AI Agent Pricing: How to Charge for Software That Does the Work

In October 2024, Salesforce launched its AI agents, Agentforce, at $2 per conversation.

Simple. Easy to explain. One number on the slide.

It lasted about seven months.

It priced out smaller organizations and nonprofits, and it felt expensive for a quick question. So in May 2025 Salesforce added Flex Credits at $0.10 per action. Then per-user add-ons for staff-facing agents. Then a fixed-price, all-you-can-eat contract called an AELA, which grew from 16 to more than 120 deals in a single quarter. Then a $2-per-resolution Help Agent in June 2026, with "no charge for escalations". Then Core, Advanced and Max editions in September 2026.

One Salesforce executive said European customers were "afraid of all these metrics" (Pricing Innovation, 6 October 2026).

Now, if one of the biggest software companies on earth needed two years and five meters to price an agent, what hope do the rest of us have?

Plenty, actually. Salesforce did the expensive experimenting for you. In this guide I'll show you the six ways agents are priced right now, with real price points, a worked margin example, and a simple framework for choosing yours.

Prices checked 27 October 2026. They change often, so see the AI Pricing Index for the latest.

Why Is Pricing an AI Agent So Hard?

A normal software subscription charges for access. You get a login, you pay $20 a month, and it doesn't matter much whether you use it once or every day.

An agent is different. It doesn't help a person do the work. It does the work.

That breaks pricing in three ways.

1. Seats shrink. If your support agent resolves half the tickets, your customer needs fewer human agents. Charge per seat and your revenue falls as your product gets better. That's exactly why Intercom moved Fin to a price per resolution (I cover the full story in the Fin case study).

2. Your cost per task swings wildly. A simple question might cost you a cent in model fees. A long research task might cost a dollar. Third-party estimates put Manus tasks at anywhere from 10 credits for a simple chat to 1,500 or more for "Wide Research". That's a 150x range on one product.

3. The customer can't see the work. With a seat, the customer knows what they bought. With an agent, they see a bill and wonder what an "action" or "credit" actually was.

Remember Procore from Chapter Four of The Subscription Playbook? Its pricing page promised "unlimited" users, and I asked the obvious question: there must be some limits to unlimited, right?

Agents bring that question back with a vengeance. Every agent task has a real cost. So every agent price has to answer two questions at once: what does the customer think they're paying for, and does it cover what each task costs you?

How Do You Price AI Agents? The Six Models

Here are the six models in use today. Most vendors now run two or three at once (ICONIQ's 2026 survey found AI companies blend an average of 1.7 pricing models).

Model #1: Per seat (the agent rides on a human license)

You charge per user per month, and the agent is part of the package. Microsoft is the champion of this approach. Microsoft 365 Copilot costs $30 per user per month for enterprises, and Copilot Business costs $21 for smaller firms (with an $18 promotion running to 31 December 2026).

Works when: the agent helps a person who still has a job to do, like drafting, searching or summarizing.

Breaks when: the agent replaces the person.

Model #2: Per conversation (or per task)

A flat fee each time the agent handles something, whether it succeeds or not. This was Agentforce's original $2 meter. Decagon offers it too, with a reported rate of around $0.99, and says "the vast majority of our customers choose per-conversation pricing."

Works when: buyers want to forecast spend from a number they already know, like monthly ticket volume.

Breaks when: tasks vary a lot in size, so a quick "where's my order?" costs the same as a 20-minute refund dispute.

Model #3: Per action or credit

You publish a rate card. Each kind of work costs a set number of credits. Salesforce Flex Credits cost $500 per 100,000, and a standard action uses 20 credits, which is $0.10. Microsoft's Copilot Studio charges $0.01 per credit on pay-as-you-go, with 1 credit for a classic answer, 2 for a generative answer and 5 for an agent action. HubSpot credits are $10 per 1,000.

Works when: the agent does many different jobs and you need one currency to price them all. (I go deeper on this in the guide to credit-based pricing.)

Breaks when: customers can't translate credits into anything they understand.

Model #4: Per minute

Voice agents are priced like phone calls. ElevenLabs charges $0.08 per agent minute on every plan, and bigger plans buy more concurrent calls rather than cheaper minutes. Deepgram's Voice Agent API is $0.075 a minute.

Works when: time is a fair proxy for work, like calls.

Breaks when: a slow agent earns more than a fast one. (That's a strange incentive to give your engineers.)

Model #5: Per outcome

The customer pays only when the agent achieves a defined result. Fin charges $0.99 per outcome. Zendesk charges $1.50 per automated resolution on a commitment or $2.00 pay-as-you-go, beyond the 5 to 10 included per agent seat. HubSpot's Customer Agent costs $0.50 per resolved conversation. Salesforce's Help Agent is $2 per resolution.

Works when: the result can be counted, verified and tied to money the customer already tracks. I've written a full guide to outcome-based pricing and a list of outcome-based pricing examples.

Breaks when: nobody can agree on what counts as "done".

Model #6: Hybrid (platform fee plus usage)

A fixed fee for access, plus one of the meters above. Decagon reportedly charges a platform fee of around $50,000 a year plus conversations or resolutions. Anthropic's Claude Enterprise is now a published $20 per seat plus usage at API rates. Glean pairs a seat with pooled "FlexCredits" for agent runs.

Works when: you sell to businesses that want a budget line and you want a floor under your revenue. This is where most of the market is heading, and I've covered it in the guide to hybrid pricing.

AI Agent Pricing Compared: Real Prices in October 2026

Here's how the main agent vendors charge today.

Prices checked 27 October 2026; they change often. See the AI Pricing Index. Figures marked "reported" or "estimate" come from third parties, not the vendor.

One corporate note: according to MarTech, Salesforce agreed to buy Fin for about $3.6 billion on 15 June 2026, and Pricing Innovation reports the deal closed on 10 September 2026. So two very different pricing philosophies (Agentforce's many meters and Fin's single $0.99) now sit under one roof. Watch this space.

How Much Does an AI Agent Cost per Month?

It depends on who's buying.

• Individuals: $20 to $250 a month for a general-purpose agent like Manus (from $20) or Genspark ($24.99 to $249.99), and $200 for Perplexity Max.
• Small and mid-sized support teams: a business handling 5,000 support conversations a month, with 65% resolved by the agent, would pay about $3,217 a month on Fin's $0.99 meter (3,250 × $0.99), plus seats for the humans handling the rest.
• Enterprises: six figures a year. Sierra's starting contracts are reportedly about $150,000. Third-party estimates of Decagon contracts run from about $105,000 to $923,000 a year (unconfirmed).

That spread is the point. The same job, answering customers, is priced from a few hundred dollars a month to close to a million a year. The meter matters less than the size of the job it's attached to.

Worked Example: Does $0.99 a Resolution Actually Make Money?

Let's run the numbers as if you were building a support agent and wanted to charge per resolution.

Three lessons jump out.

First, you pay for the failures. Every conversation costs you tokens. Only the successful ones pay you. Your resolution rate is now your gross margin.

Second, model choice moves margin more than price does. Doubling the model's token price turned a 62% margin into 30%. Routing easy questions to a small model (Claude Haiku 5.5 is $0.10 in and $0.50 out) and caching the repeated instructions can cut the cost far more than any price rise. Anthropic charges cache hits at roughly 5% of the normal input rate.

Third, competitors will undercut the headline. In April 2026 HubSpot cut its Customer Agent from $1 per conversation to $0.50 per resolved conversation. If your margin only works at $0.99, you're one competitor announcement away from trouble.

Run your own version in the AI Gross Margin Calculator. Use your real token counts from logs, not your hopes.

Is an AI Agent Cheaper Than a Human?

This is one of the most searched questions on the topic, and it's the wrong way to set your price. But it's the right way to sell it.

Fin's $0.99 works as a sales pitch because a support manager can compare it with what a human-handled ticket costs them. Harvey's legal AI reportedly sells for $1,200 to $2,000 or more per seat per month because it's measured against lawyer hours, not software budgets.

So here's my rule: set your floor from your cost, and your ceiling from the customer's alternative.

Your cost per task tells you the lowest price you can survive. The customer's cost of doing the job another way (staff time, an outsourcer, a missed sale) tells you the most they'll pay. Your price lives somewhere between the two. Where exactly depends on competition and how much proof you have.

Ask your customer for their own number. What does a ticket cost them today? What's a qualified lead worth? Don't guess it for them.

Which Agent Pricing Model Should You Choose?

Answer these four questions in order.

Question 1: Can the result be counted and checked by both sides? If yes (a resolved ticket, a booked meeting, a qualified lead), outcome pricing is on the table. If not, skip to question 2.

Question 2: Does your cost per task vary by more than about 5x? If yes, you need a meter that moves with cost: actions, credits or minutes. A flat per-conversation price will lose money on the big jobs.

Question 3: Does the buyer need a fixed number for their budget? Business buyers almost always do. Add a platform fee, a seat, or a committed volume.

Question 4: Will the agent reduce the customer's headcount? If yes, don't anchor on seats. Your success shrinks your own meter.

Here's how the answers map to models:

Still unsure? The Pricing Model Picker walks you through the same logic with your own answers.

Why Do Buyers Keep Choosing the "Boring" Meter?

Here's something the hype pieces don't tell you.

When Decagon offers customers a choice between paying per resolution and paying per conversation, most choose per conversation. When Salesforce offered five meters, the fastest-growing option was a fixed-price contract.

Why would a buyer turn down "you only pay when it works"?

Because of a problem I described in Chapter Four of the book: the prediction conundrum. Usage-based revenue is less predictable than a subscription, and that cuts both ways. Your customer's finance team has to forecast the bill. A per-resolution price that depends on a resolution rate nobody can predict yet is hard to put in a budget.

Outcome pricing is a great way to win the deal. Predictability is often how you keep it.

So offer both. Lead with the outcome story, and let the buyer convert it into a committed volume or a per-conversation rate.

Four Mistakes That Blow Up Agent Pricing

I've watched the same four mistakes repeat across the AI market over the past 18 months.

Mistake #1: Promising "unlimited"

Cursor's $20 plan once included 500 fast requests and then unlimited slower ones. Heavy users running long agent sessions cost far more than $20. When Cursor changed the plan in June 2025, the rollout went so badly its CEO apologized publicly and refunded the surprise charges. (The full story is in the Cursor case study.)

If a task costs you money, never sell it as unlimited. Say "includes X, then $Y each".

Mistake #2: Hiding the price until after the work

Replit's effort-based pricing charged on the time and compute each request used. Logical. But customers couldn't see the cost until after the agent ran, and one user reported $1,000 in a week against $180 to $200 a month before.

Show an estimate before the agent starts. Let customers approve anything above it.

Mistake #3: Charging for your own failures

Fin charges nothing when the customer asks for a human. Salesforce's Help Agent has "no charge for escalations". Clay stopped charging for lookups that find no data.

If your agent loops, fails or has to redo its own work, eat the cost. It's cheaper than the churn.

Mistake #4: Letting the meter switch a customer off

A customer who hits a limit mid-task, at a busy moment, doesn't think "fair enough". They think about switching. Use soft limits, slower fallback modes or capped automatic top-ups rather than a hard stop.

How Do You Protect Margin When Tokens Get Cheaper but Agents Use More?

Model prices keep falling for the same level of capability. Anthropic's flagship Opus went from $15/$75 per million tokens to $4/$20 for Opus 5.5.

But your bill may not fall. Agents resend their context at every step and reasoning models generate extra "thinking" tokens, so a cheaper token can still mean a more expensive task. And some prices go up. Google's Gemini 4 Argon is $2/$10 until 30 November 2026 and $4/$20 from 1 December.

Five habits protect you:

1. Cost your product at the post-increase price if your model has a scheduled rise.
2. Route by difficulty. Send easy tasks to small models and keep the expensive model for hard ones.
3. Cache what repeats. System instructions and help-center articles are the same in every call.
4. Batch what can wait. OpenAI, Anthropic and Google all discount batch processing by 50%.
5. Re-check monthly with the AI Token Cost Calculator, because model prices change more often than your pricing page should.

What Agent Pricing Does to Net Dollar Retention

In Chapter Four I wrote about the "miracle" of net dollar retention: Snowflake's 158% and JFrog's 139%, both driven by usage-based pricing. Customers landed small and spent more every year.

Agents can do the same, because every new workflow the customer hands over adds usage. Salesforce reported Agentforce ARR rising from $800 million to $1.5 billion over three quarters. Yet SaaStr estimated in September 2026 that only about 3% of Salesforce customers ran Agentforce in production.

That gap is the opportunity. The money in agents isn't the first deployment. It's the fifth.

So design for expansion from day one. Start with one narrow job that works. Price it so the second and third jobs are easy yes decisions. Then measure your net revenue retention with the NRR calculator.

And remember Factor #4 from the book: sales compensation. Pay your salespeople on committed volume and on how usage grows over the following year, not just on signing the logo. Otherwise you'll reward small deals that never expand.

What to Do Next

Here's a simple plan for this week:

1. List the jobs your agent does and pick the one your customer already counts (tickets, leads, calls, documents).
2. Pull real token counts for 50 typical tasks from your logs, and work out your cost per task on your current model and one cheaper model.
3. Calculate margin including failures using the worked example above.
4. Choose a primary meter with the four questions, and add a platform fee or commitment for business buyers.
5. Write the outcome definition (if you use one) in plain English and publish it.
6. Set default spending caps and an alert at 80% of budget.
7. Compare your price with the vendors in the table, then test it with five real customers before you put it on your website.

If you build agents for clients rather than selling a product, read my guide to how much to charge for building an AI agent. And if you're adding an agent to an existing subscription product, start with usage-based pricing for AI products.

Read the whole book free

The Subscription Playbook as a free ebook and audiobook. Australian business owners can claim a free hardcover, shipped free.

Claim the free hardcover Get the ebook

More in this section

Frequently asked questions

Get both books free

The Subscription Playbook (ebook + audiobook) and Feed a Starving Crowd, instantly. In Australia? Claim the free hardcover.

Leave this empty

First name

Email

CountryAustraliaNew ZealandUnited StatesUnited KingdomCanadaOther

What best describes you?Business owner (services)Business owner (retail / e-commerce)SaaS or AI founderHealth or allied health practiceTrades or home servicesAgency or consultantAccountant or adviserEmployee / marketerStudent / other

Yes, I'd like a free 30-minute Subscription Profit Review with a BPAG Accountants adviser (Australian businesses).

Your details go to Subscription Playbook and our sponsor BPAG Accountants Pty Ltd, who may contact you about the book and, if you ticked the box, your review. Unsubscribe any time. Privacy policy.

Send me the free books

Robert Coorey, MBA

Robert is the author of The Subscription Playbook and Feed a Starving Crowd and a Director at BPAG Accountants in Sydney. More about Robert.