How much an AI agent costs, and the five ways agent products charge

Hockeystick's published pricing starts at $49 a month, while Intercom Fin starts at $0.99 per outcome. These are different purchases with different limits. An AI agent's total bill also depends on model usage and any hosting or upkeep you handle yourself, so compare the finished work each allowance buys.

For a founder, the useful comparison is the bill for a meeting brief you can trust or a follow-up that's ready to approve. A low entry price tells you little until you know what consumes the allowance and who fixes the connection when it breaks.

What an AI agent costs today

Published entry prices describe different purchases, so compare the billing unit alongside the price. We checked the vendor pages below on November 1, 2026; annual commitments and usage limits matter as much as the headline.

Billing model	Live example and source	What to check
Per seat	Lindy pricing starts paid usage at $49.99 per active user per month, with 5,000 credits per seat pooled across the workspace.	Who becomes a billable user and how much work the pool covers.
Per credit	Lindy pricing lists a $10 top-up for 1,000 credits, with consumption varying by job.	The measured cost of your own jobs, plus the access fee.
Per task	Zapier pricing lists Professional starting at $19.99 a month; Zapier's task guide identifies that starting price as billed annually.	Which actions count and whether AI steps consume extra tasks.
Per outcome	Intercom pricing lists Fin from $0.99 per outcome.	The outcome definition, minimum commitment and any helpdesk seat costs.
Flat plan	Hockeystick pricing lists Personal at $49 a month with $10 of included AI, or Team at $199 a month flat for unlimited people with $100 of included AI.	The included AI limit and which people can use the agent.

These categories overlap. Lindy combines seats and a shared usage pool, while Intercom combines helpdesk seats and Fin outcomes when you buy its integrated platform. A pricing label alone won't describe your invoice.

Separate a trial allowance from an ongoing free offer. A trial can be useful for testing your actual records, but its balance and expiration may make it unsuitable for estimating steady monthly costs. Save the billing settings you used during evaluation so a later plan change doesn't invalidate the comparison.

This guide covers buying and running existing products. If you need custom engineering, our separate guide to the cost to build AI agents explains the work a build quote should cover.

How per-seat agent pricing works

Per-seat pricing ties the access bill to the people using the product. It suits a company where each paid user needs direct access, provided the usage allowance matches their work.

Lindy's pricing FAQ says every active user is billed, including someone who uses Lindy through a Slack mention. The same page says new teammates joining through Slack receive a trial before billing, while direct signups are billed immediately. An admin introducing the tool should understand that enrollment path before inviting the company.

For illustration, Lindy's published entry rate gives this access calculation: 4 paid users x $49.99 = $199.96 a month, before any top-ups, using Lindy's pricing. That calculation describes a selected team size, rather than an observed customer bill. It also leaves out how quickly those users consume their pooled allowance.

Seats can be a sensible purchase when teammates regularly ask questions and complete their own work. A company with occasional readers should check whether those readers need paid access, whether guests exist and whether an administrator can remove seats without losing shared records.

How credit-based agent pricing works

Credit-based pricing gives different jobs different weights inside a usage balance. You need a measured example of your workflow to translate that balance into useful output.

Lindy's pricing describes everyday asks as consuming 2-250 credits and deep work as consuming 250-1,000 credits. Those are vendor ranges across categories, so they can't establish the price of your meeting brief. Model choice and the size of the work can change consumption within the same broad category.

Ask the vendor to run representative work with your inputs, then inspect the balance before and after. Include an ordinary request and a difficult request with a long source document. Keep rejected drafts in the measurement, because your company still pays for work it can't use when the vendor meters consumption.

Expiration changes the economics too. Lindy's pricing says monthly credits don't roll over, while purchased top-ups remain until used. It also says credit-using actions pause when the pool runs low, with admins able to top up or wait for the reset.

This model fits variable work when the usage history makes the bill understandable. Be cautious when a vendor will explain the exchange rate but won't show what your finished workflow consumed.

How task pricing differs from paying for a whole workflow

Task pricing usually meters actions inside the workflow, so completing the business job can consume several tasks. Estimate the path the workflow takes before comparing allowances.

Zapier's task documentation says successful action steps count toward usage, while triggers, filters and paths don't. The same documentation says some products have different task rates and that replaying previously successful steps can consume tasks again. A long diagram therefore needs a step-by-step estimate.

AI steps require particular attention. Zapier's pricing page says the model tier affects task usage, and its task documentation distinguishes Zap workflows from Zapier Agents, which uses a separate activity quota. Don't apply a workflow task allowance to the Agents product without checking that product's rules.

A workflow execution is another possible billing unit. n8n pricing lists Starter at €20 a month billed annually, with 2,500 workflow executions and unlimited steps. That is a euro-denominated offer with an annual commitment, so it shouldn't appear beside a monthly dollar subscription as an interchangeable price.

Choose Zapier when a defined sequence across its supported apps solves the job and someone can maintain that sequence. Choose n8n when you want control over the workflow and can own its design. Both deserve evaluation on the actual path your data follows and the usage rules attached to that path.

How outcome pricing works

Outcome pricing charges against a vendor's definition of a completed result. Read that definition before assuming it matches your company's definition of success.

Intercom's Fin pricing FAQ says an outcome can count when a customer confirms resolution, stops asking for help after Fin responds, or Fin completes a Procedure, including a handoff. It also says the conversation is charged once even when multiple questions are answered. A handoff qualifying for billing is especially important when your internal scoreboard measures issues fully resolved by AI.

Using Intercom's published starting rate, an illustrative 100 billable outcomes x $0.99 = $99 in Fin usage, before other applicable charges, per Intercom pricing. The same page says using Fin with an existing helpdesk has a minimum monthly commitment. Buying Intercom's helpdesk brings seat charges as well.

Fin is worth considering for a support team that wants customer-service software with a defined outcome meter. A founder seeking internal meeting preparation is buying a different job, so Fin's unit price alone can't establish whether the internal assistant is expensive.

Agree internally on how you'll count reopened issues and human follow-up. Compare that record with the vendor's billable outcomes before treating the pricing model as evidence of savings.

What a flat agent plan includes

A flat plan makes the access fee predictable, while its AI allowance still limits consumption. Read what happens at that limit before assigning unattended work.

Hockeystick's pricing includes hosting and a company brain, with an agent that works from those records. On Team, everyone can read and ask the shared brain and talk in its threads; the agent itself works for the owner. A shared knowledge space and a personal agent for every teammate are different purchases.

The same pricing page says included AI refills monthly and unused AI doesn't roll over. When it runs out, the agent pauses until the refill unless you connect your own AI key. Your provider bills usage on that key separately, and Ask continues to use included AI.

We fit a founder who wants an agent working from company records without a server to maintain. The knowledge can move out of someone's head into notes with sources, with approval required before the agent sends, posts, deletes or spends. Scheduled jobs can't take those actions, according to our live homepage.

Check connections before buying any hosted assistant. Our applications directory distinguishes requested connections from apps we've tested, so a name in the directory doesn't establish a working integration. Bring the exact job and the tools it touches to the setup conversation.

What sits outside the subscription price

Model usage and operating work can remain separate even when the software is free or the hosting fee is fixed. Ask which party owns each dependency before comparing totals.

For a developer paying directly for model usage, Anthropic's pricing lists legacy Sonnet 4.6 at $3 per million input tokens and $15 per million output tokens. An illustrative workload of 1 million input tokens plus 100,000 output tokens therefore costs $3 + $1.50 = $4.50 at Anthropic's published rates. This is a raw token subtotal, excluding caching changes, tools, hosting and other applicable charges.

That subtotal doesn't tell you how much a customer follow-up costs. An agent can read history and call tools repeatedly before producing the final draft. Use the provider's usage record to measure the complete attempt, including revisions, instead of pricing only the visible answer.

If you self-host, add the machine or server bill and assign someone to updates and failed connections. Self-hosting can be the right choice when an operator already owns that work and wants direct control. A hosted product is worth considering when keeping the agent available would otherwise become another responsibility for the founder.

Maintenance also includes checking whether the result stays correct after a source changes. A workflow that keeps using an old contract can run successfully while producing the wrong answer. Decide who notices that failure and how the agent will be stopped while it gets fixed.

How to compare the total bill before committing

Compare vendors against the same finished work, then add the human effort that remains. A meeting brief is a useful test when it requires current calendar details and a prior customer decision with its source.

Write a definition of done before the trial starts. For that brief, specify the source records it must consult and what it should do when those records disagree. Count a missing source as a failed result even when the prose sounds polished.

Keep a record with the billed unit and whether you accepted the output. Add time spent correcting the work and restoring broken connections. This separates the provider's meter from the result your company needed.

Treat permission failures as part of that test. A brief that works only when the founder shares a private inbox with everyone may be unacceptable regardless of its price. Ask the vendor to demonstrate the intended access boundaries using the same roles your company would use after purchase.

Check the exit path while the purchase is still optional. Identify which source records and workflow instructions you can take with you, then find out who can remove app access when the trial ends. Keeping an export of the work you've approved makes a later migration easier to evaluate.

Review the limit behavior before enabling recurring jobs. Record whether the product pauses or buys additional usage, who gets notified and who can change the spending setting. A predictable ceiling may be more useful than a lower entry price with an unfamiliar overage policy.

Start with a narrow job whose inputs you understand and whose result someone will review. Expand after the usage record and accepted output justify it. The right agent price is one your company can connect to completed work and keep paying when the trial novelty wears off.