What is an AI business worth?

Valuing an AI business starts by setting aside the label. 'AI business' describes a huge range of things — a company with genuinely proprietary data and a trained model, a company that fine-tuned someone else's foundation model, and a company that's really a well-designed interface over a third-party API — and these are worth very different multiples of similar revenue, because the ownership underneath each is different. Before any earnings multiple gets applied, a buyer needs to know which of these they're actually looking at. That ownership question shapes not just the multiple applied but whether a buyer's lender will finance the purchase at all.

Ownership is the first valuation question

A business built on data it collected and owns, powering a model it trained and controls, has a defensible position a competitor can’t simply replicate by calling the same API — that defensibility is what a buyer is actually paying for in a higher multiple. A business that’s a well-built product wrapped around a third party’s model, however good the product, doesn’t have that same protection: the underlying provider can change pricing, change terms, or launch a competing feature, and the seller’s moat depends entirely on execution and customer relationships rather than on anything the buyer can’t get elsewhere. Both can be legitimately valuable businesses, but they’re priced on different logic. A useful test is to ask what happens to the business the day a key API provider changes its pricing or terms — a proprietary-data business barely notices, while a thin wrapper can see its margin disappear overnight.

Recurring revenue and customer stickiness

As with any small business, recurring revenue is worth more than one-time project or implementation fees, and a buyer will look hard at how much of an AI business’s revenue is contracted and repeating versus custom work billed once. For AI products specifically, buyers also probe how sticky customers actually are — whether switching to a competing tool, including a generic AI tool, would be easy or costly for the customer. A product deeply integrated into a customer’s workflow and data is worth more than one that’s easily substituted, even at identical revenue. Contract length matters too — a business with multi-year agreements is priced differently than one running entirely on month-to-month arrangements a customer could cancel with little notice.

What normalized earnings look like here

The same add-back exercise that applies to any small business — normalizing for owner compensation, one-time costs and non-recurring items — applies to an AI business too, but with a few sector-specific wrinkles. Cloud compute and API costs, for instance, are a real, recurring cost of running the product and shouldn’t be added back the way a genuinely one-time expense would be; a buyer inheriting the business inherits that cost structure. Founder time spent on model training or fine-tuning is often deeply embedded in day-to-day operations in a way that’s harder to normalize away than a typical owner-dependence adjustment.

Why comparables are thin right now

Because AI-specific business sales are still relatively uncommon, a seller or buyer won’t find the depth of comparable transaction data that exists for, say, a restaurant or a landscaping business. That means valuation here leans more heavily on first-principles analysis — provable earnings, defensible IP, customer retention — than on a stack of recent comparable sales. It’s honest to say this market is young, and a valuation built with humility about that fact is more credible than one that pretends the comparables exist when they don’t. Buyers and sellers who go in expecting a precise, market-tested multiple the way they might for a franchise resale are usually disappointed; a defensible range built from first principles is the more realistic goal.

Red flags that lower the number

• Training data with unclear or undocumented provenance, which raises the buyer’s legal risk regardless of current performance.
• Revenue concentrated in one or two large customers who could leave and take most of the business with them.
• A model or product that depends entirely on a single third-party API with no differentiated data or workflow underneath it.
• Missing IP assignments from contractors who built core parts of the model or code.
• Heavy founder dependence on the technical side, where the business can’t function without one specific person’s involvement.
• A product architecture that’s difficult for a new technical owner to understand or maintain without extensive documentation or the departing founder’s help.

What actually supports a strong number

The businesses that command the strongest multiples in this category combine three things: clean, documented ownership of data and code; recurring revenue from customers who’d find it genuinely costly to switch; and a technical team or process that doesn’t depend entirely on one founder. None of that is unique to AI businesses, but the ownership question is sharper here than almost anywhere else, because so much of the value is intangible and so much of the risk is in what can’t easily be seen without a proper diligence review.