Microsoft, HubSpot and Zoom Moved AI Pricing From the Seat to the Spend Limit

Microsoft switched on the meter, and October priced the envelope around it

On October 1, Microsoft posted a notice to its reseller channel about how Copilot gets billed. Starting December 1, usage-based billing is on by default for new Microsoft 365 Copilot Business licenses sold through resellers. The default spending limit is 4,000 Copilot Credits per user a month. At Microsoft's published rate of $0.01 per credit, that's up to $40 a user on top of the seat, charged only when the work runs. The seat stays. A meter now sits under it, and the meter ships switched on.

Every other pricing move in October fits around that notice. Two weeks later, HubSpot made AI spend caps the default. The account-level credit limit sits up to 50% above a customer's monthly credits, and customers can swap seats for credits mid-contract. On October 20, Zoom told CX Foundation that Zoom CX customers now commit to one annual spend that draws down across agents, seats, applications and AI. Read in order, those are a meter, a ceiling on the meter and a floor under it. A buyer’s AI contract is turning into a spending envelope, and the seat is one input inside it. I think that's the biggest change this year in how GTM software gets bought, and the defaults are where it gets decided.

The four weekly retrospectives each caught one piece of this. Put together, they show the number a GTM leader negotiates at a 2027 AI renewal moving off the seat price and onto the envelope's terms. Does the meter start on or off? Where does the ceiling sit, and how large is the committed floor? How fast does each product draw it down, and do the per-unit rates hold for the term?

A vendor paid by the envelope earns only when its agents run. So October's vendors gave agents named jobs, logins and the first contact with customers, and some sent engineers into accounts to make sure the agents got used. The proof that the metered work produces revenue didn't keep pace. That gap is where a buyer's leverage sits for the next 30 days.

This piece covers October 1 through 25. Anything that breaks in the month's last week goes into November's first weekly.

Four weeks moved the negotiable number to the ceiling and the floor

The month opened with agents disappearing into contracts customers had already approved. Between September 30 and October 1, Apollo, Gong and ZoomInfo each shipped agents on top of their data. ZoomInfo's Agent Teams came with "no new SKU or contract required."

Our weekly retrospective for that week argued that GTM data vendors had become agent platforms. The standalone agent tool selling neither the data nor the system would be the first line cut at renewal. A month later, the pricing is the detail to keep. A data vendor priced its agents at zero on top of a renewal. An agent that can't yet price its own access has to ride inside a contract someone already signed. The envelope started there, with the agent work bundled and unpriced.

The second week put a public price on the work underneath. On October 9, Andreessen Horowitz led an $870 million round in TypeSafe AI, maker of the Jev decision model, at a $7.5 billion valuation. The same day, Microsoft shipped Microsoft-Decision-1 at $0.042 per million input tokens, exactly Jev's list price. The weekly for October 5 argued that revenue teams should buy routing and scoring per decision. Their labeled outcome history is the asset that costs more than the model. The same weekly found vendors keeping the seat and switching a capped meter on underneath it. HG Insights priced a new platform on consumption credits, and ServiceNow's Amit Zavery told CNBC-TV18 that half the company's net-new revenue over the past year wasn't seat-based. HubSpot ran the same play on marketing email. Since October 1, new customers in Europe, the Middle East and Africa buy one seat-priced platform. Every email HubSpot sends for them, except one-to-one inbox mail, draws 15 credits (15 euro cents) apiece for the first 10,000 sends, according to its pricing supplement. For a list-heavy marketing team, that meter can outweigh the seat.

The third week was about the ceiling. HubSpot's default cap landed on October 15. A day earlier, G2's 2026 AI Agents Report found only 9% of more than 1,000 B2B software decision makers rate per-seat pricing fairest for AI agents. Preference for pay-as-you-go rose to 57% from 38%. The same week, four companies disclosed revenue that rises with use. Anthropic told prospective investors that third-quarter revenue rose about 45% to $16.8 billion, the BBC reported. Modal said it crossed a $1 billion annualized run rate billing compute by the second, and Ramp's co-CEOs wrote that it passed $2 billion in annualized revenue. The weekly argued that AI revenue now grows by the meter and that the term deciding a 2027 renewal is the buyer's spending ceiling rather than the seat price.

The fourth week supplied the floor. Zoom's pool and Gainsight's plan to stop selling per-seat licenses for Gainsight CS and price on the customer ARR it manages both arrived on October 20. Gainsight's Atlas service, which pairs agents with renewal managers on long-tail accounts, already manages more than $75 million of customers' ARR. That same day, the Wall Street Journal reported on a draft of Anthropic's IPO prospectus showing $16.3 billion of first-half revenue. SpaceX filed results showing $7.56 billion of third-quarter revenue in its Super Intelligence segment, which holds Cursor and Grok. Together they showed why 2027 AI renewals will ask buyers to commit. Four days later, The Information reported that OpenAI is cutting prices to pull Anthropic's biggest API customers. Replit got more than 30% off a contract worth over $100 million a year, per The Decoder's summary. The weekly argued that AI pricing's unit moved from the seat to the commitment. Model makers owe most of their compute regardless of use, so a year of committed spend is worth more to them than a higher rate card.

Read across the month, the four weeklies describe one contract being rebuilt in sequence. The agent went into the bundle first, then the work underneath got a public price per call. A meter went under the seat, the buyer got a ceiling on that meter, and the seller asked for a floor under it. By October 25, the seat had become the least interesting number in an AI renewal.

Cheap AI decisions moved the money to the spending envelope

The envelope matters because the unit inside it stopped being worth fighting over. Jev's $0.042 per million input tokens was matched by Microsoft on the day of TypeSafe's round. By our arithmetic, a routing call that sends 1,000 tokens of lead record costs about $0.000042 at that price, so 100,000 inbound leads a month comes to roughly $4.20. No procurement team will spend a meeting on $4.20.

Those prices still add up to large bills, because volume does the work that rates used to. Anthropic's $16.8 billion quarter, Modal's billion-dollar run rate and Exa's $100 million in annualized revenue on 2 billion requests a month are each a tiny unit price multiplied by usage nobody capped. Every one of those figures is company-stated or reported from an investor meeting, and none splits growth between new accounts and existing ones using more. The direction is the same in all four, though. When the unit is nearly free, the size of the bill is set by how much work runs, and how much work runs is set by the envelope's terms.

So naturally, the fight moved. Microsoft's 4,000-credit default looks generous per user and adds up to 400,000 credits a month for a 100-seat customer, which is Microsoft's own example. HubSpot's cap sits up to 50% above the monthly allowance unless an admin moves it. Zoom's pool lets a customer shift spend between seats and AI without a contract review, which also means the AI draw-down rate decides how fast the pool empties. In each case the vendor set a default, and a buyer who never touches it accepts the vendor's view of how much AI work should run.

Model suppliers face the same pressure from the other side. OpenAI's reported discounts to Replit were struck one account at a time. Replit had already made OpenAI its preferred model partner for Replit Agent, CEO Amjad Masad said on October 15, citing cost and performance. On a meter, a customer that big can move its volume without breaking a contract. The supplier's answer is to ask for a commitment and discount the rate in exchange. Accenture's buyers showed what that looks like inside an enterprise. Since October 1, it has funded $29 of each employee's monthly Claude use centrally and charged the rest to project codes, 404 Media reported. The buyer built its own envelope, with a central floor and a chargeback ceiling, because nobody upstream had.

Zoom hasn't published draw-down rates for its pool. Gainsight has no rate card yet for pricing on customer ARR. Microsoft published its default limit and the per-credit price, but which Copilot experiences draw how many credits sits in licensing guides most buyers won't read before December. The envelope arrived before its price list did, and that's the part I'd push hardest on. A buyer who signs the pool first and learns the draw-down rates later has handed the vendor the one number that decides the bill.

Agents took revenue jobs because metered vendors earn when work runs

A seat vendor gets paid whether or not a customer logs in. An envelope vendor gets paid when its agents draw down the pool, so October's launches pushed agents into jobs where they'd run every day.

On October 8, Google introduced the Gemini agent at its Gemini at Work event. A coworker agent built with it receives its own Workspace account, with an email address, a calendar, a directory listing and every action logged under its own name. A day earlier, Fin shipped Fin 4 and said its agent now drives customer success, building onboarding plans and flagging expansion. Later in the month, Anthropic opened Claude Duet as a waitlisted preview that checks a small-business owner's connected apps every few hours. It stops for approval before it pays, signs or sends anything. monday.com launched Harmony, a voice agent that calls and qualifies sales leads, priced at $0.39 a minute of agent time plus a $250 monthly platform fee. Podium said its AI agents passed $150 million in ARR with more than 20,000 deployed, sold as an add-on to plans customers already pay for. Each launch gives an agent a first contact, whether that's an inbox, an onboarding plan, an owner's morning look at the books or a new lead's first call.

Then vendors paid to make sure the agents ran. On October 7, Atlassian launched a Forward Deployed Engineering program to build Rovo agents inside enterprise accounts, with engagements of about 12 weeks. On October 13, Profound's CEO James Cadwallader posted that the company is hiring its first forward deployed marketing engineers, placed in customer success and accountable for adoption. Profound's agents draw down credits each time they run. An engineer embedded in an account is how a vendor on a meter buys the usage its revenue depends on, and neither vendor published a price for its engineers.

Pricing by draw-down turns adoption into the vendor's revenue line, so vendors invest in adoption. Nobody in October published evidence that the adoption produced revenue for the customer. Boston Consulting Group's survey of 398 sales leaders found 90% report meaningful AI use, while only 13% name win rates as a top payoff. A Salesforce regional vice president said agents working leads nobody had answered produced more than $88 million in pipeline. No close rate was attached, which is the pattern across its lead-agent disclosures. Forrester's 2027 predictions say two-thirds of providers will lose deals because buyers can't verify their claims, which turns every unsourced sales claim into a deal risk.

Those three numbers come from a consulting survey, a vendor's own count and an analyst forecast, and the self-reported ones flatter adoption. They still point the same way. Agents are running more of the revenue process, and the bill for that running is moving onto a meter. The evidence that it closes business is still a pipeline figure with no conversion rate. A buyer paying by draw-down is paying for activity until the vendor shows outcomes.

Google's design suggests the governance answer, even though Google didn't frame it that way. An agent with its own login can be inventoried, given a sponsor and audited like an employee. Microsoft's Agent 365 documentation already requires a named sponsor for every agent identity. Once each agent has an owner, the owner can be asked what the agent's draw-down bought.

Town and Revenera defend the seat, but Town's plan already holds a meter

Town and Revenera read October the other way. Town raised a $108 million Series B at a $1 billion valuation on October 19. It says paying teams grew 9x since July on a plan sold at $59 a seat a month. A Town assistant can only hand work to a colleague who has one, so the product spreads by adding seats. Revenera's Monetization Monitor, released October 14, found that only 7% of 415 software makers had lost customer seats. They rank pricing predictability ahead of near-term revenue by 47% to 16%, per Revenera's release, a finding we read alongside Kyle Poyar's survey. On that reading, the seat is healthy, buyers still like predictable bills and October's meters are a sideshow sold mostly to the largest accounts.

Part of that is right. The seat is holding in the income statement, and no vendor that published terms in October removed it. HubSpot kept the seat as the floor of the bill. Microsoft kept everyday Copilot on a per-user license and moved agents to metered credits. ServiceNow's chief executive, Bill McDermott, said on the April earnings call that many customers pick seats once they see the value. A buyer who reads October as the death of the seat will overpay for flexibility they don't use.

The reading misses what's inside Town's own seat. Town's pricing page gives each Team seat 6,000 credits pooled across the team, with the overage cap set at $0 by default. It’s an envelope in a seat-shaped wrapper, with a committed floor bought per head, a shared pool and a default ceiling. Revenera's predictability finding points the same way. Software makers that want predictable revenue and buyers that want predictable bills both end up at a committed pool with a cap. Zoom, HubSpot and Town now sell it under different names.

Buyers with the most bargaining power push the objection further. On October 5, The Information reported that Microsoft had cut its projected internal spend on Anthropic by more than a third. Meta's Claude Code users fell to about 30,000 from about 60,000. A buyer that owns a substitute can ration by switching, and an envelope with a floor locks in spend it may want to move. The risk is real for a buyer who signs a large pool without rate protection. It argues for writing the envelope carefully instead of staying on seats, because at Microsoft, HubSpot, Zoom and even Town, the seat now comes with a meter attached.

What this means for GTM leaders over the next 30 days

If I were walking into a 2027 renewal, here's the order I'd work in, because some of the defaults land before most renewals start.

By December 1, inventory every AI meter in the stack and record its default state, its limit and who can change it. Microsoft's Copilot Business default switches on for new reseller purchases on that date, and HubSpot's cap is already live for its customers. Set each limit deliberately, even if the number stays where the vendor put it, and name the admin who owns it.

Ahead of the first 2027 renewal conversation, write down the envelope terms you'll ask for. Get the draw-down rate for each product in the order form, along with rollover of unused commitment into the next term. Lock per-unit rates for the full term, so a discount on the commitment can't be clawed back through the meter. Confirm in writing that hitting the ceiling pauses work rather than upgrading the pack. Decide whether usage past the floor is funded centrally or charged to the team that ran it, the way Accenture now does, before the vendor's console decides for you.

Price the decision steps underneath any AI commitment longer than a year before signing it. Ask each vendor which model sits under its routing, scoring and qualification features and what a call costs them. Run a few hundred historical leads with known outcomes through two or three decision models and check accuracy on the cases your team argues about. A vendor charging a large envelope for work that runs on a $0.042 model has room to move, and a buyer who knows the unit cost can show it.

Write the outcome definition into the contract before the next agent pilot, and keep a holdout group. Give every agent identity a named sponsor, with RevOps for anything that writes to pipeline, customer success for adoption agents and finance for anything that touches billing. When a vendor offers forward deployed engineers, get their cost and their success measure on the order form, and agree on what counts as adoption before they arrive.

For sellers moving to pools, the order runs the other way. Publish draw-down rates before your reps quote them. Build billing that can meter each account and enforce a customer's cap, which is what Intuit agreed to buy Sequence for on October 15. Then settle whether reps earn on the committed amount or the consumed one, because the two diverge the first quarter a customer under-uses.

The large seat vendors will follow Zoom and HubSpot within 90 days

Our bet is that within 90 days, at least three of Salesforce, ServiceNow, Workday, Gong, Outreach and ZoomInfo announce either a default spend cap on their AI credits or a single annual AI commitment across products. The November retrospective should show the first of them. If none has moved by late January, October's envelopes stay a mid-market pattern. The large vendors would keep selling seats with credits bolted on and no default limit.

By October 25, the seat had become the least interesting number in an AI renewal.

By December 1, list every AI meter in your stack with its default state, its limit, and the admin who can change it, then set each limit on purpose before the first 2027 renewal.

By Business of GTM