AI Pricing Broke This Year. Check What Your Vendors Started Counting.
About four in five vendors recently introducing AI pricing chose to meter pricing on capacity over consumption. Buyers need to know what gets counted, who carries the quality risk, and whether they can audit the bill.
Chris Hart, Chief Executive Officer
||13 min read|
tl;dr
About four in five vendors recently introducing AI pricing chose capacity over consumption, and only about one in ten AI meters charges for a business result (Bain, August 2026). The label is not the meter. Three questions cut through: what am I being counted for, can I see the count, what happens when I hit the limit. Bryn's answers, in order: identified accounts, the audit log, a pause with no overage charges.
About four in five vendors recently introducing AI pricing chose to meter pricing on capacity over consumption. Buyers need to know what gets counted, who carries the quality risk, and whether they can audit the bill.
Among the AI pricing meters Bain reviewed, only about one in ten charges for a business result. The rest charge for effort or output.
Not because vendors turned cynical this year. Because the meter the industry used for twenty years stopped counting the actual valuable work produced.
Bain published some interesting numbers on August 10, from an analysis of publicly available pricing across roughly 200 B2B SaaS companies (Bain & Company, August 10, 2026).
If you buy software, their findings are more useful than any pricing headline or competitive breakdown you may be reading. If you sell it, the pricing meter is the most consequential decision on your roadmap that nobody outside the company thinks of as a product decision.
We set Bryn's meter in the spring and argued about it longer than we argued about most features. I still think it was the right thing to argue about. So, below I'm making the buyer's version of that argument.
The seat stopped counting the work
On July 1, Gartner found that up to $234 billion of enterprise application spending is exposed between now and 2030 as agents undercut the per seat model, roughly 20% of enterprise SaaS spend (Gartner, July 1, 2026).
The mechanism is simple. Agents complete tasks across several systems at once, so fewer people open the interfaces those systems charge for. George Brocklehurst, the Gartner managing vice president behind the research, put it plainly: this "breaks the link between user growth and revenue growth for many enterprise software vendors."
Note what he didn't say. He passed on the apocalypse framing, calling the disaggregation of the legacy SaaS market "less an apocalypse and more of a metamorphosis."
Not every dollar in that $234 billion will move, and a five-year forecast of software spend is directional at best. But the mechanism is already visible in your own stack. Count the seats you pay for on a tool where one person now runs work that used to take four. The seat count and the value have diverged.
When the meter stops tracking the value, someone changes the meter. That is what this year has been.
Effort, output, outcome. Only one makes payment depend on the business result.
Bain's most useful contribution is a distinction the market sometimes misses. Three things get grouped together under usage-based and outcome-based pricing, and they aren't the same:
Effort. You pay for what the work consumed: tokens, compute, agent hours. About 35% of AI meters.
Output. You pay for what the software produced: drafts generated, records updated, leads recommended. About 55%.
Outcome. You pay for a business result: a resolved conversation, recovered fraud, collected revenue. About 10%.
Shares from the Bain analysis, August 10, 2026.
The test that separates them is one question: who carries the quality risk?
Under an output meter, the vendor is paid whether or not the output was any good. Bain's example is exact. A recommended lead is an output. A qualified lead is an outcome.
Ten percent is fairly small because getting and measuring outcomes is tricky. Nearly every vendor deck this year hinted at outcomes, but only one in ten built a meter that pays only when the result arrives.
I don't read that as a failure to understand that outcomes are desirable, and I wouldn't want the number to be higher than it can honestly be. Outcome pricing needs three conditions: the result has to be observable, uniquely attributable to the software, and contractible between the two of you. Customer support, for example, can clear that bar, which is why Sierra, Fin, and Decagon all meter on "resolved conversations".
So when someone offers you outcome pricing on a pipeline, ask which of the three conditions they believe they've solved. Pipeline may be visible in a CRM. What is usually disputed is unique attribution to one agent.
This is the finding I would put in front of a CFO.
About four out of five vendors introducing AI pricing chose capacity, not consumption. You commit to a fixed amount. Unused capacity doesn't roll over and isn't refunded. Atlassian's Rovo credits, ServiceNow's Now Assist entitlements, and Adobe's Firefly credits all work this way.
Capacity preserves one important behavior of seat licensing: you pay for the commitment, not only what you consume. That buys budget predictability. It also makes the rollover, refund, and cap rules part of the product.
Which means those three rules deserve as much attention as the price. Most buyers read the numbers and skip them.
The label and the rate card don't always agree
HubSpot's April pricing change is an example of why a label is not enough. It calls both Customer Agent and Prospecting Agent outcome-based. Customer Agent charges $0.50 per resolved conversation, an outcome. Prospecting Agent charges $1 per lead recommended for outreach (HubSpot, April 2026). By the taxonomy above, a recommended lead is an output, not an outcome. Read the unit, not the headline.
GitHub announced on April 27 that all Copilot plans would move to usage-based billing on June 1. Premium request units were replaced by AI Credits, while base plan prices held (GitHub Blog). The buyer-facing unit also became harder to reason about: a Copilot code review can consume AI Credits and GitHub Actions minutes at the same time. One action, two billable meters.
I believe both companies were solving a genuine cost problem. Inference costs don't behave like the fixed costs software pricing was built around, and pretending otherwise would have caught up with them. But look at the buyer problem rather than the intent. HubSpot shows why you have to inspect the unit behind the label. GitHub shows why you have to inspect the mechanics behind the unit. In either case, the headline is not enough. What I wouldn't accept is a meter I can't audit.
Read the unit, not the headline
HubSpot Prospecting Agent · label: "outcome-based" · reveal the unit
The unit: $1 per lead recommended for outreach.
Class by the taxonomy: OUTPUT. A recommended lead is an output. A qualified lead would be an outcome.
GitHub Copilot code review · label: "usage-based billing" · reveal the unit
The unit: AI Credits (token-based) plus GitHub Actions minutes, on the same action.
One action, two billable meters. The mechanics behind the unit are part of the price.
Both cases as described in the article; units from the linked HubSpot and GitHub announcements.
A meter you cannot audit is not a price. It is a trust exercise with a number attached.
An opaque meter changes how your team uses the product
This is the cost nobody puts in the business case.
When people don't know what an action costs, they hesitate before taking it. They batch. They ask permission. They run the smaller version of the job to be safe. I've watched capable operators sit on a decision for a day because they weren't sure whether the run would blow through an allowance somebody else owned.
Hesitation is the expensive failure mode for anything you bought for speed, and it doesn't show up on a dashboard. It shows up as a tool nobody opens, and then as a renewal conversation where both sides are confused about why usage was low.
A meter isn't only a commercial instrument. It is a permission structure, and your team reads it that way whether or not you meant it to be one.
Three questions to ask about a meter
Strip out the vocabulary and there are three.
What am I being counted for?
Ask for the unit in one sentence, without the word credit in it. If the answer needs a conversion table, the unit is the table.
Can I see the count?
Not a monthly total on an invoice. The individual events, attributable to the thing that caused them, exportable. If the vendor's record is the only version and you can't pull it, you can't check the bill and you can't dispute it.
What happens when I hit the limit?
There are three honest answers: the work stops, the work continues and you're charged, or the work degrades. All three are defensible. What isn't defensible is finding out which one you bought after the invoice arrives.
Classify your vendor's meter
CHECK 1 What exactly is counted? One sentence, no conversion table.
CHECK 2 Can I independently inspect the count? Event-level, attributable, exportable.
CHECK 3 What happens at the limit? Stop, charge, or degrade: know before you sign.
Bryn: OUTPUT + CAPACITYHow we answer the three checks
What is countedIdentified accounts, the companies Bryn resolves and watches during a billing cycle. One unit, no conversion table between it and the bill.
Can you inspect itThe audit log, exportable at every point including on the way out. The CFO can export the log and reconcile the identified-account count to the invoice.
At the limitNew identifications pause, Plays already running on accounts in scope keep running, and there are no overage charges, so you know your maximum before you sign.
Monthly billing on every tier, no contract minimum, from $49. Current terms on the pricing page.
Pricing taxonomy adapted from Bain & Company, "AI Pricing: A Reality Check on Effort, Usage, and Outcomes," August 10, 2026.
What Bryn counts, and why
Bryn is the GTM agent for founders, first GTM hires, and growth teams.
By the taxonomy above, Bryn is OUTPUT + CAPACITY. Identified accounts are an output, not an outcome. And each tier includes a fixed monthly allowance, so it is capacity rather than pure consumption. That's the model we chose, and the rest of this section is why.
We had to answer those three questions before we could ship a price, so here are our answers in the same order.
What am I being counted for? Identified accounts. The companies Bryn resolves and watches for you during a billing cycle. Not seats, not sessions, not monthly active users. One unit, and no conversion table sitting between that unit and your bill.
We picked it because it's the closest available thing to the work. If Bryn watches more of your traffic, you're getting more. One operator with one login isn't penalized for covering more ground than a team of four used to cover. Leverage shouldn't raise your bill.
Can I see the count? The audit log. What Bryn saw, why it scored the way it did, what it ran, written to a record your growth team and your CFO can both read, exportable at every point, including on the way out.
That record started as the compliance answer. It became the billing answer too, which I didn't fully anticipate when we built it. Your CFO can export the log and reconcile the identified-account count to the invoice.
What happens when I hit the limit? Bryn pauses new identifications and waits. Plays already running on accounts in scope keep running, so work in flight doesn't break. There are no overage charges. You upgrade in one click or resume when the cycle resets, and the pause is written to the log either way.
So you know your maximum before you sign, and you learn you've outgrown a tier from a prompt inside the product rather than from an invoice at the end of the month.
Which is the answer to the hesitation problem three sections up. Nobody on your team should be doing arithmetic before they run a Play.
Monthly billing runs on every tier and the floor is $49. Current terms are on the pricing page. Bryn has no contract minimum. We'd rather the record earn the renewal than the contract enforce it.
If you're buying this year
Run the three questions before you run the demo. They take four minutes, and they'll tell you more about the next two years of that relationship than the feature comparison will.
Then ask one more thing: what happens to the record when I leave. A vendor who has thought hard about the meter has usually thought about the exit, because both questions come from the same place.
If you're repricing
You're choosing what you want your customers to believe you're selling. This is what a meter is.
Pick the unit that moves when your customer's value moves, then make it countable by them without your help. If you can't make it countable, you've picked the wrong unit, and renewal is where you'll find that out.
The market hasn't converged and I don't expect it to this year. Vendors are moving into and out of credit models as their cost curves change. That's uncomfortable if you were hoping for a standard to copy. It's useful if you were willing to do the work, because the meter is one of the few decisions left where a smaller company can simply be more honest than the incumbent and have the buyer notice.
If you're working through this on either side of the table, email me at chris@civic.com with the meter you chose and why.
Chris Hart is the CEO at Civic; he brings together decades of experience across technology, finance, and identity to help businesses navigate the shift to agentic AI. His Silicon Valley career spans more than 25 years, from running infrastructure at early internet and fintech startups to leading finance and operations teams at high-growth technology companies.
Beyond Civic, Chris has championed veteran leadership as Vice Chair of the Pat Tillman Foundation since 2006. When he isn't thinking about the future of identity and AI, you'll probably find him surfing or hanging out with his Dalmatian.
