AI Agent Pricing Models, Explained: Per-Resolution vs Per-Conversation vs Per-Action

A disclosure up front: Macha (the company publishing this) sells an AI agent layer that bills per action. We'll cover that model like all the others — as one option with real trade-offs, not the obvious winner. The goal here is a guide a buyer can trust, not a sales pitch.

Why the pricing model matters more than the price

Two vendors can quote wildly different "prices" and cost you the same — or quote the same price and cost you triple. That's because the pricing unit decides three things:

• Predictability. Can you forecast next month's bill, or does it swing with traffic, ticket complexity, or how the vendor defines a fuzzy event?
• Incentive alignment. Does the vendor earn more when they help you more, or when they simply do more — regardless of whether customers were helped?
• Where the risk sits. Usage-based models push volume risk onto you; outcome-based models push definitional risk (what counts as a "win") onto a number the vendor controls.

Keep those three lenses handy. Every model below is good at some and bad at others.

The five AI agent pricing models

1. Per-resolution / per-outcome

How it works: You pay a flat fee each time the AI "resolves" an issue end-to-end without a human. This is the headline model of the current AI-agent wave.

Current rates (verify before you sign):

• Intercom Fin — $0.99 per outcome, where an "outcome" is a resolution, a procedure handoff, or a disqualification; "qualifications" (lead-qualification outcomes) are billed at $9.99. You're charged once per conversation even if Fin takes several actions, with a stated minimum of around 50 resolutions/month. (Note: in June 2026 Salesforce signed a definitive agreement to acquire Fin and fold it into Agentforce, so expect this pricing to evolve as that deal closes. See our complete Intercom Fin guide.)
• Zendesk AI agents — roughly $1.50 per automated resolution (about $2 on pay-as-you-go), per current search-cited figures.
• HubSpot Customer Agent — reportedly moved to $0.50 per resolved conversation in 2026.
• Quickchat AI — around $0.50–$0.60 per resolution.
• Sierra — enterprise, outcome-based pricing tied to measurable results; Decagon offers per-resolution as one of its options.

Published per-resolution rates span roughly $0.50 to $2.00 for what's nominally the same unit — a 4x spread — which tells you the "unit" isn't as standardized as it looks.

Pros: Intuitive — you pay for results, not effort. In the cleanest case, the vendor only earns when the AI genuinely solves something, so incentives point the right way: if their model improves from resolving 70% to 80% of tickets, they earn more and you get more value.

Cons: The catch is the definition of "resolution." Most vendors use assumed resolution — if the customer doesn't reply or re-open within some window after the AI's last message, it's counted (and billed) as resolved. A customer who gave up in frustration looks identical to one who got a perfect answer. (This is the same trap that makes deflection rate a vanity metric if you don't watch it.) Cost is also hard to forecast — a viral moment or an outage spikes ticket volume and your bill with it. And the incentive can misalign: a vendor paid per resolution has a quiet reason to count generously.

Who uses it: Intercom Fin, Zendesk, HubSpot, Sierra, Decagon, Quickchat. It's the dominant model for newer AI-native agents.

2. Per-conversation

How it works: You pay for every conversation the AI handles, resolved or not — usually by committing to a volume of conversations up front, with tiered economies of scale.

Who uses it / the notable convert: Ada was an early champion of outcome-based per-resolution pricing, then shifted toward per-conversation and now publicly advocates the conversation-based model. Their argument, roughly: enterprises wanted predictable budgets, and "resolution" is too slippery a unit to bill on honestly. Decagon offers per-conversation as well; Salesforce's per-conversation option sits around $2/conversation.

Pros: Far more predictable than per-resolution — conversation volume is easier to forecast than fuzzy "resolutions," and you're billed on a clean, countable event. It also sidesteps the "what counts as resolved?" fight entirely. The vendor's incentive is neutral: they earn the same whether or not the AI did a great job, so there's no reason to game a definition.

Cons: You pay even when the AI fails and a human has to take over — so a low-quality agent costs you twice (the conversation fee and the agent's time). It rewards the vendor for handling volume, not for solving it, so quality isn't priced in at all. And a "conversation" can itself be ambiguous (does a follow-up the next day count as new?).

3. Per-seat / per-agent (legacy)

How it works: A flat monthly fee per human user — the classic SaaS helpdesk model (~$30–$80+/seat/month), extended to AI as an add-on. Zendesk's Copilot (the agent-assist feature that drafts replies and suggests next steps for human agents) is sold this way, per seat.

Pros: Maximum predictability — you know your bill the moment you count seats, with zero usage risk. Finance loves it. It's genuinely the right fit for agent-assist tools that augment a human who occupies a seat.

Cons: It fundamentally doesn't fit autonomous AI. The entire point of an AI agent is to handle volume without a human in a seat — so pricing it per seat is a category error. You either under-charge (the AI does the work of ten seats you're not paying for) or, more often, the vendor bolts usage fees on top, and you're back to a hybrid. Per-seat also caps nothing about cost-to-serve; it just decouples your bill from the work being done.

Who uses it: Legacy helpdesks for human seats, and agent-assist add-ons (Zendesk Copilot). Rare as the primary model for a fully autonomous agent.

4. Per-action / usage / credits

How it works: You're billed for each automated step the AI takes — drafting a reply, calling an API, tagging, routing, looking something up, resolving — typically via a credit system where different actions (or different underlying models) cost different amounts. This is Macha's model: Macha bills per AI action, framed as automation and orchestration rather than as a packaged "outcome." (Salesforce's Agentforce Flex Credits is the same family: standard actions ~20 credits ≈ $0.10, voice actions ~30 credits ≈ $0.15, sold at ~$500 per 100,000 credits — introduced partly because customers disliked the ambiguity of conversation-based billing.)

Pros: It's the most honest mapping of cost to what the AI actually did — you pay for work performed, full stop, with no fuzzy "was this resolved?" judgment call in the middle. It's transparent and granular: a simple "where's my order?" lookup costs less than a multi-step workflow that touches three systems. And because the vendor bills for steps, not for declaring victory, there's no incentive to over-count resolutions.

Cons: It requires you to understand your usage — a per-action bill is only predictable once you know your action-per-ticket pattern, and a chatty or badly-configured agent can rack up actions. It's less intuitive than "$0.99 when you win," and it pushes the modeling work onto you. Honestly: it rewards transparency but asks more of the buyer up front. (We break down exactly what a Macha credit covers in how Macha credits work.)

Who uses it: Macha, Salesforce Agentforce (Flex Credits), and most platforms that let you build multi-step agentic workflows where "one conversation" can mean very different amounts of work.

5. Hybrid: platform fee + usage

How it works: A fixed platform/subscription fee for access (seats, features, support) plus usage-based charges on top — per resolution, per conversation, or per action. This is where most enterprise deals actually land, whatever the marketing says.

Who uses it: Nearly everyone at scale. Intercom (seats + per-resolution Fin), Zendesk (seats + per-resolution), Microsoft (seat license + Azure-metered agents), and Salesforce (which now openly runs three models at once — per-conversation, Flex Credits, and pre-commit) all fit here.

Pros: Flexible — the platform fee covers fixed costs and the usage component scales with value. It lets a vendor serve both "predictable budget" and "pay for what you use" buyers.

Cons: It's the hardest to compare across vendors, because there are two or three dials and the headline number only shows one. Total cost of ownership is where deals get won and lost — always model the combined bill, not the per-unit rate.

Comparison table

Rates are current as of mid-2026 and change often — confirm on each vendor's pricing page before budgeting.

A worked cost example: same volume, five models

Let's run the same month through each model. Assume a mid-size support team:

• 5,000 customer conversations/month reach the AI.
• The AI resolves 60% of them autonomously (3,000 resolutions); the other 2,000 escalate to humans.
• Each resolved conversation involves, on average, 3 automated actions (understand → look up → reply).
• For per-seat comparison, say you'd otherwise need 5 agent seats at ~$50/seat for the assist tooling.

A few honest observations from this table:

• Per-seat looks absurdly cheap because it isn't buying the same thing — it prices human assistance, not autonomous resolution. Don't compare it head-to-head; it belongs to a different job.
• Per-conversation costs more than per-resolution here because you pay for the 2,000 that escalated too. If your AI resolved a lower share, per-conversation would look even worse — and per-resolution better.
• Per-action can be the cheapest of the autonomous models when tickets are simple (few actions each) — but if your workflows are complex (10+ actions per ticket), it can overtake per-resolution. It depends entirely on your action-per-ticket pattern, which is exactly why this model asks you to understand your usage.
• The "winner" flips with your resolution rate and complexity. There is no universally cheapest model — only the cheapest model for your traffic shape.

The honest take: why per-resolution is contentious

Per-resolution sounds like the obviously fair model — "only pay when it works!" — and it's the most-marketed one for exactly that reason. But it's also the most argued-about, and the criticism is legitimate:

1. "Resolution" is defined by the party who profits from it. Most vendors use assumed/automatic resolution: no reply within a window = resolved. That counts silent abandonment, "I'll just call instead," and confidently-wrong answers as paid wins. The buyer cares about actually solved; the meter measures didn't come back.
2. The incentive can quietly invert. A vendor paid per resolution benefits from a looser resolution definition. The cleanest version of the model (vendor earns only on genuine help) and the real-world version (vendor earns on a generously-counted event) can be far apart.
3. Cost is genuinely unpredictable. Tie your bill to a fuzzy event that scales with traffic and you can't forecast it. This is precisely why Ada publicly moved away from per-resolution toward per-conversation, and why Salesforce built Flex Credits — both citing customer frustration with the ambiguity.

None of this makes per-resolution bad. When the vendor defines resolution conservatively (requires positive confirmation, excludes re-contacts) and you audit it, it's a perfectly good model — arguably the most customer-aligned one. The problem is that "resolution" is doing a lot of unverified work in most contracts. If you choose this model, make them show you exactly how a resolution is counted, and check it against your own CSAT and re-contact data.

Which model is cheapest — and safest — for you

The cheapest model depends on your traffic; the safest depends on your appetite for surprise bills. A buyer's-eye summary:

• You want the most predictable bill: per-seat (for assist) or per-conversation with a committed volume. You'll trade some efficiency for a number finance can plan around.
• You're confident your AI genuinely resolves well and you'll audit the definition: per-resolution can be the most value-aligned — you pay for wins. Just verify the wins are real.
• Your tickets are simple and high-volume, and you want cost to track actual work: per-action/credits usually comes out cheapest and most transparent — provided you model your actions-per-ticket first.
• Your workflows are complex and multi-step: per-resolution can cap your cost per outcome (you pay once no matter how much work it took), whereas per-action exposes you to the step count. Model both.
• You're an enterprise: assume hybrid, and negotiate the platform fee and the usage rate as separate dials. The headline per-unit price is rarely the real cost.

Whatever you pick, do the one thing most buyers skip: run your own numbers through every model (like the table above) using your real resolution rate, conversation volume, and ticket complexity. The model that wins on a vendor's example almost never wins on yours.

Where Macha fits (honestly)

Macha is an AI agent layer that runs on top of your existing helpdesk — Zendesk or Freshdesk — rather than replacing it. It bills per AI action: each automated step the agent takes (understanding a ticket, pulling from your connected knowledge, drafting a reply, tagging, routing, resolving) draws credits, with the amount varying by the action and the underlying model used.

The honest pitch — and the reason we don't bill per "resolution" — is the argument made throughout this article: a "resolution" is a fuzzy, sometimes-fictional event, and charging for one we can't always honestly verify would be exactly the misalignment we just criticized. Outcomes vary with how good your knowledge base is and how messy your tickets are, so we price the work the agent actually does, transparently, and call it automation rather than dressing it up as a guaranteed outcome.

That model's real trade-off, stated plainly: per-action rewards transparency but asks you to understand your usage. If you want a single tidy "pay only when you win" number and you're comfortable trusting a vendor's resolution definition, a per-resolution agent may suit you better. If you want cost that maps to actual work, no incentive to over-count wins, and an agent that layers onto the helpdesk you already run, the per-action model is built for that. You can try it and watch the action-by-action usage yourself — 7-day free trial, no credit card required.

Frequently asked questions

What are the main AI agent pricing models? Five: per-resolution (pay per resolved outcome), per-conversation (pay per conversation handled), per-seat (pay per human user — legacy), per-action/credits (pay per automated step), and hybrid (platform fee plus usage). Most enterprise deployments end up hybrid.

How much do AI customer service agents cost in 2026? It varies by model. Per-resolution runs roughly $0.50–$2.00 per resolution (Intercom Fin $0.99, Zendesk ~$1.50, HubSpot ~$0.50). Per-conversation runs ~$1–$2 each. Per-action models (e.g. Salesforce Flex Credits) price individual steps at ~$0.10. Always model your own volume and complexity — the same traffic can cost very differently across models. Confirm current rates on each vendor's pricing page.

Is per-resolution pricing a good deal? It can be, if "resolution" is defined honestly. The risk is "assumed resolution" — counting a customer who didn't reply (including those who gave up or got a wrong answer) as a paid win. That makes cost unpredictable and can misalign the vendor's incentive toward counting generously. Ask exactly how a resolution is counted and audit it against your CSAT and re-contact data.

Why did Ada move away from per-resolution pricing? Ada was an early per-resolution advocate but shifted toward per-conversation pricing, citing enterprises' need for predictable budgets and the difficulty of honestly defining "resolution." It now advocates the conversation-based model.

What is per-action (credit-based) pricing? You're billed for each automated step the AI takes rather than for a packaged outcome — drafting, looking up, routing, resolving — usually via credits where different actions cost different amounts. Macha and Salesforce Agentforce (Flex Credits) use this. It's the most transparent mapping of cost to work, but it requires understanding your usage to forecast accurately.

Why doesn't per-seat pricing fit autonomous AI agents? Per-seat prices a human occupying a seat. An autonomous agent's whole purpose is to handle volume without a human in a seat, so seat-based pricing is a category mismatch — it either under-prices the work or gets bolted onto a usage fee anyway. It fits agent-assist tools (like Zendesk Copilot) that augment humans, not agents that replace the seat.

The bottom line

There is no single "best" AI agent pricing model — only the model whose risks you can live with for your traffic. Per-resolution is the most marketed and the most contentious: aligned in theory, but only as honest as the vendor's definition of a "resolution." Per-conversation buys predictability at the cost of paying for failures too. Per-seat is the wrong tool for autonomous AI. Per-action is the most transparent mapping of cost to work but asks you to know your usage. And hybrid is where most real deals land, with two dials to negotiate instead of one. Before you sign anything, run your own volume through every model, demand the exact definition behind any "outcome" you're billed for, and weigh predictability, incentive alignment, and where the risk sits. The vendor whose example math looks cheapest is rarely the one that's cheapest for you.

Vendor rates cited were current as of mid-2026 from each company's pricing pages and secondary coverage; several (Zendesk, HubSpot) are search-cited and should be confirmed directly. Pricing in this category changes frequently — next review by December 2026.