- 01. The five bands on the AI pricing spectrum — Token, Hybrid, Per-Action, Per-Outcome, Value Share — and what each band trades between transparency, risk, and value capture.
- 02. The verified 2026 anchors at each band — Salesforce Agentforce ($500 / 100k Flex Credits), Intercom Fin ($0.99 / outcome), Zendesk ($1.50–$2.00 per automated resolution) — with the operational design that makes each one work.
- 03. Why moving rightward on the spectrum compounds upside — and why most AI pricing failures come from being one band to the left of where the value capture actually lives.
- 04. The four-question diagnostic for picking your band, and the four pricing-design traps that drag teams back down the spectrum after they thought they had graduated.
The two products with the same engineering and different futures
Two AI features ship in the same quarter at two different SaaS companies. Both built on similar models. Both engineering teams of comparable strength. Both feature sets aimed at customer support — one a Copilot, one an Agent. Both technically excellent.
Six months later, one is the highest-margin product line in its company. The other has been quietly re-bundled into the base plan because, on every customer review, the line item produced a procurement conversation that ended in price compression.
The model wasn’t different. The engineering wasn’t different. The pricing model was different. The first company priced per-resolved-outcome — every successful resolution generated a clean, billable event tied to value the customer could see. The second priced per-seat, then added usage credits, then added overages, then added discounts to keep procurement happy. By month six the pricing was so complex that the customer’s CFO couldn’t tell whether they were getting value, and procurement assumed the answer was no.
This is the pattern most AI PMs underestimate: pricing is not a packaging decision made after the product is built. It’s a product decision that determines what kind of company you become. Token pricing makes you infrastructure. Outcome pricing makes you a service. The choice shapes everything downstream — your CAC, your gross margin, your customer-success motion, your competitive moat, your renewal dynamics, your exit narrative.
The argument of this post is that AI pricing in 2026 lives on a five-band spectrum, and most teams are pricing one band to the left of where their value capture actually lives. This post is the diagnostic for moving rightward — at the right pace, with the right operational infrastructure, anchored to the verified 2026 cases that show what each band looks like when it works.
The frame to carry into every pricing review
The five bands run from rawest to most aligned, from most transparent to most outcome-coupled, from most risk on the customer to most risk on the vendor. The further right you go:
- Pricing aligns more with customer value capture.
- Measurement infrastructure required goes up.
- Vendor variance risk goes up or transfers to the customer (depending on band).
- Captured share of value created goes up.
The frame:
Most AI products are priced one band to the left of where their value capture actually lives. The right move is rarely to invent a new model — it’s to graduate one band rightward, with the operational infrastructure to make the graduation real.
The structural read on AI pricing in 2026This is also the structural answer to a question that has gone from theoretical to operational across 2026: will every AI company eventually price on outcomes? No. Most AI companies will eventually price somewhere in bands 2–4. The teams that try to skip directly to band 5 without the measurement infrastructure underneath collapse into legal disputes and customer-success-as-debate-club. The teams that stay in band 1 watch their margins compress every quarter as competitors graduate rightward.
From charging for compute to charging for the result
Figure 1 — The five-band spectrum, with the canonical 2026 anchors
Each band trades transparency for value capture. Per-Outcome is highlighted because it is where the practitioner consensus is converging in 2026 — but the band you should be in is the band your measurement infrastructure can honestly defend today.
Cost transparency × customer alignment — the four archetypes mapped to segment fit
Figure 2 — The Pricing Archetype Matrix
Mature AI companies run multiple archetypes simultaneously: abstracted-value for the enterprise platform, outcome-based for high-attribution features, tiered for SMB, token for the developer API. Forcing one model on everything is the most common pricing failure in 2026.
Per-seat → tiered → outcome → abstracted — plan it before crisis
Figure 3 — The Pricing Transition Arc
By Year 3 a mature AI company runs all four archetypes simultaneously. The strategic asset isn’t the destination — it’s the discipline of planning each transition before the next cost cliff hits.
Band 1 — Token / Metered
Pricing by raw consumption. Customer pays per token, per API call, per GPU-hour, per credit-of-compute. The “infrastructure” model.
What it is
The most transparent and least aligned model. The customer sees exactly what they are consuming and pays exactly for that consumption. The vendor takes essentially no variance risk — variable cost is passed through. The customer takes all variance risk.
This is the model the underlying API providers (OpenAI, Anthropic, DeepSeek, Google) live on, and it is the model most AI products inherit by accident — the AI is wrapped in a SaaS shell, but the cost structure is metered, and when push comes to shove the metered cost shows up in the contract.
When it works
When the customer is sophisticated enough to do the cost forecasting themselves. Developer tools, research workflows, internal-platform deployments inside data-rich organisations. The customer understands tokens, accepts variance, and prefers transparency over predictability.
When it fails
When the customer doesn’t think in tokens. The most expensive public lesson on this came from Cursor in June 2025 — a credit-pool model where a single power user exhausted the pool in roughly an afternoon, producing a $7,225 invoice that became briefly famous on Hacker News. The technical cause was a missing rate limit. The actual cause was an instinct, imported from SaaS, that pricing should be transparent and metered. In SaaS that instinct is correct. In AI it routes you straight into the brand crisis we covered in L1-T02.
The PM lesson: token pricing is right when your customers think in tokens. For everyone else, it’s a brand-risk event waiting for the right power user to surface it.
Band 2 — Hybrid
Base subscription fee + usage component. The customer gets predictability for the base, the vendor captures upside on usage.
What it is
Where most of the SaaS market is consolidating in 2026. A platform fee covers access, governance, integration, support, baseline usage. A usage component captures variable consumption beyond the baseline.
The 2025–2026 SaaS pricing data shows hybrid surging from 27% to 41% of SaaS companies as their primary pricing model, with industry analysts referencing Gartner forecasts that 70% of businesses will prefer usage-based over per-seat by end of 2026 (SaaS pricing analysis 2026).
Verified anchors
Most of the major consumer AI subscriptions are hybrid: Notion AI, ChatGPT Plus, Claude Pro all use a flat subscription with usage caps that cut in at high consumption (Notion pricing; OpenAI pricing; Anthropic pricing). The cap is the variance protection — the customer experiences a flat price under normal usage, the vendor caps tail-risk above the threshold.
When it works
When the customer needs predictability for budgeting and you need upside on heavy users. When the value distribution is bimodal — some customers light, some heavy. When the cap mechanism is clear and the user understands when they will hit it.
When it fails
Three failure modes:
- The hybrid is just per-seat with a tip jar. Customers see through this in the first procurement conversation. The “platform fee” scales linearly with seats and the “usage component” is decoration.
- The cap is invisible until the user hits it. The customer pays the flat fee, hits the cap during a heavy use cycle, and experiences the cap as an unexpected punishment. Trust erodes.
- The hybrid prices the median user, not the heavy user. The flat fee and the cap are both sized on a “typical” usage profile. The most engaged customer hits the cap weekly and resents the model.
Band 3 — Per-Action / Per-Resolution
Pricing by discrete unit of work completed. Each action, each resolution, each transaction generates a clean billing event tied to a unit of customer value.
What it is
The first band where pricing is structurally tied to what the AI does, not what it costs to run. The unit can be a resolution, a generated document, a routed ticket, a classification, a transaction. The discipline: the unit is discrete, measurable, and roughly equal in customer value across instances.
Verified anchors
Zendesk Automated Resolutions. Zendesk launched per-AR pricing in August 2024 and has held the model since: $1.50 committed / $2.00 pay-as-you-go per automated resolution, with 5–15 free ARs per agent per month depending on plan (Zendesk pricing; supporting pricing analysis). The model has been stable through 2026 — a meaningful signal that per-resolution pricing is durable when the unit is well-defined.
Salesforce Agentforce — Flex Credits. Salesforce’s flexible pricing model launched in May 2025 and was repriced in 2026: $500 per 100,000 Flex Credits, where each standard action consumes 20 credits (~$0.10 per action) and a voice action consumes 30 credits (Salesforce Agentforce pricing; launch press release). Eligible Foundations customers receive 100,000 free credits as an entry point. Salesforce also offers a flat alternative at $2.00 per conversation for customer-facing agents.
The Flex Credits design is a worth-studying case in pricing abstraction. Rather than expose tokens, Salesforce abstracts the variable cost into a credit unit that customers can budget against — different actions cost different credit amounts, but the credit itself is the budgeting primitive. The customer doesn’t have to think in tokens; they think in actions and credits. The vendor preserves variance protection (because credits map to underlying cost) while the customer gets predictability of unit cost.
When it works
When the unit of work is unambiguously discrete and the customer agrees on the definition. When the cost-per-action is small enough that procurement can think in pre-purchased blocks rather than per-instance fees. When the AI’s value is concentrated in the action itself rather than in some downstream business outcome.
When it fails
Three failure modes:
- The unit is contested. The customer counts “resolutions” differently from how you count them. Dispute. Friction. Renewal compression.
- Action costs at the P90 tail exceed action prices. Same dynamic as the P90/P50 cost ratio from L1-T08 — your most engaged customers hit P90 cost on actions that you are billing at average cost. Margin compresses precisely on your best customers.
- The customer doesn’t think in actions. They think in seats, or in outcomes, or in monthly budgets. The action-based bill becomes a procurement burden, and the renewal request is “can we go back to the simpler model.”
Band 4 — Per-Outcome
Pricing by measured outcome captured. The customer pays only for results, not for actions taken on the way to results.
What it is
One step further than per-action — paying for the result, not the work. The system can take many actions to produce one outcome; the customer pays only when the outcome materialises. This is structurally where the highest gross margins in the AI economy live, because outcomes are heterogeneous, the customer’s willingness to pay tracks the value they capture, and the vendor’s incentive aligns with the customer’s success rather than fighting it.
Verified anchor — Intercom Fin
The cleanest 2026 example of per-outcome pricing at scale is Intercom Fin, which charges $0.99 per outcome — billed once per conversation, when the conversation is resolved or has no follow-up within a defined period. There is a $49/month base plan that includes 50 outcomes, with overage at the per-outcome rate (Intercom Fin pricing; Intercom platform pricing).
The reported impact has been substantial. Intercom has stated that Fin has grown from roughly $1M to over $100M ARR, that it now handles 80%+ of support volume for many customers, and that the average resolution rate is 67%+ (Growth Unhinged analysis; GTM Now interview with Intercom President).
The single most important design decision in Intercom Fin’s pricing isn’t the dollar amount — it is the outcome definition. “Resolved or no follow-up within X period” is the operational discipline that turns the ambiguous concept of “outcome” into a clean billing event. The customer can see whether the outcome was achieved. The vendor can defend the count. Disputes are bounded.
This is the pattern most teams trying to reach for outcome pricing fail to ship. They write the contract with the word “outcomes” but never define what an outcome operationally is. Intercom Fin’s $0.99 doesn’t work because of the price point. It works because of the eight or nine words after the dollar amount that turn the outcome into a billable event.
When it works
When the outcome is unambiguously measurable. When attribution to the AI is defensible. When the customer agrees on the measurement framework before signing. When the cost-per-outcome at the vendor side is small enough that the vendor can absorb the per-outcome variance.
When it fails
Three failure modes:
- Outcome is contested. The customer argues the outcome was actually delivered by their human team, not the AI. Without pre-agreed measurement, this becomes a legal dispute, not a pricing question.
- Lag between cost and revenue. The vendor incurs inference cost in January; the customer captures the outcome in May; the vendor is cash-flow underwater for four months on a deal they cannot yet bill. This is real and most outcome-pricing setups do not model it well at launch.
- Outcome attribution rate underperforms expectations. The vendor priced assuming a 60% resolution rate; actual rate is 35%; revenue lands at less than half of plan; gross margin still negative because variable cost was sized for the higher rate.
Band 5 — Value Share
Pricing as a percentage of value recovered or generated. The vendor’s revenue is a share of the customer’s outcome.
What it is
The most aligned model — and the hardest to operationalise. The vendor takes a percentage of the value their system produces for the customer. Recovered chargebacks, generated revenue, prevented churn, reduced fraud loss. The vendor’s incentive structurally matches the customer’s success. The vendor only gets paid when the customer measurably wins.
When it works
In specific verticals where the value is unambiguous, measurable, attributable, and high enough per event to justify a value share. Fraud detection, chargeback recovery, lead qualification, sales pipeline acceleration — the verticals where every “won” event is meaningful and traceable.
When it fails
Three failure modes:
- Attribution is contested at every event. Customer disputes whether the AI caused the win. Without pre-agreed attribution rules, this becomes the entire customer relationship.
- The customer wants value share to be one-way. They will accept paying X% of value generated, but they will not accept the vendor capturing X% of value prevented (e.g., churn reduction). One-way value share is much harder to justify.
- The infrastructure cost is structural. Value-share pricing demands measurement infrastructure that costs roughly as much to build as the AI itself. Most teams underestimate this and ship the contract before the infrastructure exists.
The frontier honesty: most AI products will not end up at Band 5. The infrastructure cost and attribution rigour are too high for most workflows. Bands 3 and 4 will absorb the bulk of where the AI economy lands. Band 5 is where the most ambitious vertical AI plays will live, and where the highest gross margins in the entire AI economy will be earned — the long-horizon ceiling we map in L3-T07: The Golden Quadrant — Service as Software (Coming soon).
The pricing spectrum at a glance
Five bands. Five anchors. One scoreboard.
| Band | Pricing | Customer transparency | Variance risk | Verified 2026 anchor |
|---|---|---|---|---|
| 1. Token / Metered | $/token, $/credit | Highest | Customer-borne | API providers (OpenAI, Anthropic, DeepSeek) |
| 2. Hybrid | Base + usage | High | Shared | Notion AI, ChatGPT Plus, Claude Pro |
| 3. Per-Action | $/discrete unit | Medium | Medium | Zendesk ($1.50–$2.00/AR), Salesforce Agentforce ($500/100k Flex Credits, $0.10/action) |
| 4. Per-Outcome | $/measured result | Medium-low | Vendor-borne (until success) | Intercom Fin ($0.99/outcome) |
| 5. Value Share | % of value | Lowest | Vendor-borne | Bespoke enterprise / specific verticals |
The rightward graduation rule: most teams should be one band to the right of where they are currently priced. Not two. Not five. The infrastructure to support each band has to be earned, and skipping bands collapses faster than staying put.
How architecture and pricing connect
Each architectural archetype from L2-T13: Product Architecture as a Strategy Decision maps cleanly to a pricing band — and the misalignments are where most teams quietly bleed margin:
- Augmentation → Band 1 or 2. The AI is invisible; the customer doesn’t isolate the AI value. Pricing is bundled or hybrid. Trying to price Augmentation per-action makes the customer notice something they shouldn’t have to.
- Copilot → Band 2 or 3. Suggestions are countable. Per-seat works because the user is the unit of value, but per-action is increasingly the right move as Copilots mature.
- Agent → Band 3, 4, or 5. The AI takes actions and produces outcomes. Anchoring to per-action (Band 3) is the safe entry; per-outcome (Band 4) is where the gross margin lives; value-share (Band 5) is the long-horizon ceiling.
If your architecture is Agent and your pricing is Hybrid (Band 2), you are leaving money on the table — and probably also producing a customer experience where the variable cost feels like a tax. If your architecture is Copilot and your pricing is per-Outcome (Band 4), you are claiming a value capture you cannot measurably defend, and your renewal cycle will surface the gap.
Architecture is upstream of pricing. Pick the architecture, then pick the band that fits.
The reverse-engineering exercise — pricing band first, architecture later — is how teams end up with split-personality products. The split shows up six months later as a renewal where procurement asks “what is this exactly?” and nobody on the vendor side has a clean answer.
The discipline: in every roadmap review, name the architecture first, name the band second. If the band you picked is more than one step away from the band the architecture supports, you have an alignment problem — not a pricing problem.
The consumer × API bifurcation
The other axis the band table doesn’t show: consumer apps and B2B APIs are diverging into structurally different pricing models. Consumer users want predictability and abstraction — they want to know what this costs per month and they reject anything that surfaces tokens. Developers want transparency and granular control — they want to know what each call costs and to optimise at the call level.
The right pricing for consumer is hybrid (Band 2) or value-share-style abstracted credits (Band 5). The right pricing for B2B APIs is token (Band 1) or per-action (Band 3). A product that ships both a consumer UI and a developer API has to run two pricing models in parallel — abstracted for the consumer, metered for the developer. Forcing one model on both sides usually breaks one side.
Apple Intelligence is the most extreme abstracted-value pattern. The user doesn’t pay for AI directly — they pay for the device, and AI is a value-driver inside the device experience. Monetisation through device margin (Apple — Private Cloud Compute) hides the Inference Treadmill (L1-T07) from the customer entirely; Apple optimises the cost behind the scenes. The Salesforce Flex Credits pattern is the enterprise version of the same idea — the customer budgets in credits, the vendor optimises the model mix.
The transition arc itself is what Why Per-Seat Pricing Dies (L1-T09) warned about: per-seat → tiered usage → outcome-based → abstracted-value. Each step is a one-to-two-quarter project. The cost side of the same arc is Inference FinOps (L2-T05) — the four-layer optimisation stack (distillation, 90% caching, 50–70% loop pruning, batching) that determines whether the chosen band actually produces margin. The deepest dive on Band 4 lives in Outcome-Based Pricing Playbook (L2-T09) — attribution paradox, multi-touch problem, and the 5–17% of vendors who actually ship it.
The four-question pricing-band diagnostic
Take any AI feature in your roadmap. Run these four questions in order. Each takes about 90 seconds.
Question 1 — Can you cleanly attribute a unit of value to the AI?
The unit needs three properties: discrete, measurable, valuable. If yes — Bands 3, 4, or 5 are all viable. If no — you are capped at Band 2.
Question 2 — Does the customer agree with you on what the unit is?
If yes — Band 4 (per-outcome) becomes viable, because the measurement framework is shared. If no — you are capped at Band 3 (per-action), where you bill on the work, not the result.
Question 3 — Can you absorb the variance risk?
In Bands 4 and 5, the vendor takes the variance — paid only on success, but on the hook for cost on every action whether successful or not. If you cannot absorb the variance, you are capped at Band 3 (per-action with caps and committed minimums, like Zendesk).
Question 4 — Can you measure outcomes defensibly enough to survive a dispute?
Including: pre-agreed attribution rules, audit trail, dispute-resolution process, a clear contractual definition of what counts. If yes — Band 4 or 5 unlocked. If no — Band 4 or 5 will turn into customer-success-as-legal-disputes.
The combinatorics are simple. Most AI products in 2026 land in Bands 2 or 3. The teams reaching for Bands 4 or 5 should be doing it on the back of measurement infrastructure they have already built, not on the back of a marketing slide.
If you find yourself in Band 1 with a real value-capture story, the right move is almost always to graduate to Band 2 or 3 within two quarters. If you are in Band 2 with the architecture and measurement to support Band 4, the right move is to design the operational scaffolding now and graduate within a year. The graduation is the strategic move. The right pace of graduation is the strategic discipline.
The four traps in pricing-model design
If the four-question diagnostic is the navigation, the four traps below are the rocks the diagnostic is steering you around. Each one is a specific way pricing models break in the wild — and each one has a fix that can be applied before the next renewal cycle.
Trap 1 · Token pricing for non-token products
Pricing by raw consumption when the customer doesn’t think in tokens. The Cursor 2025 case is the canonical example.
Diagnosis: your customers’ procurement teams are asking for “predictability” while you are explaining tokens to them.
The fix: graduate to Band 2 (hybrid with caps) or Band 3 (per-action) within two quarters. The token model is for infrastructure customers; you are a product company.
Trap 2 · Outcome pricing without measurement
The contract says “outcomes.” There’s no operational definition of what an outcome is, no attribution rules, no dispute process. Each customer interaction becomes a legal review.
The pricing model becomes the most expensive customer service incident in the company.
Diagnosis: your “outcome pricing” pilot has more hours of legal review than hours of customer usage.
The fix: rewind to Band 3 (per-action) until the measurement infrastructure exists. Don’t ship Band 4 contracts on a Band 3 measurement substrate.
Trap 3 · Hybrid that’s just per-seat with a tip jar
The platform fee scales linearly with seats. The usage component is decoration. Customers see through this in procurement.
Diagnosis: your Band 2 model produces revenue that maps perfectly to seat count, with usage as <10% of revenue.
The fix: the platform fee should price the platform (governance, integration, support, baseline allocation) at a flat rate per organisation. The usage component should price variable consumption as a real portion of revenue. If usage is <10% of revenue, you are not in Band 2 — you are in Band 1.5 with a marketing slide.
Trap 4 · Skipping bands
Jumping from Band 1 directly to Band 4 or 5 without the operational infrastructure for the rightward bands.
The pricing collapses on the first major customer dispute, and the team has to retreat to Band 2 or 3 — which the customers experience as a price increase even when it isn’t, because the predictability narrative reverses.
Diagnosis: your strategic deck shows the team going straight from token to value-share within four quarters.
The fix: plan for the intermediate bands. Each band has infrastructure debt. Pay it as you graduate. The goal isn’t to skip bands — it’s to be one band ahead of where most competitors are pricing, with the operational substrate to back it up. The full graduation playbook lives in L2-T19: The Outcome-Based Pricing Playbook.
Run the four-question diagnostic on your top AI feature.
For your top AI feature, answer in order:
-
1
Which band are you in today? Token / Hybrid / Per-Action / Per-Outcome / Value Share. Be honest about how revenue actually shows up on the invoice, not how the marketing page describes it.
-
2
Which band best captures the value you create? Most teams under-price by one band. The honest answer here is the strategic move.
-
3
What is the operational gap between today’s band and the right band? Measurement infrastructure, attribution rules, dispute process, contract scaffolding. Name what is missing.
-
4
What is the timeline to close the gap? One quarter? Two? Four? Pick the band you can defend in twelve months, not the band that looks best on a slide.
If your answer to question 1 and question 2 are the same, you are well-aligned. Most teams aren’t. The most common mismatch is “we’re in Band 2; we should be in Band 3.” The second-most-common is “we’re in Band 4; we don’t have the measurement to defend it.”
The most expensive pricing-model mistake is staying in Bands 1–2 when your value capture demands Band 3–4. The second-most-expensive is jumping to Band 4–5 without the measurement infrastructure to support the contract you have signed. You can do this in ten minutes. The decision the diagnostic surfaces — graduate now, build measurement first, retreat one band — is the most consequential pricing decision a PM makes this year.
The sentence to carry
Pricing is not a packaging decision made after the product is built. It’s a product decision that determines what kind of company you become. Token makes you infrastructure. Outcome makes you a service. Pick on purpose.
The frame to take into your next pricing reviewIf you remember one frame from this post, make it that one. The five bands are the structural map. The architecture-pricing alignment from L2-T13 is the discipline that prevents drift. The four-question diagnostic is the move you can make this week, before the next renewal cycle catches you priced one band to the left of where the value lives.
Sources
- Salesforce Agentforce pricing. Salesforce Agentforce pricing page — current Flex Credits ($500 / 100k) and per-conversation ($2.00) pricing.
- Salesforce Agentforce flexible-pricing announcement. Salesforce press release, May 2025 launch reference.
- Intercom Fin outcomes pricing. Fin.ai help docs — official $0.99/outcome documentation.
- Intercom platform pricing. Intercom pricing — base plan and outcome integration.
- Intercom Fin growth analysis. Kyle Poyar, “Intercom’s Bet on AI”, Growth Unhinged — ARR trajectory, support volume share.
- Intercom President interview. GTM Now interview with Archana Agrawal — Fin operational design.
- Zendesk pricing. Zendesk pricing page — current per-AR pricing and free allocations.
- Zendesk per-resolution pricing analysis. Supp.support analysis — context and durability since Aug 2024.
- Notion pricing. Notion pricing — hybrid subscription with usage caps.
- OpenAI pricing. OpenAI ChatGPT pricing — ChatGPT Plus tier hybrid model.
- Anthropic pricing. Anthropic pricing — Claude Pro tier hybrid model.
- How to Price AI Products. Aakash Gupta, Product Growth, February 2026 — pricing-models guide; Cursor $7,225 invoice context.
- Cursor June 2025 pricing change. Cursor blog, June 2025 Pricing — the token-pricing cautionary tale.
- SaaS pricing migration data 2025–2026. Zifeng Liang LinkedIn analysis — hybrid surge, per-seat decline, Gartner forecast referenced.
- Apple — Private Cloud Compute. Apple Security Blog — the canonical device-margin abstracted-value model for consumer AI.
- Andreessen Horowitz — AI Pricing Models. a16z, “The Economic Case for Generative AI” — pricing-archetype framing reference.
- L1-T08: Cost in Every PRD. P90 cost dynamics that re-enter at Bands 3 and 4.
- L1-T09: Why Per-Seat Pricing Dies. The structural why behind Band 2/3 graduation.
- L2-T13: Product Architecture as a Strategy Decision. Architecture-pricing alignment matrix.