AI revenue is not SaaS revenue

The deck arrived on a Sunday, which I respected. Series A, applied AI company, a vertical I know well. Revenue tripling year over year, logos I recognized, and on slide six the ARR chart with exactly the shape everyone in this business has been trained to admire.

On our first call I asked the founder one question. Take your ten oldest customers and tell me what it cost to serve them last month. Actuals, fully loaded. Not the plan, not the target. The bill.

He quoted me the blended gross margin from the financial model. Seventy-eight percent. I asked again, actuals this time. It took his team nine days to produce the number, which was itself an answer. When it came back it was 34 percent, and it was drifting down, because his happiest customers were also his heaviest users.

The gap between the margin in the model and the margin in the bill is where I now spend most of my diligence time on AI companies. The ARR slide looks like software. The cost of serving it frequently does not.

The habit we imported

Everyone evaluating AI companies today, me included, built their reflexes on SaaS. In that world, ARR was almost a synonym for quality. Marginal cost was close to zero, gross margins sat near eighty percent by construction, and once a customer was signed, nearly every additional euro of revenue fell through to fund growth. Every ARR multiple anyone has ever paid assumes that machinery is running underneath the number.

AI-native revenue often does not have that machinery. The number on the slide is called ARR, it recurs, the logos are real, and underneath it the economics can be something else entirely. Not fraudulent, not even hidden, just different, and different in ways that the standard diligence checklist was never designed to catch.

Three differences do most of the damage.

Where the margin actually goes

Success has a marginal cost. Inference is cost of goods sold, and it scales with usage, not with seats. Most AI products still price per seat because that is what buyers understand. Put those two facts together and you get an uncomfortable inversion: the customer who loves the product, runs it all day, and renews without a call is also the customer who costs the most to serve. In classic SaaS your best customers subsidize your worst. In a lot of AI companies it is the other way around, and growth makes the problem bigger, not smaller.

There are people inside the margin. Almost every AI company I look at has humans propping up output quality somewhere. A review team checking results before they reach the customer. Solutions engineers quietly rewriting prompts per account. An escalation desk for the cases the model fumbles. On the P&L these people live in R&D or onboarding, and in the pitch they are temporary. My test is simple: if the product quality the customer experiences depends on those people being there, they are cost of goods, whatever the accounting says. A services firm wearing a software margin is not a scandal, but it should be priced like what it is, and it almost never is.

"Models will get cheaper" does not cash the way founders assume. This is the standard defense, and it fails in three separate ways. First, consumption grows faster than prices fall. Cheaper capability per unit gets spent immediately on more units: longer context, more retries, agent chains where a single call used to be. My own stack's monthly bill has gone up over a period in which the per-unit price of the models inside it collapsed. Second, in any competitive category the savings get passed through. When everyone's cost of goods drops at once, price follows it down, and the surplus lands with the customer, not with you. Third, every time the frontier moves, the quality bar moves with it, and you spend the savings keeping up. Falling model prices are real. Treating them as a margin plan is not.

What I actually check

Four things, in order, and none of them are on slide six.

Cohort margin, in actuals, over time. The ten oldest customers, cost to serve, month by month, for as long as the data exists. The level matters less than the direction. A margin that improves with customer age tells me the system is learning: caching is working, requests are being routed to cheaper paths, humans are coming out of the loop. A margin that degrades with customer age tells me success is expensive, and scaling it will require money forever.

Cost per unit of work. Not cost per seat, cost per completed task: one document processed, one ticket resolved, one report produced, whatever the product's atomic unit is. What does it cost to deliver, and what does the customer pay for it. Founders who know this number cold are running a business. Founders who need a week to compute it are running a model with revenue attached.

The human-to-throughput ratio. How many people touch the output per thousand tasks, and which direction that ratio moved over the last two quarters as volume grew. Flat or rising means the humans are structural. Falling means the automation is real.

Where the last price drop went. I ask directly: the last time your underlying model got meaningfully cheaper, what happened to your gross margin. The honest answers are wonderful and varied. Almost none of them are "it went to margin."

The good version exists

I want to be clear that this is not a case against the category. The best AI businesses I have seen have margin curves that no SaaS company could produce, because the product genuinely improves with use. Escalations fall as the system accumulates corrections. Expensive calls get replaced by cheap ones as the team learns which paths need the frontier model and which need something smaller. Cost per task falls every quarter while price holds, and the margin expands with scale for structural reasons rather than promised ones.

When I find that curve in the actuals, I stop caring that the current margin is mediocre. A 40 percent margin heading north on a mechanism I can inspect is worth more to me than a 75 percent margin that only exists in the model. Software economics in this category are earned over time, not assumed at signing, and the earning shows up in exactly one place: the cohort data.

What I told him

I did not pass on the Sunday deck, which may be the most important part of this essay. A bad margin at Series A is usually a pricing problem and an engineering problem, and both are fixable by a good team that can see the number clearly. The dangerous thing was never the 34 percent. The dangerous thing was the nine days it took to find it.

So I told him what I would tell any founder in this position. Instrument cost per task this month, move pricing toward the unit of value rather than the seat, and come back in two quarters with cohort actuals showing the direction of travel. If the curve bends the right way, the check is easy, and the next investor's diligence becomes a formality instead of a fight.

ARR tells me that customers want the product. The margin, in actuals, over time, tells me whether a company is being built underneath them or a subsidy is being distributed. I am happy to fund the first. The second is a gift to customers, funded by whoever read slide six and stopped there.