This is The Agentic Commercial Model, a newsletter about the operating model crisis hiding inside the AI gold rush.
Here's a scenario playing out in SaaS businesses right now, and almost nobody is naming it.
You ship a better version of your AI product. It's more efficient. It does the same job for the customer using fewer tokens, fewer calls, fewer credits. The customer is delighted. They get the same outcome for less money. By every measure that matters to them, you just delivered enormous value.
And in your revenue system, that customer just showed up as contraction.
Your best, most satisfied, most deeply embedded customers, the ones extracting the most value, can register as shrinking accounts. Your net revenue retention number ticks down. The metric you use to reassure the board that everything is working just flagged your healthiest relationships as decay.
This is the efficiency paradox, and it is the clearest proof that the metric the entire industry uses to measure customer health is no longer measuring what people think it measures. When OpenAI improved GPT-4's efficiency, many customers achieved the same outputs with fewer tokens, cutting their own costs. Tremendous value creation for the customer. On the revenue line, it looked like contraction. Traditional NRR would flag those customers as at-risk at the exact moment they became more satisfied and more sticky.
Read that again, because it inverts everything NRR is supposed to tell you. In a consumption and agentic world, the better your product gets, the worse your retention metric can look. You are being punished, on the dashboard, for delivering more value per unit of spend. And as agents get more autonomous and more efficient, this paradox doesn't fade. It intensifies. Customers accomplish more with less consumption, putting downward pressure on usage-based revenue even as their dependency on you deepens.
If your headline retention metric moves in the opposite direction to the value you're creating, it is not a health metric. It's a liar. Let me show you everything else it's hiding.
What NRR was built to measure, and why that world is gone
Net revenue retention had one job: tell you whether your existing customers are worth more this year than last, before you close a single new logo. The formula is simple: starting revenue, plus expansion, minus contraction, minus churn, over starting revenue. Above 100% means your base is compounding on its own. Below means you're on a treadmill, acquiring new customers just to stand still.
It was a genuinely good metric for the world it was built for. That world had three properties: revenue came from seats, expansion came from adding more seats, and consumption was irrelevant because a seat cost the same whether the person used it every day or never logged in at all. In that world, revenue tracked relationship. If an account was growing, the customer was getting more value. NRR was a clean proxy for health.
Every one of those three properties is now false.
Revenue no longer comes from seats; it comes from consumption and, increasingly, outcomes. Expansion no longer means more seats; it means more usage, which can rise or fall weekly. And consumption is no longer irrelevant; it is the entire basis of the revenue, which means a customer's spend now fluctuates with their efficiency, their optimisation, their macro budget pressure, and the quality of your own product.
When the ground beneath a metric shifts this completely, the metric doesn't gently lose accuracy. It starts actively lying. And NRR is now lying in at least four distinct ways.
Lie one: the average hides the decay
This is the oldest flaw in NRR, and the agentic era makes it lethal.
NRR is usually reported as a single blended number, and a single blended number is almost designed to deceive. Consider the anecdote currently doing the rounds because it's so recognisable: a RevOps lead presents 118% NRR to the board. Everyone celebrates. Then someone isolates a single enterprise account that expanded from $50k to $300k, strips it out, and the number collapses to 95%. The board stops celebrating. One account was masking the fact that the entire rest of the base was shrinking.
This is not an edge case. It is the structural reality of how blended NRR behaves: companies with strong NRR can fail to recognise they're losing a significant portion of their customer base, masked by the outsized growth of a few large accounts. A single blended figure hides more than it reveals. The right discipline is to segment by account size, report median alongside average, and track organic retention separately from the distortion of a few whales and from price increases.
In a seat world, this flaw was survivable because revenue was stable. In a consumption world, where the long tail of accounts can optimise their usage down quarter after quarter, a couple of expanding enterprise accounts can hold your headline NRR above 110% while the actual foundation of your business quietly erodes. The number on the board slide is green. The business is decaying underneath it.
Lie two: it counts expansion that destroys margin
Here is the lie that should worry PE operating partners most, because it goes straight to the economics.
NRR counts revenue. It says nothing about whether that revenue is profitable. In a seat world, that didn't matter much, because the marginal cost of another seat was near zero. In an agentic world, the marginal cost of more consumption is real, and often brutal. Consider a customer expanding from $10,000 to $15,000 in monthly revenue, a 50% expansion, while their infrastructure cost rises from $4,000 to $9,000. The contribution margin barely moves, from $6,000 to $7,200. NRR records a triumphant 150% expansion on that account. The actual economic gain is marginal.
This is not hypothetical. 84% of companies report more than 6% gross margin erosion from AI costs. Which means the expansion your NRR is celebrating is, in a meaningful share of cases, expansion you're barely getting paid for. You can hit a premium NRR number and a deteriorating margin profile at the same time. NRR will show you the first and conceal the second. For a value-creation thesis built on multiple expansion, that's not a blind spot. It's a landmine.
Lie three: it can't tell efficiency from abandonment
Back to the paradox we opened with, but now as a measurement failure rather than a curiosity.
In a consumption model, a customer's spend can fall for two opposite reasons. Either they've optimised, getting the same value more efficiently, in which case they're healthier and stickier than ever. Or they're disengaging, using the product less because they're getting less from it, in which case they're about to churn. These two customers are on opposite trajectories. One is your most defensible relationship. The other is a cancellation waiting to happen.
To NRR, they look identical. Both show up as contraction. The metric cannot distinguish the customer who is succeeding from the customer who is leaving, because it only sees the dollars, not the value behind them. In a seat world this distinction barely existed. In an agentic world it is the single most important thing you need to know about every account, and your headline retention metric is structurally incapable of telling you.
Lie four: it's a lagging metric in a real-time world
NRR is calculated over a trailing twelve-month window. It tells you what already happened. In a seat world, where relationships changed slowly and renewals were annual events, a lagging metric was fine. You had a year to react.
In a consumption and agentic world, value is delivered, measured, and realised continuously, and a customer's trajectory can change in weeks. By the time a deteriorating relationship shows up in your trailing NRR, the customer has often already mentally left. The metric tells you about the fire after the building has burned down. As one industry read of the 2026 data put it bluntly, gross retention is now holding steady while net retention declines, because companies are keeping customers but can no longer expand them the way they used to. NRR is detecting that shift a year late, when the operating model needed to detect it the week it started.
The fix: stop measuring billing, start measuring value realised
If NRR is the lead symptom, the disease is this: the entire SaaS measurement stack was built to track billing, and billing used to be a faithful proxy for value. It no longer is. The fix is not to abandon NRR. It's to demote it from a health metric to what it actually is now, a backward-looking financial summary, and to put a different metric at the centre of how you run the business.
At FR Advisory, the construct I work with portfolio businesses to implement is Value-Realised Retention (VRR).
The principle is a deliberate inversion of NRR. NRR measures the value you captured in billing. VRR measures the value the customer realised in outcomes, whether or not it showed up as more spend this quarter. It is built to answer the one question NRR cannot: is this customer actually getting more out of us over time?
VRR rests on three components, and the point of all three is to make value, not billing, the unit of measurement.
First, the outcome metric, not the consumption metric. Stop measuring tokens, calls, or credits consumed. Measure the unit of value the customer recognises: resolutions delivered, decisions automated, processes completed, hours returned. When a more efficient product lets a customer achieve the same outcomes for less spend, NRR records contraction; VRR records flat-to-growing value realised, which is the truth. The efficiency paradox simply dissolves, because you're measuring the outcome, not the meter.
Second, contribution-margin weighting, not raw revenue. Every unit of retained and expanded revenue is weighted by the margin it actually carries, so that margin-destroying expansion stops masquerading as healthy growth. This is the discipline of contribution-margin retention, and in an environment of 6%-plus AI-driven margin erosion it is no longer optional. It is the only way to know whether your expansion is creating enterprise value or quietly destroying it.
Third, usage depth as a leading signal, not a lagging one. Track the breadth and depth of how the customer uses you, across workflows, across teams, across use cases, as a real-time indicator of where the relationship is heading. This is the same signal that, as the special edition argued, is the moat being dug. Rising depth distinguishes the optimising customer from the abandoning one, the thing NRR structurally cannot do, and it does so in weeks rather than in a trailing-twelve-month rear-view mirror.
VRR is harder to compute than NRR. It requires the measurement infrastructure to track outcomes, the cost data to weight by margin, and the product instrumentation to see usage depth in real time. That difficulty is precisely the point. The companies that build this capability will see decay a year before it reaches their NRR, and they will see genuine health that their NRR is busy hiding. The ones that don't will keep reading a green number off a board slide until the renewals don't come.
What this is really hiding, and who it's hiding it from
Let me be direct about who should be most uncomfortable here, because it isn't the RevOps analyst who built the slide.
It's the board, and the PE operating partners reading that slide and feeling reassured.
If you are sitting on a software board right now taking comfort from a 115% NRR, you need to sit with the possibility that the number is doing the opposite of its job. It may be concealing a base that is decaying everywhere except your largest accounts. It may be celebrating expansion that your margins never actually saw. It may be unable to tell you whether your customers are succeeding or quietly leaving. And it may be reporting all of this on a twelve-month delay, in a market that now moves in weeks.
NRR is the metric the industry reaches for to prove the agentic transition is working. In far too many businesses, it is the precise instrument by which leadership is avoiding the truth that it isn't. A reassuring retention number is the most comfortable place in the world to hide from a structural problem, and right now it is exactly where a lot of boards are hiding. The number going up is not the same as the business getting healthier. In the agentic era, those two things have come apart, and NRR is built to hide the gap.
The uncomfortable question for every board in software right now is not "what's our NRR?" It's "what is our NRR preventing us from seeing?"
Next issue: The GTM Collapse — when your buyer is an algorithm, your sales motion is obsolete. We've spent four issues on how you price, pay, and measure. Issue 05 turns to the motion itself, and to what happens to a go-to-market machine built for human buyers when the entity evaluating, purchasing, and even deploying your product is increasingly an agent acting on the customer's behalf.
First published in the Agentic Commercial Model newsletter on LinkedIn, June 26, 2026. Read the original on LinkedIn.