A Good Value Metric Grows When Your Customer's Headcount Falls

Snowflake bills for consumption and kept 125% of last year's revenue from the same customers. Asana bills for seats and kept 96%. That 29 point gap is the clearest public test available of what a value metric is worth. A metric that grows when your customer automates survives. A metric priced per human does not.

What a value metric has to do in 2026

A value metric is the unit you multiply your price by. Seats, contacts, gigabytes, transactions, resolved tickets. It is the single most consequential choice in a pricing model, and most teams inherit it rather than choose it.

For twenty years the seat was a reasonable proxy. More value delivered meant more people using the software. That proxy has now broken in one specific place, and the break is not subtle.

The three part test

Run every candidate metric through three questions, in this order.

First, does it rise when the customer gets a better result? Second, does it hold when the customer does the same work with fewer people? Third, can the buyer count it themselves, without trusting your dashboard?

A metric that fails the first is a tax. A metric that fails the second is a countdown. A metric that fails the third is a dispute waiting for a renewal date.

Why seats fail the second test

Seats are a count of humans doing work. Automation removes humans from work. The two move in opposite directions by definition, which means the better your product performs, the smaller your invoice becomes.

That is not a pricing inefficiency. It is a structural short position on your own product working. No amount of discounting or packaging fixes a metric that pays you for failure, and the pattern is already visible in the compression running through seat-based contracts.

The 29 point gap between two live metrics

Two public companies reported net revenue retention within a week of each other in February 2026. They sell different products to different buyers. They also bill on opposite sides of this argument, which makes the contrast worth reading carefully.

What 125% means on a consumption meter

Snowflake reported net revenue retention of 125% as of 31 January 2026, alongside full-year product revenue of $4.47 billion, up 29%. The same customer base spent a quarter more than it did a year earlier.

Nothing in that number required a sales conversation. Customers ran more queries, stored more data, and the meter counted it. Expansion arrived as a consequence of usage rather than as a negotiation.

What 96% means on a seat meter

Asana reported dollar-based net retention of 96% in the same quarter, with core customers at 97% and customers spending over $100,000 also at 96%. Full-year revenue reached $790.8 million, up 9%.

Read the 96% carefully. It means the installed base contracted by 4% before a single new customer was counted. Growth has to be bought from the market every year, because the meter does not generate any on its own.

The company is not badly run and the product is well regarded. The metric simply cannot expand without headcount expanding, and its customers are not adding seats.

Seats shrink because the job shrinks

The argument above only matters if buyers are genuinely removing people. They are, and the evidence is now specific rather than anecdotal.

Salesforce cut its own support seats by 44%

Marc Benioff said in September 2025 that Salesforce had reduced its support organisation from 9,000 people to about 5,000, with roughly half of customer interactions handled by agents. Support costs fell 17%.

Now put yourself on the other side of that. Any vendor billing Salesforce support by the seat lost 44% of that line item in under a year, while the software it sold worked better than ever.

The wider labour data points the same way. The US Bureau of Labor Statistics projects employment of customer service representatives to decline 5% between 2024 and 2034, from 2,814,000 jobs, a loss of about 153,700 positions. The handbook names automation as the cause.

A 5% decline over a decade sounds mild. It is not mild if your revenue is a linear function of that headcount and your contracts renew annually. It also understates the near-term effect, since the cuts cluster in exactly the large accounts that software vendors depend on.

The metrics that hold, and the ones that only look like they do

There are five metric families in common use. They behave very differently when the customer automates, and two of them are frequently mistaken for each other.

Prices are published list prices read in August 2026. The behaviour column is a structural claim about each metric, not a measured result for the named vendors.

Credits and tokens are cost recovery, not value

The credit meter is the most common mistake in AI pricing right now. It looks like usage-based pricing and it is priced like a utility, so it feels modern.

It measures the wrong side of the boundary. A credit counts what your product consumed, not what your customer received. When inference prices fall, your revenue per unit of delivered value falls with them, which is a strange thing to design on purpose. The gap between falling token prices and rising bills is its own recurring problem, and the margin consequences show up in what inference costs do to AI gross margins.

Credits are a legitimate cost-recovery layer underneath a value metric. They are a poor value metric on their own. HubSpot's structure shows the distinction clearly: contacts are the value metric, seats gate access, and credits meter the AI work sitting under both.

Score the metric before you price it

Pricing committees argue about the number and skip the unit. Force the unit into a scored decision instead, with weights agreed before anyone sees the scores.

One rule overrides the total. If a metric scores below 3 on the second criterion, it is disqualified regardless of how well it does elsewhere. A high total on a metric that shrinks with automation is a well-argued countdown.

The weights are mine, and I would defend the split between the first two rows. Alignment with the customer's result is worth more than immunity to headcount, because a metric can be automation-proof and still feel arbitrary, which is how you end up renegotiating every renewal.

The migration, in four moves

Choosing the metric is the easy half. Moving a live base onto it is where most attempts stall, usually because the team treats it as a pricing announcement rather than a twelve month sequence.

Move one is instrumentation. Count the new metric in production for a full quarter before anyone outside the company hears about it. You will find that your definition is wrong at least once, and it is much cheaper to discover that before it appears on an invoice.

Move two is free reporting. Ship the usage dashboard on its own, with no pricing attached. This separates visibility from pricing risk in the customer's mind, and it removes the single most common objection to usage-based billing, which is not the cost but the loss of control.

Shadow bill before you announce

Move three runs both meters at once. Show the customer what they pay today and what they would have paid on the new metric, on the same invoice, for two full quarters.

You are buying two things with that delay. The customer gets time to see whether the new number is volatile, and you get a list of every account where the change is punitive before they find it themselves. Expect between 10% and 20% of accounts to look worse on the new metric, and decide in advance what you will do about them.

Cap the renewal increase in writing

Move four is the cutover, and it happens one contract at a time, on each account's own renewal date. Never force a mid-term change.

HubSpot's 2024 seats migration is the cleanest published template. Existing customers stayed on legacy terms until renewal, and the company stated up front that any migration-related increase would be about 5% or less. That sentence does most of the work. It converts an open-ended risk into a bounded one, and bounded risks do not trigger procurement escalations.

Publish your own version of that cap even if your number is different. A cap you commit to in writing is worth more than a discount you offer verbally, because the buyer can take it to their finance team.

What an outcome metric actually costs you

Outcome pricing scores highest on the first two criteria and it is genuinely the direction of travel. It also has a cost that vendors consistently understate, and buyers should press on it.

The cost is definitional. When you charge per outcome, the definition of the outcome becomes the most valuable clause in the contract, and it is usually written by one party.

Intercom publishes $0.99 per outcome for its Fin agent. Its own page counts an outcome when the customer confirms resolution, or does not ask for more help, or when the agent completes a workflow including a handoff. That third case is worth reading twice. A handoff to a human can be a billable outcome.

Zendesk went first in this category, announcing outcome-based pricing for AI agents in August 2024 on the principle that customers pay only for issues resolved autonomously. The announcement carried no per-resolution price, which is a fair reflection of how much of this pricing is still negotiated rather than published.

My position, as a buyer: outcome pricing is better than seat pricing and worse than it looks. Ask for the counting rule in writing, ask what happens on an escalation, and ask for a monthly cap. A vendor who will not put the definition in the contract is selling you their measurement, not their result. The same discipline applies when the meter is a conversation rather than an outcome, as in the arithmetic behind Agentforce pricing.

Where this argument is weakest

Three objections, and the first one is serious.

Two companies is not a study. Snowflake and Asana differ in category, growth stage, buyer and competitive position. Data infrastructure is expanding and collaboration software is saturated. The metric is one variable among many, and the honest reading is that the retention gap is consistent with the argument rather than proof of it. If you want a controlled comparison, nobody has published one, and I would rather say that than dress up two data points as a finding.

Consumption metrics cut both ways. A meter that rises with usage also falls with it. Customers optimise, workloads get cheaper, and a bad quarter arrives without any churn event to explain it. Boards pay a multiple for predictability, and consumption revenue is harder to forecast than a seat contract. That is a real cost, not a rounding error.

Seats are not dead. They remain the cleanest way to price access, the easiest thing for procurement to approve, and the cheapest thing to administer. Most of the market is landing on a hybrid, where a seat anchors access and a second meter prices the automated work. HubSpot runs exactly that structure. The argument here is not that seats disappear, it is that seats can no longer be the metric that carries expansion, a point developed further in the autopsy of per-seat pricing.

There is a fourth possibility I cannot rule out. The retention gap may be downstream of category growth rather than metric design, in which case picking a better metric in a shrinking category buys you very little. Watch what happens to consumption vendors in the next demand contraction, since that is the test this argument has not yet faced.

Frequently asked questions

Where to start

Open your billing data and run one query. For your twenty largest accounts, plot billed units against the account's headcount over the last eight quarters.

If those two lines track each other, your value metric is headcount with extra steps, and every efficiency gain you ship is a future price cut you have already agreed to. That correlation is the finding. It takes an afternoon and it settles the argument faster than any pricing workshop will.

Then pick one candidate metric from the table above and instrument it this quarter. Bill nothing. You are buying a quarter of real data before you have to defend a number, and that is the cheapest option available on this whole list.

Related reading