Khaled Azar 
Every company that wants to buy your AI business asks the same question first. What happens to your revenue if the AI model your product runs on changes? In June 2026, OpenAI told developers it will shut off the first version of GPT-5 on December 11, 2026. That version came out in August 2025, only 16 months earlier.
Your answer changes the price. The buyer splits your revenue into 2 parts. The first part would keep coming in after a model change, and the buyer pays the most for it. For the second part, the buyer pays less. Or the buyer holds back some of the price and pays it later, only if that revenue keeps coming in.
By the end of this post you can run the buyer’s test yourself. It sorts your revenue into those 2 parts. You will also know how to fix the weaker part before a buyer sees it.
How to sell an AI SaaS: run the Swap Test before the buyer does
A buyer runs 3 swaps on your business in their head. Run them first.
Start with why buyers do this. They pay today for revenue they expect to collect for years. In an AI company, 3 things that produce that revenue sit with someone else: the price of the model, the model itself, and whether your feature stays yours.
Any of them can move without warning. So a careful buyer swaps each one out and watches what happens to your revenue. I call this the Swap Test.
| The swap | What the buyer asks | What proves you pass |
|---|---|---|
| Price swap | If the model cost twice as much next year, would you still make money on every customer? | Your margin worked out at today’s model price and at double that price |
| Model swap | If your model were retired, could you move to another one without customers noticing? | A saved set of real customer tasks, run on a second model, with the results kept |
| Feature swap | If the model maker or a big competitor shipped your feature for free, would customers stay? | Renewals and cancellations split by price level and plan length |
Revenue that passes all 3 is the revenue a buyer pays full price for. Everything else about the sale follows the normal steps for selling a SaaS company. This test is the part that is new.
Swap one: the price of the model
Gross margin is simply what you keep from each dollar of revenue after paying what it costs to deliver the product. In normal software that cost is small. In AI software, every time a customer uses the product, you pay the model maker a fee for the answer.
That fee is why AI margins look different. In ICONIQ’s 2026 State of AI report, a survey of more than 300 software executives, gross margins on AI products were 45% in 2025. The same companies expect 53% for 2026. About 2 in 3 of them say each request got cheaper to serve, partly by sending easy requests to cheaper models and saving the expensive ones for hard work.
To a buyer, a margin those companies expect to swing 8 points in one year can swing the other way too. If your margin depends on a price you do not set, the buyer prices the chance that it moves the wrong way. I cover the full rebuild in how buyers rework your gross margin around what each AI answer costs you. Here is the short version to have ready.
- Work out your margin by customer group at today’s model price.
- Work it out again at double that price. I use double because it is a move a buyer can picture, not a forecast.
- If any group loses money at double the price, fix that group’s pricing before you go to market.
- Show what share of your requests already go to cheaper models.
Swap two: the model itself
Go back to that OpenAI notice. OpenAI’s own policy promises at least 6 months of notice before it retires one of its standard models. Special versions of those models get at least 3 months.
A 6-month warning sounds like plenty until a buyer finds that date during diligence, which is simply the weeks a buyer spends checking everything you have told them. Now your sale has a deadline inside it that neither of you controls. The buyer will price the move to a new model, plus the risk that your product gets worse after the move.
Buyers treat this like any single supplier: if one vendor can shut off a part you depend on, every forecast you show depends on that vendor. ICONIQ’s report says running several models is now the default, with the companies it surveyed using about 3.3 AI models on average. A product that depends on exactly one model stands out, and not in a way that helps you.
The fix is a test you can rerun on any model. The numbers below are my working rules, not an industry standard. Buyers who probe how much AI risk sits inside your product ask for exactly this.
- Save 200 to 500 real customer tasks, each with the answer a good response should give.
- Run them on your current model and on at least one other provider. Keep both sets of results.
- If your model already has a shutdown date, finish the move before you go to market.
Swap three: your feature, shipped by someone else
This is the swap founders underestimate. Model makers and large software companies add features every month. If your product is a thin layer on top of a model, a customer can drop you the day a tool they already pay for ships the same thing.
2 terms help here. Gross revenue retention is the share of last year’s revenue you kept, not counting upsells. Net revenue retention is the same thing with upsells added back in, so it can go above 100%.
Net revenue retention for AI products priced above $250 a month, compared with AI products under $50 a month, in ChartMogul’s study of roughly 200 AI companies.
Those figures come from ChartMogul’s retention report on AI companies. The cheapest AI products kept only 23% of their revenue before upsells. The report calls it the curse of the AI wrapper: easy to buy means easy to cancel. The same report found that annual plans run 10 to 20 points higher on net revenue retention than monthly ones.
The extreme case shows why. A founder asked me this month how to sell an AI model with no revenue at all. With nothing renewing, a buyer has only one way to price it: what it would cost them to build it themselves.
Revenue that renews is the proof a buyer trusts. How buyers separate real AI revenue from a feature customers expect for free goes deeper on this. The prep list is short.
- Split renewals and cancellations by price level and by monthly versus annual plan.
- Move your best monthly customers to annual plans before you go to market.
- Show the work customers keep inside your product: their data, their saved workflows, the steps their team runs every day.
What the Swap Test does to your price
Here is how the test turns into dollars. This is an illustration, and the multiples in it are assumptions for the math, not market quotes. Say your AI SaaS brings in $2,000,000 a year in recurring revenue. A buyer would pay 4 times revenue for every dollar that passes all 3 swaps.
Now say $1,200,000 of it passes. The other $800,000 comes from customers on monthly plans who have never renewed through a competitor launch. The buyer values that weak pile at 2 times revenue, and only pays for it through an earn-out, which is a slice of the price paid later only if that revenue holds up.
| Step | Amount | Running total |
|---|---|---|
| All $2,000,000 valued at 4 times | $8,000,000 | $8,000,000 |
| Weak $800,000 valued at 2 times instead of 4 | -$1,600,000 | $6,400,000 |
| Headline price | $6,400,000 | |
| Weak pile’s value, paid later through the earn-out | -$1,600,000 | $4,800,000 |
| Cash at close | $4,800,000 |
Now start 12 months earlier. Say $400,000 of the weak pile comes from larger customers who just happen to pay monthly. Move them to annual plans a full year before you sell, so they renew once before a buyer shows up.
The passing pile grows to $1,600,000. At 4 times, that pays $6,400,000 at close. The remaining $400,000 of weak revenue becomes an $800,000 earn-out.
In this example, getting $400,000 of revenue through one annual renewal before the sale raised cash at close from $4,800,000 to $6,400,000. The buyer did not pay more for AI. They paid more for revenue that proved it would stay.
The earn-out is not free money either. How earn-outs work and when to walk away covers what can go wrong with it once the buyer controls the business.
When to sell an AI SaaS if the prices move every month
The prices buyers pay for software can move a lot inside one year. I spoke with the CEO of a profitable, bootstrapped software company that ran a sale this year. At the start of the year his bankers expected 5 to 6 times revenue. By the time his company was for sale, the best number he heard from peers with similar businesses was about 3.5 times.
He put the change down to the newest, most powerful AI models coming out. Then came the part that matters for timing: the buyers who had already bid would not move up, even as his revenue kept growing. They stayed anchored on the first number they offered. He pulled the sale.
You cannot predict a market that moves like that. You can decide when your own numbers are ready. These are the rules I give founders.
- Do not test the market early. Buyers anchor on their first number, and you do not get a second first bid.
- Go to market after your AI customers have been through at least one full renewal. 12 months of retention data beats any headline price.
- If a model you depend on has a shutdown date, finish the move first.
- If a large platform just announced your feature, wait until you can show one quarter of renewals after the announcement.
- Bring several buyers to the table at once. Running a competitive sale process is how you stop one buyer’s view of AI risk from setting your price.
Frequently Asked Questions
How do you sell an AI SaaS company?
You follow the normal SaaS sale steps, then add one: prove which part of your revenue survives a change in model price, a model retirement, and a competitor shipping your feature. Buyers pay full price for revenue that passes all 3. The rest gets discounted or moved into an earn-out.
Do buyers pay more for AI companies?
They pay more for AI revenue that has proven it will stay, not for the AI label. In ChartMogul’s data, AI products priced above $250 a month keep 85% net revenue retention, close to normal business software. Buyers pay for that kind of revenue.
Can you sell an AI wrapper business?
Yes, but a thin layer on top of someone else’s model is the hardest AI business to sell at a good price, because customers leave easily. The cheapest AI products in ChartMogul’s data kept only 23% of revenue before upsells. Annual plans and deeper workflows are the main fix.
What do buyers check in AI SaaS due diligence?
They check your margin at different model prices, how quickly you could move to another model, and whether customers stay when bigger players ship similar features. They also look for shutdown dates on the models you use. OpenAI, for example, promises at least 6 months of notice before retiring one of its standard models.
When is the right time to sell an AI SaaS?
After your AI customers have renewed at least once, so you can show 12 months of retention data. Finish any required model migration before you go to market, and avoid launching a sale the same quarter a large platform announces your feature.
Next Steps
Not sure how much of your AI revenue would pass the Swap Test? A value assessment sorts your revenue into the 2 piles a buyer will see, before a buyer sees them.