The distinction that now gets tested
Two years ago, describing a product as AI-enabled moved the conversation. Buyers had limited means to distinguish proprietary capability from a well-executed integration, and pricing reflected the uncertainty.
That has changed. Technical diligence on AI products is now specific, and it separates two kinds of business that used to be priced similarly. Assets with proprietary training data and defensible model work still clear at the top of the range for AI and ML platforms. Products that wrap a public model are priced as ordinary software, which is a different range entirely.
The gap is not a penalty. It reflects a real difference in what the buyer can own.
Where the training data came from
The first question is provenance, and it is now asked in writing.
Was the data licensed, and does the licence permit the use it is being put to? Was it collected from customers, and do those contracts permit training on it? Is any of it scraped, and if so from where? Can you produce the chain of rights for each material dataset?
Buyers ask because the exposure transfers with the asset and because a data source that cannot be defended is a source the buyer may have to stop using. A model that cannot be retrained on the same basis is worth considerably less than one that can.
The corollary is that clean provenance is now a value driver rather than hygiene. A seller who can document it is selling something the buyer can build on.
What happens if the underlying model changes
If the product depends on a third-party model, the buyer will want to know what happens when that provider changes its pricing, its terms, its rate limits or its model behaviour.
This is a supplier concentration question wearing different clothes, and it is priced the same way. A single-provider dependency with no tested alternative is a risk on the buyer's model. Demonstrated portability, where the product has actually been run against more than one provider rather than merely being architected to allow it, reduces the discount.
Inference cost as a margin question
Buyers now model inference cost per unit of revenue and its trajectory, because it behaves unlike traditional software cost of sales. It scales with usage rather than with customers, and heavy users can be unprofitable at a flat subscription price.
Expect the question of what the top decile of usage costs to serve. A seller who has that figure and has already priced or capped against it is answering a margin question. A seller who does not have it is presenting an unquantified liability, and buyers assume the worst case when they cannot see the number.
Whether the output is evaluated
The question that separates a serious product from a demonstration is how quality is measured.
Is there an evaluation set? What does it measure, how often is it run, and what happens when a change makes results worse? Is there human review in the loop, and at what cost? Can you show that quality has improved over time, or only that the product has shipped features?
A product with no evaluation regime cannot demonstrate that it works, which means the buyer has to take the customer retention data as the only evidence, and will weight it conservatively.
What has not changed
The rest of the diligence is the same as for any software business. Retention, concentration, contract quality, the standalone cost base, whether the engineering separates cleanly. An AI product with poor retention is a poor software business, and no amount of model sophistication compensates.
That is worth holding on to, because the temptation in preparation is to build the narrative around the technology. Buyers who pay the top of the band are paying for defensibility they can verify, not for the description.
Preparing for it
Audit the data rights before a buyer does. Document the model architecture and what in it is actually proprietary. Produce the inference cost curve. Write down the evaluation methodology, and if there is not one, build it, because it takes a quarter and it changes how the entire technical diligence reads.
None of that is work you would regret doing if you decided not to sell.
Divestitures.com Research · Published for orientation, not as advice on a specific transaction. Any figure cited is orientation, not a valuation. See market notes.