Separating two markets
Figures quoted for artificial intelligence businesses are frequently drawn from a different market than the one a middle-market seller is in.
Venture rounds for frontier model developers, and public multiples for a handful of listed companies, produce numbers that have nothing to do with what an acquirer pays for a profitable applied AI business with twenty million of revenue. Quoting the first at a seller in the second sets an expectation that no bidder will meet.
For middle-market technology assets, AI and machine learning platforms can command the highest revenue multiples of any software vertical, and the spread inside the category is the interesting part. What decides where an asset lands inside it is whether the buyer is acquiring something they could not otherwise build.
What actually carries value
Data that cannot be replicated. Models can be retrained and architectures can be copied. A dataset assembled over years from a position competitors do not have is the one asset that transfers and cannot be recreated. This is the single largest determinant of where an asset sits in the band.
Demonstrated performance, measured rather than asserted. An evaluation regime with a held-out set, run regularly, with results tracked over time. A business that cannot show its quality has improved is asking the buyer to take retention data as the only evidence.
Revenue from customers who renew. Pilots, proofs of concept and discounted first-year deals are not the same thing, and buyers now separate them explicitly in diligence.
A team that will stay, which in this field means checking who has already left as much as who remains.
Cost structure that improves with scale. Inference cost that rises linearly with usage is a margin problem rather than a technology problem, and it is priced as one.
What buyers discount
Dependency on a single third-party model provider, with no tested alternative. This is supplier concentration, and it is priced the way supplier concentration is always priced.
Training data with unclear provenance. If the rights cannot be evidenced, the buyer has to assume they may have to stop using it, and a model that cannot be retrained on the same basis is worth materially less.
Products where the AI component is a feature of a conventional software business rather than the thing customers buy. These are valued as conventional software, which is a lower band, and describing them otherwise in a memorandum damages credibility on everything else.
Heavy usage concentrated in a few accounts that are unprofitable to serve. Buyers now ask what the top decile of usage costs, and a seller without the number is presenting an unquantified liability.
How the methods apply
Revenue multiples remain the primary approach, with the definition of revenue mattering more than usual. Recurring platform revenue and one-off implementation work are not the same and should not be presented together.
Discounted cash flow is difficult here because the cost curve is truly uncertain. Inference costs have fallen substantially and may continue to, or may not. A model that assumes either is making the assumption that determines the answer.
Comparable transactions are thin at this size and frequently undisclosed, which means the range around any comparable is wide. Treat a single reference point with appropriate caution.
Where a business is being bought principally for its team, the valuation logic is different again and closer to a recruitment calculation than an enterprise valuation. Sellers should know which conversation they are in, because the two produce very different structures.
Preparing an AI business for sale
Audit the data rights before a buyer does, and be able to produce the chain of rights for every material dataset.
Document what is proprietary in the model work and what is configuration of something publicly available. Buyers will establish this in diligence, and a seller who has drawn the line honestly is more credible than one who has not.
Produce the inference cost curve, including the expensive tail.
Build an evaluation regime if there is not one. It takes a quarter and it changes how the whole technical diligence reads.
Separate recurring revenue from services in the reporting, and do it several quarters before going to market so the history is consistent.
The honest summary
The premium for AI capability is real and it is narrower than it was. It attaches to businesses where a buyer is acquiring data, evaluated performance and paying customers that they could not assemble themselves.
Everything else is software, valued as software.
Editorial Team · Published for orientation, not as advice on a specific transaction. Any figure cited is orientation, not a valuation. See market notes.