What outcome-based AI pricing really costs — and how to budget for it

AI has quietly rewritten the rules of software pricing. For thirty years, business software was sold by the seat: a predictable per-user, per-month licence. That model made sense when the marginal cost of another user was near zero. But an AI agent that resolves a customer's problem consumes real, variable compute every time it works — so the industry has moved toward pricing that tracks value delivered, most visibly per-resolution pricing. It aligns incentives beautifully: you pay when the AI succeeds. It is also genuinely harder to budget for than a seat count. This is a guide to the models, the traps, and how to think about the number.

Why pricing had to change

Two forces broke the per-seat model for AI. First, cost: every resolution runs model inference and retrieval, so usage has a real underlying cost that per-seat pricing ignores. Second, value: the whole promise of an agent is to do work a person used to do, so pricing it like a passive tool undersells it and misaligns incentives. The result is a shift from "pay for access" to "pay for outcomes."

The five models you'll encounter

Almost every AI vendor uses some version of these:

• Per-seat. Predictable and easy to budget, but disconnected from value — you pay whether or not the AI helps anyone. Increasingly rare for autonomous agents.
• Per-conversation / per-session. You pay for every interaction, resolved or not. Simple, but you're charged for failures — the incentives point the wrong way.
• Per-resolution. You pay only when the interaction is solved end to end. Best incentive alignment, but your bill depends on a rate you can't fully predict up front.
• Per-action / credit-based. Each step the agent takes consumes credits. Flexible, but opaque — it's hard to translate "credits" into a forecast, and a single resolution may burn many actions.
• Hybrid (base + usage). A modest platform fee plus usage. This is now the industry standard, because it gives the vendor predictable revenue and the buyer a predictable floor.

Why hybrid won

Hybrid pricing — a base subscription plus per-outcome usage — has become the default because it balances both sides. The buyer gets a predictable base that finance can plan around, plus a variable component that scales with value received. The vendor gets recurring revenue that doesn't collapse in a slow month. If you're evaluating options, a transparent hybrid with a published usage rate is usually the easiest to model and the hardest to be surprised by.

The variable everyone underestimates: your resolution rate

Here is the uncomfortable truth of outcome pricing: your single biggest cost driver is your resolution rate, and you cannot know it precisely until you deploy on your own data, content and traffic mix. A vendor can quote you a per-resolution rate, but your monthly bill is that rate times a number that only reveals itself in production. This is why buyers should model a range, not a point estimate, and revisit it after the first month of real data.

The setup-fee and opacity problem

At the enterprise end, two things make budgeting worse. First, implementation fees: some platforms charge substantial one-time sums — often tens to low-hundreds of thousands — before you resolve a single ticket. Second, opacity: several leading vendors publish no pricing at all, so you can't model cost until you're deep in a sales process. A hidden price and a large setup fee are both signals to scrutinise the total cost of ownership, not just the sticker rate.

What "a resolution" should actually mean

Because the unit is the money, the definition of the unit is everything. A fair, unambiguous per-resolution model should specify:

• One charge per resolved conversation, even if it took several steps — never per message or per action.
• No charge for failures — if the agent hands off to a human or doesn't resolve, you don't pay.
• No charge for testing and internal traffic.
• A clear definition of "resolved" — the customer's question was answered from your data, or the requested action was completed.

If a vendor can't state these plainly, that ambiguity will show up on your invoice.

A checklist for evaluating any AI quote

• What exactly is the billable unit, and is it defined in writing?
• Am I charged when the AI fails or hands off?
• Is there a setup or implementation fee? How much?
• Is the rate published, or quote-only? Can I model it before signing?
• What's the overage behaviour — hard stop, alert, or silent charges?
• Are there per-seat fees hiding underneath the usage price?
• How are different channels (e.g. voice) metered?

A worked example

Take a transparent hybrid: a flat monthly platform fee plus a small per-resolution charge. Your bill is simply base + (rate × resolutions) — fully predictable once you know your volume. Because the base is fixed, your effective cost per resolution falls as volume rises: at low volume the base dominates; at higher volume it amortises away and you converge toward the marginal rate. That's the opposite of per-conversation models, where cost rises linearly with every interaction whether it succeeded or not.

Benchmark against the real alternative: a person

The honest comparison isn't vendor-to-vendor — it's AI-to-human. A fully-loaded support agent (salary, benefits, tooling, management, facilities) costs well over a hundred dollars a day. Spread across the interactions they can actually resolve, that's on the order of a few dollars per resolved interaction. A per-resolution rate measured in cents is a different order of magnitude — which is why the ROI conversation should always anchor on cost per resolution, human versus AI, not on the monthly subscription in isolation.

Budgeting in practice

• Estimate volume first. Pull your resolvable interaction count by month and channel.
• Model a range. Assume a conservative and an optimistic resolution rate and compute both bills.
• Set a cap or alert. Agree a monthly spend alert so usage never surprises finance.
• Revisit after 30 days. Replace your assumed resolution rate with the real one and re-forecast.

Red flags in AI pricing

• Being charged per message rather than per outcome.
• Paying for conversations the AI failed to resolve.
• Credit or action systems you can't translate into a forecast.
• Large setup fees bundled with a "call us" price.
• Overages that bill silently with no cap or alert.

Negotiation levers worth knowing

Even with published pricing, there's room to shape a deal: annual commitments in exchange for a lower effective rate; volume tiers that reduce the marginal rate as you scale; and price protection so your rate is locked for the term. If you're heading toward high volume, an enterprise agreement with volume discounts usually beats pay-as-you-go.

Where outcome pricing gets nuanced

Per-resolution isn't a universal unit. Voice is the clearest exception: a phone call's cost is driven by minutes (speech-to-text, the model, text-to-speech, telephony), so voice is typically metered per minute on top of the platform fee. And at very low volume, a flat base fee makes the effective per-resolution cost high — outcome pricing rewards real usage, so it fits teams past a modest threshold better than tiny pilots.

Outcome-based pricing is genuinely good for buyers — when it's transparent and predictable. Demand a clear unit, a published rate you can model, no charge for failures, and no surprise setup fee. Then the model does exactly what it promises: you pay for results.