What an AI agent costs in 2026
There is no single price for an AI agent because vendors do not agree on the unit. Salesforce sells conversations and credits, Microsoft sells credits, Intercom sells outcomes and Zendesk sells automated resolutions. On published list prices a customer-facing agent costs between about one and two dollars for each conversation it handles or resolves, and an internal agent that runs a handful of actions costs cents per task.
The unit matters more than the number. A price per outcome looks cheap until you read what counts as an outcome, and a price per action looks tiny until one task needs ten of them. So this page gives each vendor's published price, the definition behind it, and the arithmetic to turn it into a cost per completed task.
I build and run production AI agents for operational teams, so my bias is towards owning the system when the work is core to the business and renting it when it is not. No vendor has paid to be here. Prices are the vendors' own published list prices, checked on 12 October 2026.
1. Salesforce Agentforce
Salesforce publishes three ways to pay. Conversations cost $2 each. Flex Credits cost $500 for 100,000, and one Agentforce action uses 20 credits, which makes an action 10 cents. Per user add-ons for Sales, Service, Field Service and Operations cost $125 a user a month, and industry add-ons $150.
In sterling Salesforce lists £1.60 a conversation and £400 for 100,000 Flex Credits. Voice actions use 30 credits. Its top editions with Agentforce included start from $550 a user a month and come with 2.75 million Flex Credits a year for the organisation.
It fits a company already running its service or sales team on Salesforce, because the agent works on the data and permissions that are already there. It is the wrong buy if your customer records live elsewhere, since you would pay to move the data before the agent could use it. Count the actions in a typical task before you choose credits over conversations.
2. Microsoft Copilot Studio
Microsoft sells Copilot Credits. A pack of 25,000 credits costs $200 a month, which works out at 0.8 cents a credit, and unused credits do not carry over to the next month. There is also a pay as you go meter through an Azure subscription, and a one year prepaid plan that Microsoft says saves up to 20 percent.
What a credit buys is published. A scripted answer uses 1 credit, a generative answer 2, an agent action 5 and an answer grounded in your Microsoft 365 data 10. Microsoft's own example of an agent that processes an incoming order with four actions comes to 20 credits, about 16 cents at the pack rate. Reasoning models add a premium charge by the token on top.
It fits a company on Microsoft 365 and Dynamics. Staff who hold a Microsoft 365 Copilot licence, $30 a user a month paid yearly, use employee-facing agents without drawing on the credits. Watch the limit. When usage reaches 125 percent of the prepaid capacity, Microsoft disables the agents until you add more.
3. Intercom Fin
Fin costs 99 cents per outcome, on top of Intercom seats at $29, $85 or $132 a seat a month depending on the plan. It can also run in front of another helpdesk, where Intercom says a minimum monthly commitment applies.
Read the definition. Intercom counts an outcome when the customer confirms the issue is resolved, when the customer stops asking for help after Fin replies, or when Fin completes a workflow, including one that ends in a handoff. Each conversation is charged once at most.
It fits support teams with many repeat questions and good help content. For work that ends in a change to a system of record, such as an order or a refund, budget for building and testing those workflows as well as the outcome fee.
4. Zendesk AI agents
Zendesk charges per automated resolution. Its UK price list shows £1.30 a resolution when you commit to a volume in advance and $2.00 a resolution on pay as you go. The Suite Team and Support Team plans include 5 automated resolutions per agent a month and Suite Professional includes 10.
Seats are separate. Suite Team is £45 an agent a month and Suite Professional £89, both billed annually, and the Copilot add-on for human agents is £40 an agent a month. So a small team pays mostly for seats, and the resolution price only starts to matter at volume.
It fits a team already on Zendesk. The gap between the committed and the pay as you go price is wide, so forecast volume from a real month of tickets before you sign.
5. Sierra and Decagon
The two best-known specialist vendors of customer-facing agents publish how they charge but not what they charge. Sierra describes outcome-based pricing, where it is paid when the agent completes a task such as a resolved conversation, and says that in most cases an escalated conversation is not charged.
Decagon offers a fixed price per conversation or a higher price per fully resolved conversation with no charge for escalations. It says most of its customers choose the price per conversation because it is predictable and avoids disputes over what counts as resolved. Both sell by custom quote, so expect an enterprise sales process.
6. Building your own agent
When you build the agent yourself you pay for three things, the model, the infrastructure it runs on and the people who keep it right. Model prices are per token and change often, so I track them daily for every major model on the LLM API pricing page of this site.
Infrastructure is cheap by comparison. Amazon Bedrock AgentCore, for example, lists its runtime from $0.0895 per virtual CPU hour and $0.00945 per gigabyte hour, and its gateway at half a cent per thousand tool invocations. OpenAI lists web search at $10 per thousand calls and file search at $2.50 per thousand, with tokens billed at the model's own rates.
The real cost is the build and the upkeep. Someone has to connect it to your systems, write the checks, test every change against past cases and review what it did last week. That is the right spend when the work is specific to your business and the wrong one when a vendor above already does the job.
Why the unit matters more than the price
A price per conversation charges you whether or not the customer was helped. A price per outcome charges only for success, but the vendor writes the definition of success, and a customer who gives up and leaves can look the same as one who was helped. A price per action or credit follows the work done, which is fair, and also means an agent that takes ten steps costs ten times one that takes one.
Per user pricing is the easiest to budget and the least connected to results. It suits agents that help staff with their own work. It does not suit an agent that runs on its own in the background, because nobody is sitting in the seat.
None of these units includes the work of making the agent correct. Connecting systems, writing the rules it must follow, building a set of test cases and reviewing its output are yours to pay for on every platform.
How to compare on your own numbers
Pick one real task and count it for a month. How many arrive, how many steps each takes, and how many end the way you want. Then price it under each model. A task that needs four actions costs 40 cents on Agentforce credits and about 16 cents on a Copilot Studio credit pack, before any premium for a reasoning model.
For a customer-facing agent, work out the cost per conversation that was truly resolved. If an agent charged at $2 a conversation fully resolves half of them, each resolved conversation costs $4. If one charged at 99 cents an outcome counts a customer who walked away, the real figure is higher than 99 cents too. Ask each vendor for the definition in writing and test it on your own transcripts.
Then add what the price leaves out, the seats you still need, the integration work and the hours your team will spend checking results. The free ROI calculator on this site does that sum.
Comparison at a glance
| Vendor | Unit you pay for | Published price | Watch for |
|---|---|---|---|
| Salesforce Agentforce | Conversation, or credits per action | $2 a conversation, or 10 cents an action | Actions per task add up |
| Microsoft Copilot Studio | Copilot Credits | $200 for 25,000 credits a month | Agents disabled at 125 percent of capacity |
| Intercom Fin | Outcome | 99 cents an outcome, plus seats | A customer who stops replying counts |
| Zendesk AI agents | Automated resolution | £1.30 committed, $2.00 pay as you go | The gap between the two rates |
| Sierra | Outcome | Custom quote | Agree the outcome definition first |
| Decagon | Conversation or resolution | Custom quote | Which of the two models you sign |
| Build your own | Model tokens and infrastructure | Billed by usage | Build, testing and upkeep are the real cost |
How to choose
If the job is answering customers and you already run Intercom, Zendesk or Salesforce, start with the agent your helpdesk sells. The data is already there and you can measure it against real tickets within a month.
If the job is internal and you are a Microsoft company, Copilot Studio is the cheapest credible start, and staff with a Copilot licence can use the agents without drawing on credits.
If the job is the core of how you operate, such as pricing a quote, entering an order or running payroll, a platform agent will get you part of the way and then stop at your rules. That is where a system built around your own process earns its cost, and where I spend my time.
At every size, measure cost per completed task, not the list price, and decide who owns the checking before you sign anything.