← ALL WRITING
DILIGENCE

Wrapper or moat? 12 questions that show whose AI it really is

Almost every AI product is built on someone else's model. The diligence question is what the company owns on top of it, and what happens when the model underneath changes.

WRITTEN
READINGNine minutes
FORPE deal teams, VCs, corporate development, investment committees

A startup is "just a ChatGPT wrapper" when the third-party model is doing nearly all the work and the company owns nothing a competitor couldn't rebuild in a quarter. Being built on a foundation model is not the problem: almost every AI product is. Defensibility comes from what sits around the model (proprietary data and the rights to use it, how deeply the product is embedded in a customer's workflow, distribution, evals and switching costs), and from how exposed the business is if its model provider changes price, terms or the model itself.

The word "wrapper" has become a lazy insult in investment committees. I hear it used to dismiss companies that have real advantages, and I see it missed in companies that have none. It is a better question than it is an answer. The model is the part of an AI product a competitor can buy on the same terms you can. Diligence is about everything else.

The model is the part of an AI product a competitor can buy on the same terms you can.

What "wrapper" really means

A wrapper, in the narrow sense, is a product whose main function is a call to someone else's large language model, with a prompt, an interface and perhaps a search over the customer's own documents added on top. Strictly, that describes most of the AI application market.

Menlo Ventures' 2025 State of Generative AI in the Enterprise estimated enterprise spending on generative AI at $37bn in 2025, of which $19bn went on applications and $12.5bn on foundation model APIs. A large share of that $19bn is, technically, wrappers. Some are excellent businesses.

The useful distinction is between a thin wrapper and a thick one. A thin wrapper adds a prompt and a nice interface. A user with a ChatGPT or Claude subscription and an afternoon could get 80% of the value. A thick wrapper adds things the model provider doesn't have and won't build: data, integrations, domain-specific workflow, review processes, accountability to a regulated standard. The model is a component, the way Postgres is a component of a SaaS product. Nobody calls a CRM a "Postgres wrapper".

Why being built on a foundation model isn't itself a problem

Training a frontier model costs more than nearly any start-up will raise, and the result goes out of date within a year. Building on someone else's model is the rational choice for almost every company, and investors who insist on "proprietary models" are often asking a company to waste their money.

The economics also move in the application builder's favour. Epoch AI's 2025 analysis found that the price of reaching a fixed level of model performance fell by between 9x and 900x a year depending on the benchmark, and about 40x a year for GPT-4-level performance on a graduate-level science test. A company whose product gets cheaper to deliver every quarter, without doing anything, is in a good position, as long as it keeps the saving and its customers don't take it.

That proviso is the whole question.

Where AI defensibility actually comes from

In my experience the advantage almost never sits in the model or the prompts. Prompts leak, get reverse-engineered, or stop mattering when the next model is released. The durable sources are more boring.

Proprietary data, and the rights to it

Data a competitor can't get is the strongest moat in AI, and also the most overclaimed. "We have proprietary data" often means "our customers upload documents", which is the customer's data, under the customer's contract, and probably can't be used to train anything. Check the contracts. Data rights that permit aggregation and model improvement across customers are rare and valuable. Data that is public, scraped, or licensed on terms a competitor could also sign is not a moat.

The clearest example of data as defensibility is content licensing. In June 2025 LexisNexis and Harvey announced an alliance bringing LexisNexis primary law content and Shepard's Citations into Harvey's platform. Harvey's product sits on foundation models. What a new entrant can't easily replicate is that content agreement.

Workflow embedding

A product that sits inside the customer's daily process, writes back into their systems of record and is configured around their approvals is expensive to remove even if a better model appears. A product that sits in a separate browser tab is not.

Distribution

The best-built product often loses to the one that is already in front of the buyer. Exclusive channel relationships, a strong position with a specific professional community, or inclusion in a platform a market already uses can matter more than any technical feature.

Evals

An eval set is a library of real cases with known good answers, run against every model and prompt change. It is the least glamorous asset in an AI company and one of the most valuable. It lets the team change model providers in days rather than months, prove quality to customers, and catch regressions before they ship. A competitor starting from scratch has no equivalent, and building one needs domain expertise and time.

Switching costs

These come from all of the above: integrations, configuration, historical data, trained users and process built around the tool. Measure them through retention. Net revenue retention and logo churn say more about switching costs than any slide.

Model and provider dependency risk

The other half of the question is dependency. An AI product that relies on one provider carries three exposures most businesses don't think about until something happens.

Pricing is the first. Model prices have generally fallen, but the provider sets them, and the most capable models are usually the most expensive. A product priced on the assumption of last year's cheapest model can see its margin move with one announcement.

Terms and access are the second. In June 2025 Windsurf, the AI coding company, said Anthropic had sharply cut its first-party access to Claude 3.5 Sonnet and 3.7 Sonnet with little notice, at a time when Windsurf was reportedly being acquired by OpenAI. Most companies will never be in a provider's competitive line of fire, but the case shows that access is a commercial decision by the provider.

Deprecation is the third, and the one that actually happens to everyone. OpenAI's deprecations page gave notice on 14 October 2026 that several GPT-4o and GPT-4.1 snapshots will be shut down on 14 January 2027. Its published policy is at least six months' notice for generally available models, three months for specialised variants, and as little as two weeks for previews. Anthropic's policy commits to at least 60 days' notice for publicly released models. A replacement model is not a drop-in: it answers differently, follows instructions differently and breaks prompts that were tuned for its predecessor.

The market also moves under everyone's feet. Menlo's estimates put Anthropic at 12% of enterprise LLM API usage in 2023 and 40% in 2025, with OpenAI falling from 50% to 27% over the same period. The leading provider changed in under three years. Companies that could switch quickly captured the improvements. Companies hard-wired to one provider waited.

Some of the strongest application companies have responded by building models of their own for specific jobs. Cursor, the coding tool, released Composer, its own coding model, in October 2025, and its technical report on Composer 2 describes building it on an open-weight base model followed by large-scale reinforcement learning. That is what a mature answer to dependency looks like: own the parts where you have an edge and buy the rest. Few targets in a mid-market deal will be there. Most should at least be able to run on two providers.

12 questions that show whether a company's AI is really theirs

These are the questions I use in AI due diligence, in roughly the order I ask them. Each has a version of a good answer. A weak answer to one is a finding. Weak answers to most of them mean the AI is a feature anyone could ship. The full set of 87, with the evidence to request for each, is in our AI due diligence questions.

# Question What a good answer looks like
1 If you swapped your current model for the best alternative tomorrow, what would break and how long would it take? A specific answer: "we've run it on two providers, quality drops 3% on our evals, about a week of prompt work"
2 What does your product do that a user with a ChatGPT or Claude subscription can't do in an afternoon? Integrations, data, workflow and accountability they can name and demonstrate
3 What data do you have that a competitor can't get, and what do your contracts let you do with it? Contract clauses permitting aggregated use, plus data generated by the product itself
4 How do you measure quality, and can I see the eval set? Hundreds or thousands of real cases, scored on every release, with history
5 What does one unit of work cost you, all in, and how has it moved over 12 months? A per-task cost including model, infrastructure and human review, falling or stable
6 What share of revenue is gross margin after model costs, at today's usage and at three times today's usage? A modelled answer, with model costs in cost of sales rather than buried in R&D
7 Which provider terms have you read, and what do they say about data use, rate limits and price changes? Someone senior has read them and can summarise the risks
8 What happens to customers if your provider has an outage? A fallback provider, graceful degradation, or at least a measured history of outages
9 Where does the product write back into the customer's systems? Named integrations with systems of record, used by most customers
10 What do customers lose if they leave? History, configuration, trained models or workflows that don't export
11 What would the leading model providers have to build to compete with you, and why haven't they? A credible reason rooted in data, regulation, distribution or a market too specialised for them
12 How has net revenue retention moved since the last major model release? Stable or rising through each release, which shows customers aren't defecting to the raw model

Question 12 is my favourite, because it uses the market as the test. Every time a new frontier model lands, thin wrappers lose customers to it. Companies with real defensibility see no change, or gain, because the better model makes their product better.

Question 1 is the one I would put to the engineer rather than the founder. Founders answer it with confidence. Engineers answer it with the number of prompts that are tuned to one model's quirks.

How to score the answers

I don't think a numerical score adds much here. The answers usually sort a company into one of three groups, and each group means something different for the deal.

A company with strong answers on data rights, evals, integration and retention is a software business that uses AI well. Price it as software, and put the dependency questions in the 100-day plan.

A company with a good product but weak answers on data and switching costs is a well-executed thin wrapper. It can be a fine investment at the right price, with a clear plan to build what it lacks. Structure helps: milestones tied to retention through model releases, or a value creation plan that funds the eval set and the integrations.

A company whose answers retreat under questioning, where quality isn't measured, costs aren't known and nobody has read the provider's terms, has a product that the next model release may make redundant. That is a price conversation, and sometimes a walk away. The overlap with AI washing is large here, and the techniques for testing claims are the same.

Where AI due diligence fits

For a buyer, these twelve questions sit inside a wider technical due diligence review: architecture, code, security, team and IP, set out in the technology due diligence checklist. The AI layer needs its own attention because the risks are different and change faster. The separate question of whether AI threatens the target's market, rather than whether its own AI is real, is covered in the AI disruption risk assessment.

Our AI due diligence works through these questions with the target's engineers and data, tests portability and cost on the live system, and reads the provider contracts. It runs as a standalone review at £10k to £20k over one to two weeks, or as an addition to a full technical due diligence.

If a deal's equity story leans on AI and you want to know how much of it the company actually owns, the AI due diligence page explains how we run it and what you get at the end.

FREE RESOURCE · PDF + DOCX + XLSX
AI due diligence question bank
DOWNLOAD FREE →

Questions

What is a ChatGPT wrapper?

A ChatGPT wrapper is a product whose core function is a call to a third-party large language model, such as OpenAI's GPT models or Anthropic's Claude, with prompts, a user interface and sometimes retrieval over the customer's documents added on top. The term is usually used as criticism, meaning the product adds little a user couldn't get from the model directly. In practice it describes most AI applications, good and bad.

Does fine-tuning a model make an AI product proprietary?

Not by itself. Fine-tuning an open or commercial model on your own data can improve quality on a narrow task, but the advantage is only as durable as the data behind it. If the training data is public or easily bought, a competitor can fine-tune the same base model in weeks. Fine-tuning matters when it is built on data the company has the rights to and others can't get.

How do you value an AI company built on OpenAI or Anthropic?

Value it like any software company, then adjust for three things: gross margin after model costs at realistic usage, how much revenue sits on workflows and data a competitor can't copy quickly, and the cost of switching providers if prices or terms change. A company with high retention, owned data and tested portability deserves a software multiple. A thin interface on one vendor's model does not.

What happens to an AI startup when its model provider retires a model?

It has to migrate to a replacement, usually within a few months. OpenAI's published policy is at least six months' notice for generally available models, and Anthropic's is at least 60 days. Migration means re-running evals, rewriting prompts and checking behaviour, because newer models often answer differently. A company with a good eval set does this in days. One without can spend weeks finding regressions in production.

Can a wrapper company build a moat over time?

Yes, and the good ones plan for it from the start. Usage generates corrections, edge cases and outcome data that improve the product and can't be copied. Integrations deepen, customers build processes around the tool, and the eval set grows. The test in diligence is whether that is happening now, with evidence, or only promised in the roadmap slide.

Brannan Coady, founder of Anyway Ventures
Brannan Coady
Founder of Anyway Ventures. Exited founder, former CEO of YourParkingSpace, investor and board member; now building AI inside PE-backed and mid-market businesses. More →

Keep reading

ALL WRITING →
DILIGENCE · 9 MIN
AI washing: how to test a target's AI claims in due diligence
When the equity story leans on AI, the diligence question is simple. Take the AI away: what's left, and what does it cost to keep it running?
DILIGENCE · 10 MIN
AI due diligence for private equity: what to test before you sign
Most AI diligence reports check whether the target's AI is real and stop there. That is one job of three, and it is rarely the one that moves the price.
DILIGENCE · 10 MIN
AI disruption risk assessment: will AI hollow out the target?
AI washing covers targets that overclaim AI. This covers the other problem: targets that AI threatens. An eight-factor grid, the evidence to request, and what the score should do to the deal.

Talk about this piece.

The first conversation is free and takes under an hour. If there's an engagement worth having you'll leave knowing its shape and its price. If there isn't, I'll say so.

Book 30 minutes brannan@anywayventures.com