The NRR is real. The bill just has not arrived.
Borrowed NRR is a retention number that climbs while the economics underneath it weaken. The headline holds because expansion and price cover the cracks. Value, cost to serve, and proof were never measured. So the number runs on credit.
NRR is a board metric, so it gets defended. Expansion and price increases can lift it. Gross margin can slip at the same time, and renewals get harder to win. Nobody set out to borrow. No one built the view that tells durable retention from propped-up retention.
NRR and gross margin are meant to move together. When they split, the retention number is borrowing against future margin. You can hold NRR for a while. Spend more to serve, discount to keep accounts. The board sees a strong number. The economics tell a different story. It compounds.
The number gets measured. Value delivered, cost to serve, and proof become visible together. NRR stops hiding what it costs to hold. The goal is retention that pays for itself rather than retention bought with margin. A human still makes the call. The product shows the true cost of holding the account.
NRR tells you customers are paying more. It does not tell you they are getting more, that the revenue is profitable, or that it will still be there next year.
Those things used to move together in SaaS. When a customer expanded, it was usually because they got more value, and the margin structure held. AI breaks that link. An AI premium can raise NRR this quarter while the value delivered stays flat, gross margin drops underneath it, and renewal risk builds for next year.
A board reads NRR as proof of expansion quality. It may only be proof of pricing power. Those are not the same thing, and the difference is the entire risk.
Healthy AI expansion has to pass three tests. Run them on your own book. Each one has a clear pass, and each one fails inside the operating model, not the price.
The Value Test. When it passes, the customer can point to a real outcome and so can you: a number that moved, work that went away, a decision that got faster.
When it fails, you cannot say, because you never tracked usage. The AI sits in the base SKU, so no one knows who uses it. Value gets claimed in a QBR. It never gets proven. This is Hollow Usage: a feature that shows up in the package but not in the customer's outcomes.
The Cost Test. When it passes, you know the gross margin of the AI revenue and you priced the premium with that cost in view.
When it fails, AI COGS hides inside a general hosting line and you are flying blind. AI has real usage cost that classic SaaS did not. Bessemer pegs AI gross margins near 50 to 60 percent, against 80 to 90 for SaaS. And the cost is climbing. Gartner expects inference cost to rise, because token demand grows faster than token prices fall, and an agentic task can burn 5 to 30 times the tokens of a simple chatbot. Left untracked, that gap eats the expansion you just booked.
The Durability Test. When it passes, the customer walks into the renewal already convinced, because the value showed itself all year.
When it fails, you have no record of outcome and the renewal becomes a fight. The first premium comes from packaging power. The second has to survive proof. The odds are bad. MIT found 95 percent of companies get no measurable return on AI. If your customer is one of them, you reach the renewal with nothing to show.
Fail one test, and the NRR is borrowed.
Not on its own. NRR rising while gross margin falls is a warning, not a win. The pair matters more than the single number.
Track it against gross margin over several quarters. If retention holds while margin erodes, the number is resting on economics that are getting weaker.
Borrowed NRR is AI expansion revenue you book now that the renewal cannot sustain. Net revenue retention rises because the price went up, but the value was never proven, the cost to serve was never measured, and the premium was never built to survive the next renewal. It is real revenue with unproven quality, and the gap shows up when the renewal comes due.
No. AI can create real value, and when it does you should charge for it. The problem is booking the premium as expansion before you have proven the customer got value, priced the cost to serve it, and built a reason the premium survives the renewal. Charging for AI is fine. Counting undefended AI revenue as durable growth is the mistake.
Because NRR measures the price the customer pays, not the value they receive or the cost to serve it. AI premiums can lift retention this quarter while AI COGS — which runs far higher than classic SaaS at roughly 50 to 60 percent gross margin versus 80 to 90 — drags margin down underneath the number. The two metrics split, and only one of them is being reported as growth.
Run the three-question test on your own book. The Value Test: can the customer and you both point to a real outcome? The Cost Test: do you know the gross margin of the AI revenue? The Durability Test: will the premium survive the next renewal on proof, not packaging power? Fail any one, and the NRR is borrowed.
Seven questions. Five minutes. A pattern read on the spot, no call to see it.