What you are buying
Most AI startups for sale at the small-business end of the market are applications built on top of third-party models: a workflow, prompt pipeline, retrieval layer, or interface that calls an API from a model provider. What you acquire is that application layer, plus its customers, brand, data, and evaluation know-how. You generally do not acquire the model itself, so the product's quality, cost, and availability depend partly on contracts and roadmaps you do not control.
That dependency is concrete. OpenAI's deprecation policy states that, unless safety or compliance concerns require a faster timeline, generally available models get at least six months' notice before shutdown and preview models can get much shorter notice. Anthropic's documentation says it gives at least 60 days' notice before retiring publicly released models. Either way, a product tied to one model version will need re-testing and prompt changes on a schedule set by the provider. When an AI business changes hands, the buyer also needs new API accounts or a transfer of the existing ones, fresh API keys, and a review of whether the provider's terms permit the product's use case under the new owner.
What to check before making an offer
Map every model and API dependency
Ask for a list of each model, provider, and model version the product calls, with monthly volume. Check each against the provider's deprecation page for scheduled retirements. Products pinned to a deprecated or preview model carry near-term migration work.
Rebuild gross margin from inference bills
Request several months of model-provider invoices alongside revenue. Calculate inference cost per paying customer and per heavy user. Flat-price plans with unlimited usage can lose money on power users even when the total looks profitable.
Test switching cost to another model
Ask whether the product has an evaluation set and how it performed on alternative models. If there is no evaluation harness, assume switching providers will take real engineering time. A business that can swap models with measured quality is more resilient.
Confirm data rights and customer commitments
Review the privacy policy, terms of service, and any customer contracts that promise data will not be used for training or will stay in a region. Make sure those promises are compatible with the providers you will use after the sale. Ask how user-uploaded data is stored and deleted.
Look for durability beyond the base model
Identify what a competitor could not copy by calling the same API: proprietary data, integrations, distribution, workflow depth, or a loyal niche audience. If the main feature is now available natively in the model provider's own app, discount future growth accordingly.
How these businesses are valued
AI businesses are valued with the same income, market, and asset lenses as any other small company, but buyers usually adjust harder for two risks: margin volatility, because inference prices and usage patterns change, and platform risk, because a model provider can ship a competing feature or retire the model the product depends on. A business with revenue that grew quickly in a short window deserves extra scrutiny of retention, because early interest in AI tools does not always become durable subscription revenue.
Use verified trailing revenue to confirm the top line, then spend most of your diligence on cost of revenue and retention. Normalize inference spend to current provider pricing, model what happens if your largest provider raises prices or retires the model in use, and give more weight to recurring revenue from customers who have stayed for several months. RevenueBug does not appraise listings or publish AI valuation multiples.
Model a range with the SaaS valuation calculator, then work through the buyer due diligence checklist.