Season 3 · Episode 5

Salesforce Sells It By the Action. It Reports It As Recurring Revenue.

Three of the companies furthest into the crossing tell the market about it in instruments built for the thing they are leaving. Salesforce prices Agentforce at ten cents an action and reports it as annual recurring revenue. HubSpot bills credits, knows its resolution counts, and publishes neither. ServiceNow reported a 98% renewal rate for the quarter the market answered with the worst day in its history. One company built its own instruments and gets priced on its own arithmetic. Here is what it published, and what the crossing version has to contain.

The Agentic Commercial Model Newsletter · Fessal Rahman · November 20, 2026 · 12 min read

From FR Advisory.

The author spent twenty years running commercial organisations inside software and private equity portfolios. He no longer works in the technology industry, holds no position in any company named in this series, has no clients in it, and is selling nothing to anyone in it.

Start with the seller, because the seller is the one holding the bag.

An enterprise account executive closes a deal in which the customer moves four hundred thousand pounds of committed seat spend onto a meter. The customer is delighted. They now pay for what the agents actually do, the finance director has stopped treating the renewal as a hostage negotiation, and the relationship has eighteen months of runway it did not have in March.

Inside the vendor, that deal lands as a reduction. Committed contract value falls by four hundred thousand. The retention metric takes the hit this quarter. The consumption it produces arrives over the following year, in a different line, with no history to compare it against, attached to no quota and visible in no board pack.

The seller is not confused about what happened. Neither is the customer. The only parties who cannot see it are the compensation plan and the market.

Episode Three ended on a single instruction: build the second number yourself, because nobody is coming to invent consumption-era disclosure for you. This is the construction. It is also, unavoidably, an audit of what the companies in front of you have chosen not to publish.

What three crossing companies actually tell the market

This is not a survey of opinion. These are the disclosed key business metrics, from the filings, for the most recent quarters available.

Salesforce. Its own Flex Credits rate card, updated 17 June 2026, says: "Customer consumes Flex Credits at the applicable multiplier rate based on the Usage Type identified below." A standard Agentforce action costs twenty credits. Salesforce's own announcement of the model puts it plainly: "Each of these actions consumes just 20 Flex Credits ($0.10 per action)." Credits sell in packs of a hundred thousand for five hundred dollars. Salesforce describes this, in its own words, as a "consumption-based model that aligns cost with business outcomes."

Then it reports the result to the market: "Agentforce ARR exceeded $1.5 billion, up over 240% Y/Y."

Annual recurring revenue. For a product the company itself sells by the action.

It gets worse on inspection. In the same release: "Effective Q2 FY27, Agentforce ARR includes our AI offerings, Slackbot and Headless 360." The definition widened in the quarter that printed 240% growth. Whatever that figure measures, it does not measure the same thing it measured a year earlier, and the comparison is doing work it cannot support.

HubSpot. Three key business metrics as of 30 September 2026. Customers, defined as "the number of business entities with one or more paid subscriptions to our customer platform," at 311,704. Average Subscription Revenue per Customer, at $12,038 for the quarter. Net Revenue Retention, calculated on Contractual Monthly Subscription Revenue.

Every instrument is a subscription instrument, at a company that repriced its Customer Agent to fifty cents a resolution, published the rate card, and let customers turn committed seats into credits.

The filing is not silent on the meter. It discusses credits consumed by agent use and usage fees deferred until consumption, because the accounting has to handle it. What it does not disclose is a single credit volume, a single resolution count, or any measure of customers converting from seats to usage. In August the Chief Financial Officer explained part of the retention decline by pointing at customers swapping seats for credits. That swap has no published number. The company named the cause and withheld the measurement.

Two of those three definitions also moved during 2026. From the first quarter, Net Revenue Retention began including Solutions Partner commissions and using fixed exchange rates refreshed annually. From the third, the Customers count began including paid subscribers to YouSpot, an early product out of HubSpot's internal incubation group aimed at one-person businesses, whose size the company does not disclose.

Neither change is an inflation scandal, and the YouSpot addition is almost certainly immaterial against a count of 311,704. The point is narrower and worse. The retention metric at the centre of this entire story had its calculation altered two quarters before it fell to 102% and took a fifth of the company with it. When the instrument moves and the thing it measures is also moving, nobody outside the company can separate the two.

ServiceNow. For the quarter ended 30 September 2026: remaining performance obligations of $29.7bn, up 22%. Current RPO, defined in the filing as "RPO that will be recognized as revenue in the next 12 months," at 46% of the total and up 20%. Customers above $5m of annual contract value, 680 against 553. And a renewal rate of 98%.

That is the disclosure for the quarter ServiceNow reported on 28 October 2026. The following day the stock fell 27%, the worst single day in its history, after the company guided fourth quarter subscription revenue growth of 21.5% in constant currency against contracted backlog growth of 18.5%.

Same quarter. The renewal rate says 98%. The market said the worst day on record. Both were correct, because the renewal rate measures whether contracts came back, and the market was pricing what those contracts will be worth once they are meters rather than commitments. A renewal rate cannot tell you that a renewed customer has renewed onto a smaller floor.

No consumption, credit, or usage metric appears in any of it.

The one that built its own instruments

Snowflake has never had seats to lose, and it could not have used the standard set if it wanted to.

Its own annual report is blunt about the consequence: the company "generally recognize[s] revenue on consumption," and "because our customers have flexibility in the timing of their consumption, we do not have the visibility into the timing of revenue recognition" that a subscription business has.

That is an admission of forecasting weakness, published voluntarily, by a company that then went and built the instruments to compensate.

For the quarter ended 31 July 2026 it disclosed product revenue of $1,491.9m, up 37%, broken out separately from total revenue of $1,546.8m, up 35%. Net revenue retention of 126%, calculated on consumption rather than contract. Remaining performance obligations of $9.00bn, up 30%. And 828 customers producing more than a million dollars of trailing twelve-month product revenue, up 27%.

Four instruments. Not one of them is a seat. Revenue split so that consumption can be read on its own. Retention redefined around usage. A commitment measure retained, because commitments still exist in a consumption business and pretending otherwise is its own dishonesty. And a concentration measure anchored to what customers actually consumed over twelve months rather than what they agreed to buy.

A business growing thirty-seven per cent on pure consumption, priced by the market without drama, because it handed over a scoreboard the market could read.

The instruments are not missing. The will to publish them is.

Why this is a choice, not a limitation

The defence offered for the subscription scoreboard is that consumption data is immature, unaudited, hard to define, and would confuse investors.

Examine that against what these companies already do.

HubSpot bills credits. It cannot bill them without counting them. It repriced the Customer Agent at fifty credits per resolution and published that rate card publicly, which means it knows, to the credit, how many resolutions it delivered last quarter and what each one earned. Salesforce operates a published multiplier table covering standard actions at twenty credits, voice actions at thirty, and prompt tiers from two to sixteen. You cannot run that table without a per-action ledger.

The data exists. It is operational, it is billed against, and it is precise enough to invoice enterprise customers who dispute invoices for a living.

What does not exist is the decision to publish it.

That decision has a cost, and the cost is not theoretical. HubSpot lost 19% on the day it announced the meter and 19% again on a quarter where it beat revenue, operating income and adjusted earnings, with the fall on 6 August the steepest single day in its history on the exchange. In both cases the market was reading subscription instruments describing a business that had stopped being one. The company supplied no alternative reading. The market used what it had.

There is a version of this that is simply prudence. There is another version in which a company would rather be mispriced than publish a realised rate that shows what its agents actually complete. That second version is the one every board should be asking about, and almost none are.

What the crossing instrument has to contain

This is the part an operator can take and build. Six numbers. They are not a replacement scoreboard, they are the bridge, published alongside the standard set for as long as the crossing takes.

One. Converted spend. Committed contract value that customers moved from fixed to variable in the period, stated separately, with the prior-period base it came out of. This is the single most important number and the one nobody publishes. Without it, every conversion is indistinguishable from a contraction, and your retention metric is reporting your best commercial outcome as damage. State it gross, not netted against anything.

Two. Retention on two bases. Publish the standard calculation exactly as defined, and publish the same calculation with converted spend treated as retained. Then publish the gap between them. The gap is the transition, measured. Do not replace the standard figure, because replacing it reads as evasion and invites the assumption that the real number is worse. Show both and let the gap tell the story. HubSpot's retention fell a point to 102% for a reason its own CFO described out loud; a two-basis disclosure would have turned that quarter from an unexplained decline into a demonstrated conversion.

Three. Consumption volume and consumption revenue, separately. Credits, resolutions, actions, tokens, whatever your meter counts. Volume as its own line with its own history, revenue as another. Published separately so that price and volume can be told apart, because a business whose volume is compounding while its realised price falls is in a completely different position from one where both are flat, and a single blended revenue number hides which one you are. Snowflake does the structural version of this by breaking product revenue out from total revenue. Do it at the unit level.

Four. Realised rate. What a unit actually earned, net of what you did not charge for. If you price per resolution, publish resolutions attempted and resolutions billed. If you price per action, publish actions executed and actions charged.

This is the number that frightens people, and it is worth more than the other five combined. A vendor charging for results is making a claim about its product working. The realised rate is the evidence. Publish it and every conversation with a sceptical buyer changes, because you have put the thing they doubt into a filing. Withhold it and you are asking the market to take outcome pricing on trust from an industry that has never earned it.

Five. Consumption concentration. Customers above a spend threshold, measured on trailing twelve-month consumption revenue rather than contract value. Snowflake's 828 customers above a million dollars of trailing twelve-month product revenue is precisely this instrument, and it works because it cannot be gamed by signing a large commitment that never gets used. Contract-value concentration measures what people promised. Consumption concentration measures what happened.

Six. The cohort curve. For customers who converted, consumption revenue in each of the four quarters after conversion, indexed to the committed spend they gave up.

This is the only number that answers the question the market is actually asking, which is whether a converted customer ends up worth more than an unconverted one. It appears in none of the filings examined for this piece. It is also, for most vendors, eighteen months of history away, which is the honest reason to start publishing the other five now: the cohort curve is worthless without the cohorts, and the cohorts start the day you begin counting.

Whoever publishes a credible version of this first will define it for everyone who follows, and will be the only company in the sector whose crossing can be read rather than guessed at.

What this does to the sales floor

A disclosure set is not a finance exercise. It changes what a seller is paid for within a quarter of being published, and if it does not, it was decoration.

The seller needs a conversion to be a win in the system, not just in the story. If converted spend is published as its own number, it can be carried as its own target. The account executive who moves four hundred thousand from committed to variable has done the thing the strategy requires, and the plan has to say so in money. Until converted spend has a line in the disclosure, no compensation committee will give it a line in the plan, because nobody designs a comp component around a quantity the company refuses to count.

The renewal stops being the unit of account. A renewal rate of 98% describes a world where the contract is the event. On a meter, the event is consumption, continuously, and the thing that threatens you is not non-renewal but a customer quietly running fewer actions for two quarters. That is invisible to a renewal rate and obvious in a consumption cohort. Whoever owns the number has to change with it, which is Layer Seven, and almost nobody has moved it.

The board question changes. Today the Chief Revenue Officer is asked why retention fell. On this scoreboard the question becomes how much committed spend converted, what the converted cohort is consuming against what it gave up, and what the realised rate did. Those are answerable questions about a transition. "Why did retention fall" is a question about a subscription business, and answering it honestly requires explaining that the metric is measuring the wrong thing, which no CRO has ever survived doing twice.

Underneath all three sit the layers this series has been building: metering and entitlement, because none of these numbers exist without a meter you trust; compensation, because the plan has to pay for conversion before sellers will do it willingly; revenue architecture, because converted spend has to be modelled before it can be forecast; and commercial systems, because a number you cannot produce from your own stack every month is a number you will quietly stop publishing by the third quarter.

The first ninety days

Publishing to the market is the end of this, not the beginning. The sequence that survives contact:

Days one to thirty. Count it internally and tell nobody. Produce all six numbers for the last four quarters from existing billing data. Most vendors discover two things here: the data is there, and it is worse than the story being told internally. Better to find that in a spreadsheet than in a filing.

Days thirty to sixty. Run it past the people who will be held to it. The CRO, the FP&A lead, the investor relations function, and the compensation committee. Every one of them will want a definition pinned down in writing, which is the point. A metric whose definition can move is a metric that will move, and both Salesforce and HubSpot widened a definition mid-crossing this year.

Days sixty to ninety. Publish the definitions before you publish the numbers. Put out the methodology a quarter ahead, exactly as a company does when it changes a segment. It costs nothing, it removes the suspicion that the metrics were reverse-engineered to flatter a bad quarter, and it means the first print lands as evidence rather than as excuse.

Then publish, and keep publishing through the quarters where the numbers are ugly. A disclosure set that appears when results are good and vanishes when they are not is worse than never having published it, because you have taught the market to read the silence.

The position

The argument of this season has been that the crossing is correct and the instruments cannot see it. That remains true. What this episode adds is that for the companies furthest into it, the instruments are not merely inherited. They are selected, quarter after quarter, from a set that describes the business being left behind, while the data describing the business being built sits in the billing system, counted to the credit, unpublished.

Salesforce sells Agentforce by the action and tells the market a recurring revenue number, then widens what that number includes. HubSpot bills fifty credits a resolution and publishes three subscription metrics. ServiceNow reports a 98% renewal rate for the quarter the market answered with the worst day in its history. Snowflake built four instruments nobody required it to build and is priced on its own arithmetic.

The gap between those two postures is not technical capability. It is willingness to be measured on the thing you claim to be doing.

In the Special Edition I committed to a test: if a frontier lab publishes a durable consumption measure that survives a slow quarter and the market accepts it, my argument about the instruments is wrong. That test stands, and this is the specification. Converted spend, retention on two bases, volume and revenue separated, realised rate, consumption concentration, and the cohort curve. Anything less is a press release with a number in it.

The companies crossing right now are going to be mispriced for four to eight quarters whatever they do. The only variable is whether they spend that time explaining a scoreboard that cannot see them, or publishing one that can.

Nobody is coming to invent this for you. The first vendor that builds it will spend two quarters being accused of inventing metrics to flatter a bad year.

Then everyone else will copy it, and the accusation will land on whoever is last.

Next: Season 3, Episode 6. The J-Curve. If you decide to cross, this is the order you move in, and this is what breaks in which quarter.

The Agentic Commercial Model. agenticcommercialmodel.com

Sources

All Season 3 episodes →Subscribe on LinkedIn →