A new breakdown every week — read the newsletter →
ΘNavigator
PeopleBusinessesToolsNewsletter
StrategiesMental ModelsDecision ToolsBusiness ModelsFrameworksMoats
LearnBook SummariesReading ListsGuidesQuote Collections
Artificial Intelligence Learning PathsWhat should I study? AboutContact
Business Model

BUSINESS MODEL — Usage-Based / AI-Metered Pricing

Business Models · Model #56 (2026-native) · Consumption / outcome monetization · 20 min read

Thesis: For thirty years software was sold by the seat. When one AI agent does the work of ten people, the seat becomes the wrong thing to count — so the industry is re-pricing itself around what the software does, metered by token, action, conversation, resolution, and increasingly by outcome.

1. What this model is

Usage-based pricing charges customers for what they consume, not for how many people have logins. AI-metered pricing is its 2026 mutation: the unit of consumption is no longer just storage or API calls but units of work an AI performs — a token generated, an action taken, a conversation handled, a ticket resolved, an outcome delivered.

The shift matters because it breaks the central assumption of the SaaS era. Per-seat pricing worked when software was a tool a human operated: more humans, more seats, more revenue. But an AI agent is not a tool a human operates — it is the worker. If a support team of fifty shrinks to five because agents handle the rest, per-seat pricing means the vendor's revenue collapses exactly as the vendor delivers more value. The seat count and the value delivered now move in opposite directions. Metering the work instead of the workers realigns them.

2. How the money flows — the metering spectrum

There is not one AI-metered model; there is a spectrum, ordered by how closely the charge tracks delivered value:

LOOSELY TIED TO VALUE ───────────────────────────────► TIGHTLY TIED TO VALUE
per-token      per-action     per-conversation   per-resolution   per-outcome   % of value
(API compute)  (each step)    (each dialogue)    (solved E2E)     (result)      (revenue share)
   ▲                ▲               ▲                  ▲               ▲             ▲
 raw model      "Flex Credits"   pay whether or   pay only when    pay when the  vendor takes a
 access; you    ~$0.10/action    not it's solved  the AI resolves  business goal  cut of revenue
 pay for the                     ($2/convo)       end-to-end       is achieved    it produces/recovers
 fuel                                             (~$0.99–$2)
  • Per-token — the rawest form: you pay for the model's compute (input + output tokens). This is how foundation-model APIs are sold. Pure fuel pricing; the customer bears all the risk of turning tokens into value.
  • Per-action — each discrete step an agent takes (look up an order, update a record) is metered. Salesforce packages this as "Flex Credits," roughly a dime per standard action.
  • Per-conversation — you pay for every dialogue the agent handles, resolved or not. Salesforce Agentforce launched here at about $2 per conversation.
  • Per-resolution — you pay only when the agent solves the issue end-to-end without a human. Intercom's Fin publishes $0.99 per outcome; Zendesk sits around $1.50; HubSpot cut to $0.50 in 2026; Quickchat around $0.50. This is the first rung where vendor revenue tracks the metric the buyer actually cares about.
  • Per-outcome / % of value — the vendor is paid only when a defined business result lands: a deal closed, a cart recovered, an invoice collected, a claim processed. Sierra built its whole pitch on "pay for the work the agent completes, not the technology it consumes." The purest form takes a cut of the revenue the agent produces or recovers.

Almost every real enterprise deal is actually a hybrid: a fixed platform fee (for access, seats, support) plus usage on top. That's where deals land whatever the marketing says.

3. The mechanics — and the unit-economics trap

The elegance of metered pricing hides a hard problem the seat era never had: your cost of goods scales with usage. Every token the agent generates costs you inference compute. In per-seat SaaS, serving one more heavy user was nearly free; in AI-metered software, every unit of work carries a real marginal cost. So the model only works if the price per unit sits durably above the compute cost per unit — and that spread is being squeezed from both sides as buyers demand lower per-resolution rates and vendors race each other down (a 4x gap already exists on the same unit, from ~$0.50 to ~$2.00 per resolution).

Three levers keep the model healthy: 1. Falling inference cost — model prices per token have fallen fast; each drop widens the vendor's margin on a fixed sticker price. 2. Rising resolution quality — if the agent resolves 80% instead of 70%, the vendor earns more and the buyer gets a better outcome, on the same traffic. Quality improvement is revenue growth with no new customers. 3. Definition control — "resolution" is vendor-defined. Two vendors both quoting $0.99 can bill very differently depending on what counts as a resolution (some count non-escalated conversations; the honest ones count only genuine positive resolutions). This is the model's biggest integrity risk and its biggest margin lever.

4. Where it wins

  • When value per unit is high and variable. A resolved support ticket, a collected invoice, a qualified lead — each has clear standalone worth, so metering it feels fair to the buyer.
  • When the buyer can't model their own costs. Vague usage plus a demo is how most contracts get signed; a vendor who can show cost tracking value wins procurement.
  • When the vendor's product genuinely improves over time. Metered pricing turns quality gains directly into revenue — a flywheel per-seat pricing can't capture.
  • When adoption risk is the blocker. "Pay only when it works" collapses the buyer's risk and shortens the sales cycle. Intercom even backs Fin with a performance guarantee of up to $1M if it misses resolution targets — a contract shape impossible under seat licensing.

5. Where it breaks

  • Revenue unpredictability (both sides). Buyers hate variable bills they can't forecast; vendors get lumpy, harder-to-model revenue. Zendesk's 2026 move to automatic overage billing with no notice is the cautionary tale — metered pricing without transparency breeds distrust fast.
  • The COGS floor. If inference cost ever rises or quality stalls, the spread evaporates and you're selling dollars for ninety cents.
  • Definitional disputes. When the invoice depends on what counts as a "resolution" or "outcome," every billing cycle is a potential argument. Trust becomes the product.
  • Adverse gaming. Per-conversation pricing rewards the vendor even for failures (you pay whether or not it's solved); per-outcome pricing can tempt vendors to define "outcome" loosely. The model's incentives are only as clean as its definitions.
  • Cannibalizing your own seats. If you also sell per-seat software, an agent that reduces the customer's headcount reduces your seat revenue — early Agentforce deployments already show ~10% support-seat reductions. You must be willing to disrupt your own pricing before someone else does.

6. Worked example — the support-agent economics

A mid-market company runs 50,000 support conversations a month. Under the old model they paid for ~40 agent seats. Under AI-metered pricing:

  • The vendor deploys an agent that resolves 75% of conversations end-to-end.
  • At $0.99 per resolution, that's 37,500 resolutions × $0.99 ≈ $37,000/month to the vendor.
  • The buyer redeploys most of its 40 seats; even after the agent's bill, total cost drops and coverage goes 24/7.
  • As the agent improves ~1%/month, resolutions climb toward 80–84% — the vendor's revenue rises with no new customer and the buyer's cost-per-contact keeps falling.

Now run the same traffic at $2/conversation (paid whether solved or not): 50,000 × $2 = $100,000/month, plus platform and data-cloud fees. Same traffic, 3x the bill, and the incentives no longer point the same way. The spread between these two invoices — on identical work — is the entire strategic debate of 2026 enterprise software.

7. Real companies (2026)

  • Foundation-model APIs — per-token pricing; the fuel layer everyone else meters on top of.
  • Twilio, Snowflake, AWS, Stripe — the pre-AI usage-based canon (per-message, per-compute-credit, pay-as-you-go); the proof the model scales to giants.
  • Intercom Fin — ~$0.99 per outcome, ~76% average resolution across ~12,000 customers, ~1%/month improvement, $1M performance guarantee; reportedly being folded into Salesforce in 2026.
  • Salesforce Agentforce — the tell of the whole market: it runs three pricing models at once (per-seat, per-conversation at ~$2, Flex Credits at ~$0.10/action) and added pay-per-resolution for its Help Agent — a giant hedging because it knows which way this goes. Its agent business reportedly reached ~$540M ARR in roughly a year.
  • Sierra (Bret Taylor) — outcome-based, enterprise, opaque pricing centered on the argument that vendors should be paid when the work gets done.
  • Zendesk, HubSpot, Ada, Decagon, Quickchat — the per-resolution field, from ~$0.50 to ~$1.50, each defining "resolution" a little differently.

8. The incentive-alignment thesis (why this is more than pricing)

The deepest thing about outcome pricing is that it changes the contract between software and buyer. Per-seat licensing pays the vendor whether or not the software works. Per-outcome pricing pays the vendor only when it does. That single change realigns the entire relationship: the vendor now has a direct financial stake in the customer's result, which pushes product, support, and roadmap toward delivered value instead of booked seats. It is, quietly, the most buyer-friendly shift in enterprise software economics in decades — and the reason incumbents are terrified of it, because it exposes exactly how much of their revenue was never tied to value at all.

9. Adjacent models

  • Freemium (#18) — often the front door to usage-based: free tier, then meter the heavy users.
  • All-you-can-use / Flat-rate (#15) — the opposite bet: hide usage variance behind one price; increasingly hard to sustain when your COGS is metered compute.
  • Razor-and-blade — a fixed platform fee (razor) plus metered usage (blades) is the dominant hybrid shape.
  • Affiliate / % of value — outcome pricing at its limit becomes revenue-share, kin to affiliate economics.
  • Subscription — what usage-based is eating; the two increasingly coexist as "platform fee + usage."

10. Analyst's Take

The seat is dying, but slowly, and the winners will be the ones who manage the transition rather than the ones who are loudest about the destination. Three convictions:

First, pick your rung on the value ladder honestly. Per-token pricing is fair but transfers all the value-creation risk to your customer, which caps how much you can charge. Per-outcome pricing captures the most value but demands you define the outcome cleanly and stake your margin on delivering it. Most companies should start one rung looser than they think and tighten as their quality data earns the right — you cannot credibly sell "pay only when it works" until you can prove how often it works.

Second, "resolution" is a product decision, not a billing footnote. The vendors who win the trust war will publish honest definitions and count only genuine successes; the ones who count non-escalated conversations as resolutions will win a few quarters and lose the relationship. In a model where the invoice depends on a definition, transparency is the moat. Build the billing so the customer can audit it, and you turn the model's biggest weakness into a differentiator.

Third, watch the COGS floor like it's the only number that matters — because eventually it is. Metered AI pricing is the first software model in a generation with a real, usage-scaling cost of goods. Falling inference prices are papering over thin spreads right now; if that trend pauses, a lot of per-resolution businesses discover they've been selling below cost. The durable players are the ones widening the gap between price-per-unit and cost-per-unit through genuine quality gains and efficiency, not the ones racing to the bottom on sticker price.

The uncomfortable truth for incumbents: if your revenue is tied to seats and your customers are cutting seats because your own AI works, you are already short your own stock. The only defense is to disrupt your pricing before the market does it for you — which is exactly why the biggest incumbent is running three models at once and won't say which one wins.

11. 2026 and beyond

Expect the spectrum to slide rightward, toward outcome and value-share, as agents get reliable enough to stake margin on. Expect a wave of billing-integrity tooling — third-party verification of what counts as a resolution — because buyers won't accept vendor-defined units forever. Expect hybrid platform-fee-plus-usage to be the steady state, not pure metering. And expect the hardest strategic question of the late 2020s to be the one this model forces: when your software replaces the buyer's labor, how much of the savings do you get to keep? The answer the market is converging on — some, but not all, and only when it works — is the most honest pricing software has ever had.

12. Top Resources

  1. Monetizely — SaaS Pricing Benchmark studies. The clearest running data on per-seat vs. outcome-based adoption.
  2. "Understanding Outcome-Based Pricing" — Pragmatic Institute. The framework for tiering outcome pricing without giving away margin.
  3. Intercom Fin + Stripe pricing case study. How a per-outcome model is actually metered and billed in production.
  4. Public Agentforce pricing evolution (conversation → Flex Credits → per-resolution). The single best real-world map of a giant hedging across the whole spectrum.
  5. "Why AI Is Killing Per-Seat SaaS Pricing." The clearest articulation of why seat-count became the wrong number to count.