what the company itself has built. That distinction determines whether you will develop new capability or defend a wrapper as its features become commonplace.

Who owns the training data, and whether the chain of title survives scrutiny

Ask for a provenance map showing what percentage of the training corpus is proprietary, licensed, synthetic, or web-scraped. For synthetic data, ask how validation covers real-world edge cases. Licensing agreements should support the ownership claims, and a named person should own data governance. Public benchmarks are not proprietary training assets.

Owning the data is not enough to stay ahead over time. A strong answer also explains how using the product feeds a data loop that gets better as customers interact with it, backed by work that competitors cannot easily copy. A cleaned Common Crawl dataset does not meet that standard. Undocumented provenance creates legal and funding risk, and delays in financing can reach directly into the runway supporting your job.

The unit economics of inference, and what happens when no one modeled them before you arrived

Ask for inference cost per active user alongside the model showing how gross margin improves with scale. The assumptions need to be explicit. A promise to optimize later is not a cost model, and the missing work is likely to become your assignment after joining.

Architecture makes this an engineering question requiring scrutiny beyond the CFO. Multi-agent systems with memory and tool use introduce compounding token costs and latency that need active management. A strong answer spells out those costs and shows that margin will stay above the floor the company needs at the scale it plans to reach. A weak answer assumes that growing revenue is enough without explaining what it will cost to serve that growth.

Whether the codebase can be explained by a human — and why AI-generated code requires a new diligence standard

Ask the founding engineer or CTO to walk through authentication, billing, and data handling without opening an AI assistant. Who wrote the code is not the decisive issue. Who understands it is the better test.

These paths expose whether the team has engineered a product or assembled code it cannot explain. Thin tests and stale dependencies are fixable gaps, but client-side authentication, secrets in git history, or only one person understanding billing change the risk of the offer itself. They are repricing events. If nobody can explain the critical paths, the new hire becomes the implicit owner of technical debt, diverting their work from contributing new capability.

What the eval infrastructure reveals about whether the team can tell when the model breaks

Ask how the team diagnosed its last production model failure. A strong answer traces the error to the training data and code version behind it and goes beyond describing an output that fell short. Request the eval harness, model registry, and version-control tools such as DVC, LakeFS, or an equivalent. Also ask about drift detection and retraining triggers.

The incident response needs equal scrutiny here. Ask what happens when a model produces harmful output or leaks data. Request a postmortem from an actual incident. A tested playbook and continuous monitoring show operational maturity. Release-only benchmarks or claims that the model has never failed leave open whether anyone could detect a failure at all.

Foundation model dependency and what happens when the underlying API changes without warning

Ask for the documented list of model and provider dependencies, then ask what happens if pricing, availability, or terms change. A prepared team has modeled at least two scenarios and can name fallbacks for its top two providers. Someone should own that contingency work explicitly.

This question extends beyond the immediate API bill. Relying on a single cloud provider’s AI stack, with no plan for multi-cloud or on-prem deployment, can make it harder to sell to enterprises. A strong answer includes a roadmap for moving core capabilities to fine-tuned open-source alternatives where vendor risk is highest. A weak one leaves an external company in control of a dependency your team would have to replace under pressure.

Runway, burn, and whether the headcount figure the founder quotes includes yours

Ask whether the quoted runway includes your fully loaded compensation along with the headcount plan for the next 12 months. Current burn can make the answer look reassuring while planned hiring shortens the actual window. Refusing to share runway after an offer is a red flag.

Also establish whether the company is in active Series A diligence, pre-process, or post-close. Those are different financial positions, despite the stage label used in recruiting. Ask for the burn multiple and compare it against industry benchmarks rather than taking the size of the raise as proof of health. A company with 12 months of runway and no clear funding path presents a cliff.

The full equity picture — diluted share count, preference stack, strike price, and exercise window

Ask what percentage of the fully diluted share count the grant is. An option count alone tells you little. Then ask for the total liquidation preference and whether it is participating or non-participating, because preferred investors are typically paid before common holders in a sale.

Ask how the strike price compares with the 409A valuation at the time of grant. A large gap raises tax and economic questions that belong in the offer discussion. Finally, ask for the post-termination exercise window and whether an extension is available, under what conditions. The standard 90-day window can force an employee who leaves to decide whether to exercise, which can cost a lot and may have tax consequences.

Which of the three ML archetypes you are actually being hired for — and why the distinction shapes your day-to-day work

Ask which job the company actually needs done. An applied AI engineer turns existing models into useful products, an ML engineer makes ML systems reliable and scalable, and a research scientist creates or validates new methods. Those roles need different skills and success criteria. The title does not settle the distinction.

Ask what percentage of the role involves building new ML capabilities versus maintaining and improving existing ones. Then ask why the company chose the advertised title. Too many researchers, too few product-minded engineers, and nobody owning evaluation or inference cost create gaps that the next hire absorbs. At Series A, a structural mis-hire can stall the roadmap for a quarter.

EU AI Act compliance posture — and why it is an engineering deliverable, not a legal team problem

Ask whether the company has worked out which of the EU AI Act’s risk tiers its AI systems fall under and started a conformity assessment for high-risk applications. Engineers need to take part in risk classification, documentation, transparency labeling, and monitoring. Lawyers cannot supply the technical work alone.

A strong answer names each product’s risk tier and gives a compliance roadmap that says who is in charge and how much money is set aside for it. It also shows that legal and engineering have discussed the obligations together. For a company selling to enterprises in regulated verticals, ask about a dedicated compliance engineer or a firm, funded hiring commitment. A vague referral to legal signals that governance work may arrive on your desk retroactively and under deadline pressure.

Whether a forward-deployed engineering function exists — or whether that work will fall to you by default

Ask who owns customer deployment after the sale and who responds when an integration breaks. A forward-deployed engineer embeds with customers, builds integrations, owns deployment, and brings product intelligence back to the core team. Without that function, ML engineers often absorb the work. It is valuable work, but a different assignment.

The answer should reveal whether ownership is staffed, informal, or unresolved. An absent function calls for explicit negotiation about customer-facing work versus core ML development. Keep that answer for the delivery discussion later. A company cannot credibly assign you deployment responsibility while planning as though all your time belongs to the product roadmap.

How the remote or hybrid policy actually functions — and what the compensation geography means for your offer

Ask whether compensation is location-based or the role has a single national pay band. Do not infer pay policy from “remote.” The label can describe a salary equivalent to that in a high-paying region, a location-adjusted salary, or a decision the company has not made.

For working conditions, ask about core hours, documented asynchronous practices, and how decisions get made across time zones. A hybrid or remote policy is only useful if the operating rules are clear. Then ask who handles production incidents outside business hours and how on-call is assigned. “Everyone wears many hats” does not explain coverage, and an undefined answer suggests that writing the policy may become part of your job.

What technical and leadership departures in the last 12 months reveal about the team's actual health

Ask what happened to previous people in the role and whether founding team members or early technical leads left in the last 12 months. A repeatedly reopened position deserves a direct explanation. The vacancy itself is not the whole story.

Ask whether the pattern reflects hiring standards, undefined scope, or work that differed from what candidates were promised. A strong answer explains what changed in the role and the team after a departure. A weak one skips over what happened while promising easy AI milestones or model-performance timelines that leave out the time needed for research. Those answers suggest the next engineer may inherit the conditions that drove out the last one.

What you are expected to have shipped in 90 days and in 12 months — and why vague answers are a warning, not modesty

Ask what you should have shipped at both milestones. At Series A, the bottleneck is often repeatability rather than the ability to ship another bespoke feature. A useful goal could be to raise experiment throughput while keeping incidents, regressions, and model spend per active customer from rising.

Ask which non-engineering duties will interrupt that work and how often. Ask what gives way when they do. Compare the answer with deployment ownership alongside the history of departures. Undefined expectations plus repeated backfills suggest a structurally misspecified role. By this point, your questions have reached details a pitch deck cannot settle, giving you concrete leverage to negotiate the actual job.

What strong and weak answers across the fourteen questions tell you about which companies are worth joining

Judge the answers by layer and by pattern. When data provenance and eval infrastructure are missing, the company has gaps in how it is built that go beyond trouble explaining its work. Candor about unfinished work is more credible than confidence that collapses under questioning, but candor alone does not clear a technical kill criterion. Attractive equity cannot offset an absent cost model or an inexplicable codebase.

Treat correlated failures as grounds for walking away from negotiations. Proceed when the technical foundation holds, runway includes your headcount, and the company defines the first 90 days specifically. Then negotiate equity, the exercise window, the scope of the role, and the policy for remote compensation. The aim is not to reject every imperfect startup, but to choose knowingly, and the best companies will welcome that scrutiny.

Filed underEquity Norms

More in Equity Norms