AI Agent Platform Pricing: The Real 2026 Cost Math

Open any AI agent platform’s pricing page and you’ll learn almost nothing. You’ll see three or four plan names, a column of checkmarks, and a “Contact Sales” button that appears right where the number should be. That’s not an oversight — it’s the business model. When pricing is opaque, you can’t build a spreadsheet, and without a spreadsheet you can’t say no. Whether you’re comparing the top AI agent platforms on paper or trying to work out which is the best AI agent platform for your actual support volume, the sticker price is never the real number.

This article builds that spreadsheet. We’re going to work through the six ways agent vendors actually meter usage, show what the named platforms publish and what they hide, and then go one layer deeper than any vendor comparison: what a grounded support conversation actually costs in raw LLM tokens — the number everything else is priced against. Disclosure up front: we build AI Chat Agent, a self-hosted chat widget that sits at one extreme of this cost curve — one-time licence, bring-your-own API key. The numbers below are cited to public pricing pages and Anthropic’s published rates, not to our own conclusions. If you’re also weighing skipping vendors entirely, we priced out what building it yourself costs in a companion piece.

Why AI agent platform pricing is hard to compare

Ask five AI agent software vendors for a quote and you’ll get five different units of measurement. One prices per seat. Another prices per resolution. A third sells credits that different models burn at different rates. A fourth counts executions. A fifth just says “talk to sales” and comes back with a number shaped by your company’s LinkedIn headcount. None of these numbers convert into each other cleanly, which is the point — a unit mismatch is much harder to shop against than a bad number.

The contact-sales wall does real work here. Enterprise platforms like Sierra and Decagon don’t publish self-serve pricing at all; deal size is negotiated against your support volume, your industry, and how badly the vendor wants the logo for their own case studies. That’s not necessarily a scam — outcome-based deals can make sense at enterprise scale — but it means you cannot build a like-for-like table without picking up the phone, and the vendor knows it.

Even where prices are published, they describe different things. Intercom Fin’s $0.99 per resolution looks comparable to Salesforce Agentforce’s roughly $2 per resolution, until you check what counts as a “resolution” at each vendor — defined by the vendor, measured by the vendor, rarely audited by anyone else. A side-by-side table of headline numbers is only honest if the units are honest first. They usually aren’t, so most published comparisons end up comparing apples to units of a different fruit entirely.

The six pricing models

Every AI agent platform prices itself with one of six meters, sometimes stacked on top of each other. Here’s how each one actually works, who it favours, and where it breaks.

Per-seat

You pay for a named human licence — an agent, an admin, an editor — regardless of how much the AI does. Intercom’s entry plan starts from $29/seat/month billed annually, Zendesk’s suite runs roughly $19 to $169 per agent/month depending on tier, and Voiceflow and Lindy layer per-user pricing under their credit systems. Seats favour teams with stable headcount and punish teams that scale support volume without scaling headcount — precisely the scenario AI agents are supposed to enable. You end up paying human-shaped prices for AI-shaped work.

Per-resolution

You pay only when the bot closes a conversation without escalating to a human — Intercom Fin at $0.99/resolution, Salesforce Agentforce commonly reported around $2/resolution, sold in pre-purchased blocks. It favours vendors, because “resolution” is a metric they define and grade themselves, and it favours buyers only when volume is genuinely unpredictable and low. It breaks at scale, because the price per unit never drops.

Per-conversation

Older live-chat platforms and some bot builders meter every conversational thread the AI touches, resolved or not. It’s simpler to forecast than per-resolution, but gameable in the other direction — a vendor benefits from you having more, shorter conversations, not fewer, better ones. Most 2026-era platforms have moved away from pure per-conversation billing toward resolution or credit metering, because a raw thread count is too easy to inflate through UI friction.

Per-credit / per-action

Chatbase, Voiceflow, and Lindy/Relevance AI all sell a monthly credit allowance, then charge more credits for “complex” steps — a call to a premium model, a multi-step tool action, a long context window. Premium models consume several credits per reply, so effective message volume on the same plan can drop sharply the moment you upgrade model quality. The credit is a variable unit dressed up as a stable one.

Per-execution / per-task

Automation platforms like n8n (one execution equals one workflow run) and Zapier (task-based, roughly 100 free per month, scaling from there) meter the workflow layer, not the conversation layer. LLM calls inside that workflow are billed separately by your own model provider — neither platform marks up your tokens. This is the most transparent of the six models, but it only covers orchestration; you still budget model spend on top.

Per-token

The rawest meter: you pay your model provider directly for input and output tokens. Anthropic’s current rates: Claude Opus 5 at $5/$25 per million input/output tokens, Claude Sonnet 5 at $3/$15 (introductory $2/$10 through 31 August 2026), and Claude Haiku 4.5 at $1/$5. Every other model on this list is a markup built on top of this one — which is why we use it as the baseline later in this article.

What the platforms actually charge

Here’s what AI agent providers actually publish versus what requires a sales call, as of today.

Platform	Pricing model	Published rate	What’s hidden
Intercom Fin	Per resolution	$0.99/resolution	No separate Fin platform fee, but Fin sits on top of Intercom seats — see our Intercom comparison
Intercom (seats)	Per seat	From $29/seat/month, annual entry plan	AI add-ons priced separately from seats
Zendesk	Per seat + resolution	Seats ~$19–$169/agent/month by tier	Per-resolution AI rate not published; third-party teardowns estimate $1.20–$1.50/resolution
Salesforce Agentforce	Per resolution (pre-purchased blocks)	Commonly reported around $2/resolution	Requires an underlying Service Cloud licence
Sierra / Decagon	Outcome-based, enterprise	No self-serve pricing	Publicly reported first-year deals run into six figures
Chatbase	Credit / message tiers	Tiered plans	Premium models consume several credits per reply, cutting effective volume
Voiceflow	Per-editor seat + credits	Seat-based with a monthly credit allowance	Complex AI steps consume multiple credits per response
Lindy / Relevance AI	Per-user + usage credits	Subscription plus credits	Credit burn rate scales with task complexity
n8n	Per execution	Free self-hosted community edition; cloud from ~€24/month	LLM calls billed separately by your provider
Zapier	Per task	Free tier ~100 tasks/month, paid tiers scale up	Agent features priced as a separate add-on

The pattern: platforms selling to enterprises with procurement teams hide the number. Platforms selling to solo builders and marketers publish it, because self-serve tools that can’t be checked out with a credit card don’t survive. Zendesk sits in an odd middle ground — seats are published, the AI layer is not, which is why the comparisons that matter most for Zendesk shoppers are the ones third parties have to reverse-engineer.

If you want help narrowing down which platform actually fits your team beyond the sticker price, that’s a separate exercise — see our buyer’s guide to AI agent tools and our agentic AI platforms breakdown. This piece stays on the number.

The hidden cost lines

The headline rate is never the bill. Nine line items regularly show up after the contract is signed:

• Seat minimums. Agentforce still needs a Service Cloud licence underneath, Fin still needs an Intercom plan — the AI price is additive, not a replacement.
• Annual commitments. Entry-level rates like Intercom’s from-$29 seat price are usually the annual-billing number. Month-to-month costs more, and cancelling mid-term rarely gets a refund.
• Overage multipliers. Go past your resolution block or credit allotment and the marginal unit price often jumps — the block exists so the vendor gets paid upfront and you get penalised for underestimating.
• Implementation and onboarding fees. Enterprise, outcome-based deployments commonly bundle in professional services — the reported six-figure first-year number often includes setup, not just usage.
• Premium-model multipliers. Switching a credit-based bot to a stronger model doesn’t change the monthly fee — it changes how fast credits disappear, several per reply instead of one.
• SSO and enterprise gating. Single sign-on, audit logs, and advanced permissioning are routinely locked behind the top tier, regardless of how many seats you actually need.
• Data-retention add-ons. Longer conversation history, for compliance or QA, is frequently a paid extra rather than a default.
• Connector surcharges. Pulling data from your CRM, ticketing system, or a private API sometimes carries its own per-connector fee.
• Human escalation cost. Every resolution rate implies a non-resolution rate — those conversations still need a human, and that seat and that time never show up on the AI pricing page.

None of these show up in the pricing page’s first screenshot. All of them show up in the invoice by month three. The only cost-per-conversation number worth trusting is the one you measure in a pilot on your own traffic mix.

Unit economics at 1k, 10k, and 50k conversations

Here’s where per-resolution pricing and a self-hosted licence diverge, worked out at three volumes. Assume, generously, that 70% of conversations resolve without a human handoff — a rate vendors like to imply in marketing and rarely put in a contract. The self-hosted column uses Claude Sonnet 5 token costs (derived in the next section at roughly $0.048 per grounded conversation) plus a modest VPS, and treats AI Chat Agent’s one-time €79 licence as noise after month one. The core question is simple: does your cost per resolution fall as you grow, or stay flat forever?

Monthly conversations	Resolutions billed (@70%)	Per-resolution vendor cost (@$0.99)	Self-hosted: tokens + VPS	Self-hosted cost/conversation
1,000	700	$693	$48 tokens + ~$20 VPS = $68	~$0.068
10,000	7,000	$6,930	$480 tokens + ~$30 VPS = $510	~$0.051
50,000	35,000	$34,650	$2,400 tokens + ~$60 VPS = $2,460	~$0.049

Two curves, two shapes. The per-resolution line is linear — $0.99 times resolutions, forever, with no volume discount built into the public rate. The self-hosted line is a fixed cost amortised over a growing number of conversations, so its per-unit cost falls as volume rises and flattens toward the pure token rate. At 1,000 conversations a month the gap is real but survivable — $693 versus $68. At 50,000, it’s $34,650 versus roughly $2,460, a difference that funds a hire. Note the two columns aren’t measuring the same thing: the vendor bills only the 70% it resolves, while the self-hosted figure covers every conversation including the ones a human finishes. This is the entire argument in self-hosted versus SaaS chatbots reduced to a single table.

The raw cost floor: what the tokens actually cost

Every price in this article — $0.99 per resolution, $2 per resolution, a bucket of credits — is a markup on top of a floor anyone can calculate: the LLM API cost of actually answering the question. Here’s the maths for one grounded support conversation, meaning the bot retrieves relevant knowledge-base chunks and answers from them rather than from parametric memory alone.

A typical exchange sends the model a system prompt, retrieved context, and recent conversation history as input, and gets back a written answer as output. Call it roughly 2,500 input tokens and 300 output tokens per turn, over an average four-turn conversation:

Per grounded support conversation (~4 exchanges):
Input tokens:   ~2,500/turn × 4 turns = 10,000 tokens
Output tokens:  ~300/turn × 4 turns   =  1,200 tokens
Claude Sonnet 5 ($3 / $15 per million input/output):
Input:  10,000 × $3  / 1,000,000 = $0.030
Output:  1,200 × $15 / 1,000,000 = $0.018
Total per conversation          = $0.048
Same conversation on other Anthropic models:
Claude Haiku 4.5 ($1 / $5 per million):  ≈ $0.016
Claude Opus 5    ($5 / $25 per million): ≈ $0.080

Even the premium model, Opus 5, lands at eight cents. The mid-tier model most teams would actually run in production, Sonnet 5, lands under five cents — call it twenty times cheaper than Intercom Fin’s published rate and roughly forty times cheaper than the commonly reported Agentforce rate. OpenAI and Google publish comparable per-token rates across their own tiers; check their pricing pages directly rather than trust a number repeated secondhand.

Two costs sit on top of this and don’t change the order of magnitude. Embedding a knowledge base for retrieval is a one-off job — for a typical help centre it costs cents to a few dollars, and you only pay it again when the content changes, not per conversation. And the compute to run retrieval and chat logic needs a server, not a fleet: a modest VPS handles this workload because the LLM call, not local compute, does the heavy lifting. Tokens are the floor. Everything a vendor charges above roughly five cents per conversation is margin, infrastructure, sales cost, or all three.

Why per-resolution pricing breaks your budget

Per-resolution pricing has one structural problem that no amount of vendor transparency fixes: you cannot forecast it. Your bill is a function of how many people talk to your bot, which is a function of traffic, seasonality, a product launch going well, or a product launch going badly and generating a support spike exactly when your budget is tightest. A finance team can model a seat count. It cannot reliably model how many people will need help next quarter, because nobody can, three months out.

There’s a subtler problem underneath the forecasting one: the incentive is backwards. You pay more precisely when the bot succeeds. A well-tuned bot that resolves 80% of conversations costs more per month than a badly-tuned one that resolves 40% and pushes the rest to unpaid human agents. The vendor’s revenue rises exactly when their product works best — a fine model for the vendor, and an uncomfortable one to explain to a CFO who just asked why the AI bill jumped 30% the same month the bot got better.

What to model instead of the vendor’s own resolution-rate promise: your historical conversation volume, not resolutions; a range of plausible resolution rates rather than a single optimistic one; and the cost at the top of that range, not the bottom. Budget for the worst case a per-resolution vendor can hand you, because that’s the one you’ll actually be billed.

None of this means per-resolution pricing is wrong for everyone. It’s a defensible choice for high-ACV enterprise deals where a single resolved conversation might be worth thousands of dollars in retained revenue, or for teams with zero engineering capacity who need a live bot next week and are willing to pay for that speed and for someone else’s on-call rotation. The model fits a specific buyer. It’s just not the buyer running high-volume, low-ACV support who’s counting every conversation.

Where a one-time licence changes the maths

AI Chat Agent takes a different position on this curve entirely: €79 once, self-hosted, no monthly vendor fee and no per-resolution fee ever. You bring your own API key — OpenAI, Anthropic, Google Gemini, OpenRouter, or any OpenAI-compatible endpoint including Groq, Ollama, or a self-hosted model — and pay your model provider directly at the rates in the table above. Switching providers doesn’t require a data migration; the knowledge base and conversation history stay put.

The retrieval layer is built to avoid the two failure modes that make grounded bots expensive to get wrong: hallucinating answers, costly in support tickets and trust, and missing relevant context, costly in escalations. It fuses dense vector search with lexical full-text search, rewrites the query, runs every candidate through an LLM reranker that gates relevance before anything reaches the model, and expands matches with neighbouring chunks so answers don’t get cut off mid-thought. Markdown-aware chunking keeps code blocks and tables intact instead of shredding them across chunk boundaries. When the knowledge base doesn’t cover a question, grounding instructions tell the bot to say so rather than invent an answer — which matters more to support quality than any pricing model does.

Beyond retrieval: multi-bot management for agencies running several clients off one licence, human takeover mid-conversation with automatic release back to the AI, white-label branding, UTM and visitor-identity capture for attribution, and a 26.4 KB widget that doesn’t bloat page load. Five services, one Docker Compose command, 1,646 automated tests behind it.

What it is not: a general autonomous-agent orchestration framework. There’s no voice channel, no visual multi-step tool-calling workflow builder, no subagent orchestration across a dozen SaaS integrations. If you need an agent that books a flight, files an expense report, and posts a Slack update across a chain of tools, this is the wrong product for that job. AI Chat Agent is a grounded support and lead-capture widget that happens to remove the per-resolution meter entirely. That’s a narrower job, done at a cost structure the rest of this article should have made obvious by now.

A three-year TCO worksheet

Everything above collapses into one exercise: build your own three-year total cost of ownership, using your own volume, not the vendor’s demo numbers. These are the line items to actually fill in — the AI agent TCO maths that pricing pages skip:

1. Base licence or subscription fee — annual seat cost × seats, or one-time licence fee, for years one through three.
2. Usage fees — resolutions, credits, executions, or tokens, projected at your actual conversation volume across a realistic range, not the vendor’s best case.
3. Overage risk — what happens, in dollars, if you exceed your committed block or credit allowance by 20%, 50%, 100%.
4. Model/token spend — if you’re bringing your own key, your provider’s per-token rate at your projected volume.
5. Implementation and onboarding — one-time, often the largest line item on enterprise outcome-based deals.
6. Infrastructure — VPS or hosting cost if self-hosted; effectively zero marginal infrastructure if fully SaaS.
7. Enterprise add-ons — SSO, audit logs, data retention, connector fees, priced at the tier you’ll actually need, not the tier you start on.
8. Human escalation cost — the seat-hours of agents handling everything the AI doesn’t resolve, at your current resolution-rate assumption.
9. Internal engineering time — setup, integration, and ongoing maintenance, valued at a real internal hourly rate, not zero.
10. Switching cost — what it costs in time and data migration if you need to leave this vendor in year two.

Run each line for year one, two, and three, and watch where the totals cross. Per-resolution and per-credit platforms tend to scale their bill in a straight line with your growth. Self-hosted and per-token models tend to front-load cost in year one and flatten hard afterward. Which shape wins depends entirely on your volume curve — which is exactly why the worksheet, not the pricing page, should drive the decision.

If you want to see how AI Chat Agent handles this cost curve in person, the live demo opens the full admin panel with no registration, and the retrieval, reranking, and lead-capture flow described above is running exactly as described. If the maths already answered your question, grab the €79 licence and skip next quarter’s per-resolution invoice entirely. And if you’re still comparing options, the rest of the pricing and platform breakdowns are on the blog.

Frequently Asked Questions