Seats, Tokens, or Outcomes: Making Sense of AI Tool Pricing Before You Buy

I've sat through a lot of AI vendor evaluations over the past two years, and the strangest part is that "what does this cost" no longer gets a straight answer. One vendor quotes $40 per seat per month. Another quotes per thousand tokens, or per credit, or per "action," which turns out to mean whatever their meter says it means. A third wants 99 cents every time their agent resolves a support ticket. All three are pitching roughly the same capability, and for the same workload the three bills can differ by 10x. If you're signing one of these contracts this year, the pricing model deserves more scrutiny than the product demo.

Some quick context on why buying well is the skill worth building. MIT's NANDA initiative reported in August 2025 that about 95% of enterprise generative AI pilots produced no measurable P&L impact, and one of the more useful details in that research was that tools purchased from vendors reached successful deployment about two-thirds of the time, roughly triple the rate of internal builds. For most mid-market companies the takeaway is to buy rather than build, which puts the leverage in procurement. And pricing models are where procurement gets won or lost, because the model shapes behavior on both sides. Per-seat pricing makes you ration people. Usage pricing makes you ration activity. Outcome pricing makes you argue about definitions.

What the vendor is actually paying for

Behind every AI product there's a meter running at a model provider. Vendors pay for inference, billed per token in and per token out, and that cost scales with things you influence but never see itemized. Context length is the big one. An agent that reads your forty-page returns policy before answering a customer consumes many times the tokens of one answering from a two-line prompt, even though both look like "one answer" from your side. Premium models cost the vendor several times more per token than the workhorse models. Long conversations compound, because many systems re-send prior context on every turn. And agents multiply all of it, since a single request can fan out into a dozen model calls, retrievals, and tool invocations.

Every pricing model is a different way of passing that variable cost to you. Per-seat vendors absorb the variance, price for the average user, and protect themselves with fair-use clauses and feature gates. Usage vendors pass the meter through with a markup. Outcome vendors bundle the cost of all the failed attempts into the price of the successful ones. None of this is sinister, but knowing which cost-transfer game you're in tells you what to check before signing.

Per-seat: predictable, and it punishes broad rollout

Microsoft lists Microsoft 365 Copilot at $30 per user per month on top of the base Microsoft 365 license, and that structure is the template for most copilot-style tools. The appeal is genuine. Finance can budget to the dollar, procurement runs the same motions as any SaaS renewal, and no meter anxiety suppresses adoption, since nobody hesitates to use a tool that's already paid for.

The weakness shows up at rollout. Seat pricing charges for people, and usage across people is wildly uneven. In every deployment I've seen actually measured, a small group of power users generates most of the activity while a long tail logs in twice in the first month and drifts away. Put a $30 add-on on 300 employees and you're at $108,000 a year whether the tail ever comes back. So you either restrict seats to proven users, which caps your upside, or you pay for shelfware. There's a slower structural problem too. As tools shift from assisting humans to doing work autonomously, "per human" stops mapping to value, and vendors quietly respond by gating the expensive features behind credits inside the seat. At that point you're on usage pricing wearing a seat-pricing costume, and the predictability you paid for is gone.

Per-seat fits tools a human touches every time (writing assistance, meeting summaries, coding copilots) where usage is broad and steady. If you go this route, audit utilization quarterly and put a seat-reduction right in the order form, because vendors will happily let dormant licenses renew forever.

Usage-based: the meter runs on things you don't control

Usage pricing arrives in several units (tokens, credits, actions, conversations) and the unit matters more than the rate. Salesforce is the instructive case here. SaaStr's coverage of Agentforce pricing traces how Salesforce launched at $2 per conversation, then introduced Flex Credits at 10 cents per action in May 2025, partly because a single conversation could trigger a dozen backend processes. Look at that shift from the buyer's chair for a moment. A conversation is something you can count in your own systems. An action is whatever the vendor's meter records. Every step down in unit granularity moves the count further from your ability to verify it.

The classic failure mode is bill shock, and the cruel part is that it's usually caused by success. The tool works, the team routes more work to it, usage triples, and finance flags an overrun that is actually adoption doing exactly what you hoped. I've watched mid-market rollouts stall at precisely this point, with a CFO freezing expansion of a tool that was earning its keep, because nobody modeled the growth case before signing.

Credit systems deserve extra suspicion. Credits obscure unit prices, they expire, and their exchange rate can change underneath you. A vendor upgrading to a newer underlying model can double the credits consumed per task without touching the published price list. Before signing a credit-based deal, get in writing what the quoted price assumes about model tier and context size, and what notice you receive when either changes.

Usage pricing fits spiky or seasonal workloads, low-volume starts, and back-office processing you can throttle or batch. If you take it, insist on a real-time spend dashboard, alerts at thresholds you set, and a hard cap that throttles service rather than running an open tab.

Outcome-based: paying for results, once you agree what a result is

Intercom prices its Fin support agent at $0.99 per resolution, the most visible example of the newest model, where you pay when the AI finishes the job and failed attempts cost nothing. Document processing vendors run the same play per processed invoice or extracted contract. The alignment is genuinely attractive. The vendor now shares your interest in the tool actually working, and the CFO can hold 99 cents directly against what a human resolution costs.

The entire negotiation lives inside the definition. Intercom, to its credit, publishes one: a resolution counts when the customer confirms the answer helped, or when they leave the conversation without asking for more. That second clause is worth sitting with, because a customer who gives up in frustration and emails you instead can look identical, in the logs, to one who got a great answer and left satisfied. In document processing, ask whether a "processed" invoice with a wrongly extracted total still bills. Minimums also creep in around the edges. Intercom's standalone plan carries a 50-outcome monthly minimum, small in itself, but a signal that pure pay-per-result rarely stays pure at scale.

Outcome pricing fits high-volume, well-defined, verifiable units of work. It also demands that you keep your own count, because if you can't audit the vendor's tally against your own logs, you've agreed to pay a bill you can't check.

The same workload, priced three ways

Here's the exercise I'd run before any shortlist call, with round numbers. Say you run a 40-person support team handling 30,000 conversations a month, at a fully loaded cost of about $6 per human-handled conversation. Your pilot suggests AI can fully resolve about 40% of the mix, call it 12,000 conversations a month. Now price that same world under each model.

• Per-seat copilot at $50 per agent per month. $2,000 a month, $24,000 a year. It makes agents faster rather than resolving anything on its own, so the value arrives as shorter handle times. Trim 15% off average handle time and you've freed roughly six agents' worth of capacity, which dwarfs the license cost. But every dollar of that value depends on agents actually using the thing.
• Usage-based agent at 10 cents per action, averaging eight actions per conversation. Let it attempt all 30,000 conversations and you're at $24,000 a month, or $288,000 a year. Route it only the 12,000 it can plausibly handle and you're at $9,600 a month, about $115,000 a year. Your routing rules and the size of the knowledge base it reads now swing the bill by more than double, and those are engineering decisions nobody in finance ever sees.
• Outcome-based at $0.99 per resolution. 12,000 resolutions cost $11,880 a month, roughly $143,000 a year, with failed attempts free. Predictable per unit, though the total still rises with ticket volume, and the definition of resolution decides whether you're paying for 12,000 satisfied customers or 12,000 conversations that merely ended.

Same workload, and the annual bills come out at $24,000, somewhere between $115,000 and $288,000, and $143,000. No single number means much in isolation. The comparison that matters is dollars per successful unit of work measured against your current cost per unit. Those 12,000 conversations were costing about $72,000 a month in human handling, so even the priciest option can clear the bar comfortably, while the cheapest option might deliver the least actual displacement. Run that division before you compare any headlines.

Line items that don't show up in the demo

Whatever the model, the quote you see rarely equals the invoice you get. Here are the six places the gap usually hides.

• Overage rates. The blended rate inside your committed volume is often a fraction of the marginal rate above it. Get the overage price per unit in writing, not just the bundle price.
• Premium model tiers. Many quotes assume the vendor's cheaper workhorse model. The advanced-reasoning toggle that made the demo impressive can burn credits at several times the base rate.
• Implementation and integration fees. Connectors to your ERP, CRM, or ticketing system are often priced separately, and one-time setup fees have a way of coming back annually as maintenance.
• Indexing and data fees. Some vendors charge to embed and store your knowledge base, then charge again to re-index it when it changes.
• Support and SLA tiers. Response times you'd consider table stakes often live one tier up.
• Renewal uplift. Year-one pricing in a category this young is frequently a teaser. Ask for year-two unit pricing in the original order form, not at renewal.

Negotiation levers that actually move

Vendors in this category are still discovering their own cost structures, which cuts both ways. Pricing is soft, and reps have more room than their rate cards admit. A few levers I've seen work repeatedly in mid-market deals.

• Caps and collars. A hard monthly spend ceiling, with throttling instead of open-ended overage, traded against a modest committed floor. Vendors accept this more often than buyers expect.
• Pilot-to-production steps. Lock production unit pricing at signature, contingent on the pilot hitting agreed volumes. Otherwise a successful pilot becomes the vendor's leverage, because by renewal the tool is embedded and switching hurts.
• A written definition of the billable unit. Resolution, action, credit, processed document, whatever it is, define it in the contract and take the right to audit counts against your own logs.
• Model substitution protection. If the vendor swaps the underlying model, per-task consumption shouldn't rise without notice and a re-baselining of the rates.
• Data rights. Your prompts, documents, and outputs stay out of training sets unless you opt in, get deleted on exit, and export in a format you can actually use. Vendors expect this ask in 2026, and the ones who resist it are telling you something about their business model.

Before the next vendor call

The preparation fits in a spreadsheet and takes an afternoon.

1. Compute your current fully loaded cost per unit of the work you're automating, whether that's per ticket, per document, or per report.
2. Forecast volume at three levels (pilot, expected, and double expected), because if the tool works, doubling is what happens.
3. Price every vendor quote at all three volumes, with overages, premium model rates, and implementation fees included.
4. Ask each vendor to re-quote under a different model than the one they led with. How they respond tells you where their margin lives and how confident they are in their own product.
5. Write the billable-unit definition, the spend cap, and year-two pricing into the order form before signature.

We run a version of this exercise at FirmAdapt before clients take vendor calls, and the consistent result is shorter shortlists and noticeably better terms, mostly because so few buyers show up with unit math in hand. Treat AI contracts the way you already treat freight or payment processing, as unit-cost agreements to model and negotiate, and the seats-versus-tokens-versus-outcomes question largely answers itself once you know your own cost per unit.