HubSpot Concedes the AI Credit Pricing Argument in Q2

HubSpot’s Q2 2026 results are the clearest concession yet from a major CRM that credit-metered AI pricing has been holding buyers back. The company reported revenue of $911.7 million for the quarter ended June 30, 2026 — up 20% as reported and 17% in constant currency — and framed a product, pricing and go-to-market overhaul around a single idea: customers want predictable pricing when they adopt AI agents.

That framing matters more than the revenue line for anyone who has spent the past year trying to write a defensible budget for an agent rollout. Metered AI pricing has been the standing objection in every procurement conversation we have been part of: nobody can approve a line item whose ceiling is a function of how enthusiastically their own staff use the product. When the vendor with roughly 306,000 customers starts describing outcome-based pricing and spend thresholds as the direction of travel — as it did on this quarter's earnings call, per coverage of it — the objection has been heard.

This piece separates what HubSpot actually filed from what was said on the earnings call and picked up by call coverage, reads the per-agent adoption split honestly rather than as one uniform adoption story, sets HubSpot’s structure against what Salesforce and Zoho publish on their own pricing pages, and gives you a worksheet for pricing a metered agent rollout before you sign anything. It is analysis for operators budgeting agent work — not investment commentary.

Start with the tier of fact that carries the most weight. HubSpot’s Q2 2026 results were published on August 5, 2026 and filed as an exhibit to a current report on Form 8-K, with the same release mirrored on HubSpot’s investor relations site. Everything in this section comes from that document.

Total revenue was $911.7 million for the quarter ended June 30, 2026, up 20% as reported and 17% in constant currency. Both figures belong together: the three-point gap is currency movement, not demand, and quoting only the as-reported number overstates the underlying trajectory. Subscription revenue was $894.0 million, also up 20% as reported, with professional services and other revenue at $17.7 million.

Profitability swung. GAAP operating income was $43.3 million against a GAAP operating loss of $24.6 million a year earlier — a GAAP operating margin of 4.8% versus negative 3.2%. Separately, and this is a different measure that is easy to conflate, non-GAAP operating income was $185.3 million, up 44%, for a non-GAAP operating margin of 20.3% against 17.0% a year earlier. Operating cash flow was $222.8 million and non-GAAP free cash flow $167.9 million.

HubSpot also guided Q3 2026 revenue to $924.0 million to $925.0 million, which is 14% growth as reported and 15% in constant currency, with non-GAAP operating income of $187.0 million to $188.0 million. Full-year guidance is $3.678 billion to $3.686 billion. Read as operating context rather than market commentary, the shape is a company expanding margin while its growth rate steps down — which is exactly the condition under which a vendor reworks packaging and entry prices rather than defending them.

One governance item from the same filing week: the board expanded to eleven directors and appointed Jerry Dischler as a Class I director effective August 5, 2026. Dischler previously led Google Cloud Applications and AI development for Google Workspace. Board appointments are not usually worth an operator’s attention, but a product-and-AI operator joining the board of a company mid-way through an AI packaging overhaul is a reasonable signal about where the next few quarters of roadmap attention go.

The load-bearing sentence in the release is a direct quote from CEO Yamini Rangan. It is worth reading as a procurement document rather than as marketing copy, because it names the two things buyers have been asking for by name — outcomes and predictability — and commits product, pricing and go-to-market to delivering them.

On the earnings call, per coverage of that call, Rangan went further on the same theme. She reportedly told analysts that businesses have been hit with unpredictable token costs and want pricing that is transparent and tied to value, and described introducing outcome-based pricing for several agents, lowering entry price points, and giving customers visibility and control over usage and spend — including the ability to set thresholds. We are relaying those as reported statements rather than quoting them: they reach us through call coverage rather than a transcript we could read directly, and the wording may not be exact.

Strip the earnings-call register away and the operator translation is short. Spend thresholds mean a metered product can be capped. Lower entry price points mean the first agent no longer requires a tier upgrade to try. Outcome-based pricing means the billable event moves from something the vendor’s software does to something your business actually got. Those three changes address, in order, the budget-ceiling objection, the pilot-cost objection and the value-attribution objection — the three that kill agent deals in our experience.

The numbers that travelled furthest from this quarter are per-agent activated-customer counts. They are genuinely useful — very few CRM vendors publish per-product agent adoption at all — but they need a label attached every time they are used, and that label is not “HubSpot reported.”

These figures were discussed on the Q2 2026 earnings call. They do not appear in the press release or the 8-K exhibit. We could not read the corrected call transcript directly at the time of writing; what we have is near-identical reporting of the same figures across three independent outlets covering the call, including stockanalysis.com and Investing.com. Three-outlet agreement on exact figures is strong corroboration. It is still not the same confidence tier as a filed document, and we treat it accordingly.

Plotted against the prior quarter, the trajectory is the interesting part. The Data Agent moved from a base above 9,000 customers in Q1 to a reported 16,000-plus in Q2, and the Prospecting Agent from roughly 14,000 to roughly 17,000. Note what happens to the growth rates as the bases get larger: the Data Agent’s quarter-over-quarter growth was reported at 122% in Q1 and 80% in Q2. That is a completely normal shape for a product moving from early adopters into the main install base, and it is more informative than either quarter’s absolute count in isolation. One caution if you rebuild the arithmetic yourself: the reported growth rates and the rounded counts do not reconcile exactly for the Prospecting Agent, which is what you should expect from figures relayed through call coverage rather than a filed table. Read the direction, not the decimal.

Set those counts against the 306,446-customer base and the honest read is that agent activation is still an early-innings number for every product in the line-up. Roughly 17,000 activated customers on a 306,000-customer base is a single-digit share. The story here is direction and velocity, not saturation — and any vendor pitch that implies the market has already standardised on agents should be checked against that ratio.

The most useful thing in the adoption data is not the totals, it is the ordering. Two agents that operate on internal data and internal workflows are being switched on faster than the one agent that talks to your customers. That ordering is not an accident and it is not a statement about relative product quality — it is a statement about blast radius.

Our reading, from running these deployments rather than from the call: the adoption curve follows reversibility. A data agent’s mistakes are correctable in the CRM and visible in a field-history audit. A prospecting agent’s mistakes are embarrassing but bounded, and most teams gate the send. A customer agent’s mistakes are live, in front of a paying customer, and unrecoverable in the sense that matters. Buyers are sequencing their rollouts by how expensive an error is, which is exactly what a competent operator should do. We wrote up the mechanics of that sequencing in our guide to HubSpot AI agent workflows.

There is a pricing consequence hiding in this ordering. Outcome-based pricing is easiest to define for the customer-facing agent — a resolved conversation is a clean, countable outcome — and hardest for the data agent, where the “outcome” is a field that is now correct. So the agent with the cleanest billable outcome is the one buyers are slowest to deploy, and the agents buyers are fastest to deploy are the ones where outcome pricing is hardest to define. That tension is the real design problem behind any credit system, and it is why per-action credits have persisted despite the complaints.

HubSpot’s Agent Hub and Agent Builder entered public beta on July 23, 2026 for Professional and Enterprise customers, per trade coverage of the launch, with Agent Hub included in those editions during the beta. HubSpot’s own product page describes Agent Hub as managing marketing, sales and service agents inside the Smart CRM, with HubSpot Credits supporting pay-for-results usage. What it does not do is publish per-credit rates, and the dedicated Agent Hub pricing URL did not resolve when we checked it.

That gap matters, so we are being explicit about it: per-credit figures for HubSpot Agent Hub are circulating on third-party sites, and we are not printing them, because we could not confirm a single one of them on a HubSpot-owned page. A number you cannot trace to the vendor is not a number you can budget against. Where we can trace rates — Salesforce and Zoho both publish theirs — the table below reports them and says so.

Two structural observations. First, Salesforce runs three billing models simultaneously, which is a hedge rather than a position: credits for teams that want granularity, per-conversation for teams that want outcome alignment, and flat seats for teams that want a number their finance function will approve without a usage forecast. Second, Zoho meters five times finer per credit — a tenth of a cent per credit against Salesforce’s half-cent — while bundling the core AI features into existing seat tiers. Those are genuinely different philosophies about where AI value should be captured, and we compared their practical consequences in our CRM AI agent comparison across Salesforce, HubSpot and Zoho.

The unrolled-credit clause in the Salesforce terms deserves its own line in your evaluation. Credits that expire at term end convert a usage-based product into a use-it-or-lose-it commitment, which is the worst of both structures: you carry the forecasting burden of metered pricing and the sunk-cost dynamics of a seat licence. Ask every vendor the same two questions — do unused credits roll over, and can I set a hard spend cap that stops work rather than merely alerting me.

Here is the exercise we run with clients before any agent contract is signed. Because HubSpot’s own per-credit rates are not published on a page we could verify, the worksheet is calibrated against the two rates Salesforce does publish — $0.10 per standard billable action and $2.00 per conversation. Substitute your own vendor’s confirmed rates when you have them; the structure is what transfers, not the numbers.

The inputs are deliberately simple. Assume 4,000 support conversations per month routed to an agent, an autonomous resolution share of 72% of routed conversations — the rate reported for HubSpot’s Customer Agent, used here purely as an illustrative input — and, under the per-action model, three billable actions per conversation. Every derived cell below is computed from the formula stated beside it.

The spread is the whole point. The same month of work, the same resolution rate, the same customers — and a 6.7× difference in cost driven entirely by which billing unit was in the contract. Per-conversation pricing looks cleaner and aligns better with value, which is exactly why it is priced higher per unit; per-action pricing looks cheaper until someone changes an agent’s configuration to take five actions per conversation instead of three, at which point Rate A becomes $2,000 a month without anyone approving an increase.

Two lines deserve to be read twice. The escalation line is the one most budgets forget: 1,120 conversations a month still land on a human, so the agent’s cost sits on top of a support cost that has not gone to zero. And the seat-equivalence line is the argument your finance function will actually engage with — Rate A costs about what ten seats cost, Rate B about what sixty-four cost. If the agent does not plausibly displace that much human capacity, the metered model is not the cheaper option, it is just the one with the less visible invoice. Our agent token budget and cost-control framework extends this into a full per-workload model, and the same discipline applies to regulated deployments like the Army HRC Agentforce rollout at IL5, where the volume assumptions are enormous and the escalation path is mandatory.

If HubSpot is conceding an argument, it is worth asking how strong the argument was. The honest answer is that the evidence base is thinner and more commercially interested than the confidence with which it gets quoted, so we will name every source and let you discount accordingly.

Taken together with what HubSpot said this quarter, the picture is consistent even if no single data point is decisive: buyers report being surprised by AI charges, vendors are converging on hybrid structures that put a predictable floor under a variable component, and the largest CRM vendors are now saying out loud that predictability is a feature. The strongest evidence is not the survey data at all — it is that a vendor with a metered product told its investors, per coverage of the call, that it is lowering entry prices and adding spend thresholds. Companies do not restructure pricing that is working.

The forward read, and this is our projection rather than anything HubSpot stated: expect the industry to land on a hybrid default within a few quarters — a seat or platform fee that includes a real allowance of agent work, a hard threshold you configure yourself, and outcome pricing reserved for the narrow set of agents where an outcome is genuinely countable. Pure per-token or per-credit pricing will survive at the developer-platform layer, where the buyer is technical and the volume is forecastable, and will keep losing ground in the CRM seat market, where the buyer is a revenue leader who needs a number for a budget line. The vendors that get there first will do it by making the cap a first-class product feature rather than a billing-page footnote.

Nothing in this quarter changes what you should be doing, but it improves your negotiating position. A vendor that has publicly committed to predictable pricing has given you language to quote back during procurement. Use it.

Three practical steps regardless of which column you land in. Run the worksheet above with your own volumes and the vendor’s confirmed rates before the first call, so you enter the conversation with a number rather than reacting to theirs. Instrument the agent from day one — actions per conversation, escalation rate, resolution rate — because every one of those is a billing input and none of them is stable. And sequence your rollout the way the adoption data suggests the market already is: internal-data agents first, drafting agents second, customer-facing agents when your escalation path is genuinely tested. If you want that sequencing designed against your own CRM rather than in the abstract, that is the core of our CRM automation work, and the broader operating-model questions sit inside our AI transformation engagements.