Angel investors backing AI startups check the same fundamentals they check anywhere — team, market, traction, terms — plus a short list of questions specific to machine learning businesses: what the company actually owns, what its compute and model costs do to gross margin, and whether its advantage survives the next release from a large model provider. An angel investor is an individual putting their own money into early-stage companies, usually at pre-seed or seed stage, often alongside other angels or a lead fund.
The hard part is not finding AI deals. It is separating a company with a durable business from one whose product is a thin layer over somebody else's model, priced below what it costs to run. The checks below are the ones that tend to surface that difference before a wire goes out.
Table of Contents
- What does the company actually own?
- Unit economics when inference costs real money
- The platform-risk question
- Reading traction honestly
- Terms, dilution, and what an angel check actually buys
- A practical diligence sequence
- Frequently Asked Questions
What does the company actually own?
Start with the stack. Ask which parts of the product the startup built, which parts are licensed, and which parts are API calls to an outside model provider. All three can support a good business, but they carry different risks, and founders sometimes describe the third as if it were the first. If the company fine-tunes or trains its own models, ask what data it trained on and what rights it has to that data.
Data obtained by scraping, by purchase, or through a customer contract each come with different limits — a customer agreement may permit using that customer's data to serve them, while forbidding use in a shared model that benefits competitors. Ask to see the relevant clauses rather than accepting a summary. Then ask what accumulates. Proprietary data that gets better with usage, workflow integration that makes switching painful, distribution into a hard-to-reach buyer, or genuine regulatory clearance are all things a competitor cannot copy in a weekend. A prompt and a nice interface are not.
Unit economics when inference costs real money
Software businesses have trained investors to assume near-zero marginal cost. AI products often do not work that way, because every user request may trigger paid model inference — the compute cost of running a model to produce an answer. Ask for gross margin, and ask specifically whether the cost of serving customers is inside it. Request the cost per unit of whatever the customer buys: per seat, per document processed, per support ticket resolved, per minute of audio.
Then ask how that cost has moved over the past few quarters and why. Model prices have generally fallen over time, which helps, but heavy users, long context windows, and agent-style products that make many calls per task can push costs up faster than prices fall. Watch for the pattern where a flat monthly price meets unlimited usage. It converts the company's best customers into its biggest losses, and it is often invisible in blended averages. Ask for margin on the top decile of users, not just the mean.
- Cost to serve one unit of the thing customers pay for, stated in dollars
- Gross margin after inference, hosting, and any human review in the loop
- Margin on the heaviest 10% of accounts
- What happens to pricing if a model provider raises rates or deprecates a model
- Whether a cheaper or self-hosted model has been tested as a fallback
The platform-risk question
Every AI startup building on a third-party model faces the same question: what happens when the provider ships the feature as a native capability, or adds it to a general assistant for free? This risk is real but frequently overstated and sometimes understated, so it is worth making concrete instead of philosophical. Ask the founders directly which parts of their product a major model release could absorb, and what would remain. Good answers are specific and unglamorous: the integrations with a customer's internal systems, the labeled data from years of corrections, the compliance documentation, the trust of a buyer who will not send data to a general-purpose tool.
Weak answers lean on speed of execution alone. Also check single-provider dependence as an operational matter. Deprecated model versions, rate limits, pricing changes, and terms-of-service restrictions on specific uses have all disrupted products built on one API. Ask whether the architecture allows swapping models, and whether anyone has actually tested it.
Reading traction honestly
AI products generate unusual amounts of curiosity-driven signup. Headline user numbers therefore tell you less than usual, and retention tells you more. Ask for a cohort retention curve — the share of each month's new customers still active and paying several months later — and look for whether it flattens or keeps sliding toward zero. Separate pilots from production.
A large company name on a slide may represent a paid, deployed deployment with a renewal date, or a free trial with one curious team and no budget line. Ask how many pilots have converted, how long conversion takes, and who inside the customer signs the renewal. For businesses selling to enterprises, ask whether the product has cleared security review, where the customer's data lives, and whether the customer has agreed to its data being used for training. These reviews often determine the real sales cycle length, and a founder who has been through several can describe them in detail.
Terms, dilution, and what an angel check actually buys
Most angel investments at this stage use either priced equity rounds or convertible instruments — a SAFE or convertible note that turns into shares at a later round. Read the valuation cap, the discount, and whether it is pre-money or post-money, because post-money SAFEs shift dilution onto founders and earlier holders in ways that compound when a company stacks several of them. Ask for a capitalization table that includes all outstanding convertibles, not just issued shares. A company that has raised a large amount on stacked SAFEs at different caps can be far more diluted than its stated valuation implies.
Also ask about pro rata rights, which let you maintain your ownership percentage in later rounds, and about information rights, which determine whether you get financial updates at all. Be clear-eyed about the base rate. Most angel investments return nothing, returns concentrate in a small number of winners, and holding periods commonly run many years with no way to sell. Sizing individual checks so that a total loss is survivable matters more than any single diligence insight — and in the United States, most private rounds are open only to accredited investors, a status defined by income or net worth thresholds under SEC rules.
A practical diligence sequence
Work in a fixed order so you spend time only on companies that pass the cheap tests first. Most of this can be done in two or three conversations plus a few reference calls.
If founders cannot produce cost-to-serve figures, treat that as information rather than an oversight. Teams running a real AI product watch that number weekly, because it is the one that decides whether growth makes them stronger or poorer.
- Confirm the problem exists by talking to two customers the founders did not pick for you
- Get cost-to-serve and gross margin in writing, including heavy users
- Ask which parts of the product survive a major model provider shipping the same feature
- Review data rights: training data provenance and customer contract clauses
- Request cohort retention and pilot-to-paid conversion numbers
Frequently Asked Questions
How much should an individual angel put into a single AI startup?
There is no correct figure, but the standard discipline is to size each check so that losing all of it changes nothing material about your finances, and to plan on a portfolio of many small positions rather than a few large ones.
Is it a bad sign if a startup uses a third-party model instead of training its own?
No. Training a model from scratch is expensive and rarely the right use of seed capital. What matters is whether the company accumulates something defensible — data, workflow depth, distribution, or regulatory position — on top of whichever model it uses.
Can I rely on the lead investor's diligence?
Partly. A credible lead reduces your work on legal terms and structure, but a fund's return math differs from yours. They can afford a portfolio built on a few outsized outcomes; your position may be too small to follow on into those winners unless you secured pro rata rights.
What is the single most common thing angels miss in AI deals?
Gross margin measured after inference costs. Revenue growth on a product that loses money per heavy user looks identical to healthy growth in a top-line chart.