The first time I tried to compare two AI agent vendors side by side, I gave up after an hour. One quoted me a per-seat number. The other quoted per resolution. A third wanted a platform fee plus “task credits” plus an overage rate I had to dig out of a footnote. There was no apples-to-apples version of the question “what will this cost me next year.”
That’s the state of AI agent pricing in 2026. The old per-seat model that ran SaaS for fifteen years is breaking, and nothing has cleanly replaced it. If you’re the person who has to sign off on the spend, you’re now expected to understand four different billing philosophies and spot the traps in each. Here’s how I’d think about it.
Why per-seat pricing is falling apart
Per-seat made sense when software was a tool a human operated. You pay for a Salesforce login because a salesperson sits in front of it all day. The seat maps to the value.
Agents break that mapping. An agent doesn’t occupy a seat — it does work. When a support agent resolves 4,000 tickets a month, charging “per user” is nonsensical, because there’s no user. The vendor either has to invent a fake seat or admit the seat was never the thing you were paying for.
The numbers show the shift is real, not just a talking point. Per-seat pricing fell from 21% to 15% of SaaS in roughly twelve months, according to Bessemer’s 2026 AI Pricing Playbook. On the buyer side, a Futurum 1H 2026 survey found 43% of buyers now prefer consumption-based models and 27% favor outcome-based structures — fewer than one in five still want classic per-user pricing.
So the seat isn’t dead, but it’s clearly demoted. Most vendors I’ve looked at this year either layer usage on top of seats or drop seats entirely for agent products. The interesting question isn’t whether per-seat survives. It’s what you should actually want in its place.
One more thing worth saying about per-seat before we move on: its real virtue was never accuracy, it was predictability. A CFO could look at headcount, multiply by a number, and know the line item for the year. Every model that replaces it trades that certainty for better alignment with value. That trade might be worth making — but go in knowing you’re giving up the one thing finance teams loved most about the old way.
The four models, and what each one really charges for
There are four billing shapes in the wild right now. Most real contracts are a blend, but it helps to understand them in pure form first.
Per-seat. You pay per licensed user, flat, regardless of how much work happens. Predictable, easy to budget, and increasingly mismatched to agent products. Still common for copilots that genuinely sit alongside a human — a coding assistant, a writing tool — because there a human really is in the loop.
Usage-based (per-task or consumption). You pay for what the agent consumes: tasks run, tokens burned, API calls, workflow executions. This is the model most aligned with how agents actually operate. It’s also the one most likely to surprise you on the invoice, because usage scales with adoption and adoption is hard to forecast in year one.
Outcome-based (per-resolution). You pay only when the agent delivers a defined result — a resolved ticket, a booked meeting, a closed case. Intercom’s Fin charges $0.99 per resolved conversation; HubSpot dropped its equivalent to $0.50 in April 2026. Zendesk shifted to charging for successful AI-driven resolutions instead of AI seats. The pitch is irresistible: you pay for results, not promises. The catch is in the word “defined,” and we’ll get there.
Hybrid. A base subscription plus usage or outcome overage. This is now the default — 41% of AI vendors use it, up from 27% in 2025 per Bessemer. Salesforce’s Agentforce tracks “Agentic Work Units” as consumption credits on top of platform fees. Adobe and ServiceNow blend all three. Hybrid exists because pure models each have a failure mode, and stacking them lets the vendor hedge.
Outcome-based pricing sounds perfect — read the fine print
I want to like outcome-based pricing. On paper it solves the alignment problem: the vendor only wins when you win. Sierra has built its whole brand on this, billing per successful resolution rather than per seat. There’s a genuine appeal to a vendor putting their revenue on the line for your results.
But “outcome” is a negotiated definition, not a fact of nature, and that’s where it gets messy.
Start with what counts as a resolution. For a simple FAQ, fine — the agent answered, the customer left happy, charge me. But Sierra’s own documentation admits the model is blended: some interactions bill per conversation (routing, greetings), others bill per outcome, and multi-touch or follow-up conversations get counted differently by contract. A “resolution” that involves an identity check, an account change, and an approval is not the same unit as “what are your hours,” yet both might show up on the invoice as one outcome.
Then there’s attribution. When a deal closes, the AI tool, your sales rep, and three other tools all touched it. Who gets credit? Vendors in adjacent spaces have already hit this — no-shows logged as “meetings booked” creating ghost charges, or a rep manually working an account and the agent claiming the result anyway. One pricing analyst called these “attribution knife fights,” and the phrase fits. The cleaner the outcome metric, the less room for the fight. Pin it down before you sign.
The irony is that outcome-based pricing is partly a defense against cost overruns. Deloitte’s 2026 State of Generative AI in the Enterprise found 72% of enterprise AI projects blow their original budget by at least 30%. Paying per confirmed result caps your downside in a way raw usage billing doesn’t. So it’s not that outcome pricing is bad — it’s that the contract language does all the work, and a vague definition turns “pay for results” into “pay for whatever the vendor’s dashboard decides to call a result.”
Usage-based pricing and the runaway-bill problem
Usage billing is the most honest reflection of how agents work, and the easiest one to lose control of.
The mechanism that makes it fair — you pay for exactly what you use — is the same one that makes it volatile. Roll an agent out to one team, costs are tiny. Roll it out company-wide after a good quarter, and a sudden adoption spike becomes a price-shock invoice nobody forecasted. With humans, usage is naturally bounded by how many hours people work. With agents, there’s no such ceiling unless you build one.
So if you go usage-based, the spending controls aren’t optional. Before signing, I’d want to know: Can I set hard caps that stop the agent, or only soft alerts that email me after the money’s spent? Is there throttling? Can I get per-team budgets instead of one company-wide meter? A vendor that can’t answer those is telling you the cost is your problem, not theirs.
This is also why pure usage pricing is drifting toward hybrid. A base subscription gives the vendor predictable revenue and gives you a floor you can budget against, while overage handles the variable part. It’s less elegant than pure usage, but “elegant” isn’t what your finance team wants — predictable is.
How to actually compare vendors
The real difficulty isn’t picking a favorite model in the abstract. It’s that vendors quote in different units on purpose, and the only way through is to normalize everything to one number you control: total annual cost at your expected volume.
Here’s the rough process I use.
Estimate your real volume first — tickets, tasks, resolutions, whatever the unit is — for year one and a realistic year two after adoption grows. Then force every quote into that frame. A $0.50 per-resolution rate and a $40-per-seat plan are comparable only once you’ve multiplied each out against your actual numbers. Do the arithmetic at low, expected, and high volume, because the model that wins at 1,000 tasks a month can be the loser at 50,000.
Watch for the costs that don’t show up in the headline rate. Outcome and usage vendors love a quiet platform fee, an implementation charge, or an annual renewal hike that lands after you’re locked in. Custom-quoted enterprise tools are the worst here — Sierra doesn’t publish pricing at all, and third-party estimates put real year-one budgets at $200,000–$350,000+ once setup fees are counted. The sticker rate is rarely the spend.
A few questions I’d put to every vendor before signing:
- Exactly what unit am I billed on, and who measures it — your dashboard or a source I can audit?
- What happens at 3x my expected volume? Show me that invoice.
- Can I cap spend hard, or only get warned after the fact?
- What’s the renewal price, in writing, not just year one?
- For outcome pricing: write down the precise definition of a billable outcome, including follow-ups, partial resolutions, and disputed ones.
If a vendor gets cagey on any of these, that’s information too.
What this means for your 2026 budget
The practical takeaway is that AI spend is becoming variable in a way software spend never was, and your budgeting has to follow. A fixed annual SaaS line item is turning into something closer to a cloud bill — partly committed, partly metered, capable of surprising you if nobody’s watching the meter.
Match the model to the work. For agents doing high-volume, clearly-defined tasks where you can pin down a clean success metric, outcome-based pricing genuinely aligns incentives and caps your downside — take it, but nail the definition. For agents whose work is variable or hard to attribute, usage-based with hard caps is more honest, as long as you actually set the caps. For copilots a human drives all day, per-seat is still fine and simpler than the alternatives. And hybrid is where most of you will land, because most real deployments are a mix.
The vendors have spent a year figuring out how to price agents in their favor. The least you can do is force every quote into your own spreadsheet before the renewal locks in. If you’re evaluating anything this quarter, build that low/expected/high volume model first — it’ll change which vendor looks cheapest more often than you’d think.
Sources: Futurum Group on outcome-based and hybrid pricing, Sierra on outcome-based pricing, Fin AI pricing comparison, Monetizely 2026 pricing guide, MindStudio on per-seat collapse, Featurebase on Sierra AI pricing.