How Enterprise AI Agents Are Priced: Models, Real Prices, and TCO

Microsoft's billing documentation for Copilot Studio contains a rule most pricing pages leave out. When a tenant on prepaid capacity reaches 125% of that capacity, its custom agents are disabled, and people who try to use one are told the agent has reached its usage limit. The conversation already in progress finishes. The next one does not start.

That is not a criticism of Microsoft. It publishes the rule, the per-feature rates, and a usage estimator, which is more than many vendors do. It is a reminder that AI agent pricing is two questions, not one: what the meter charges, and what happens to the work when the meter runs out.

This guide is written for the people who have to put a number in a budget and defend it later: the line-of-business owner sponsoring the project, and the finance, procurement, and platform teams reviewing it. It covers the six pricing models in use today, real list prices taken from each vendor's own pricing page in September 2026, a worked example of how design choices move a bill, the total cost of ownership beyond the license, and the questions that expose the difference between a quote and a cost.

What is AI agent pricing?

AI agent pricing is the way a vendor charges for software that plans and takes actions with delegated access, rather than only answering questions. Vendors meter one of six things: the people who have access (seats), the interactions the agent handles (conversations, actions, executions, or credits), the compute and tokens it consumes, the agents or bots you deploy, the outcomes it delivers, or a scoped engagement that bundles platform access with implementation. Most combine two or more.

Pricing is unsettled because the old unit does not fit. Software seats assume a person logs in. An agent triggered by a new order, a ticket, or a schedule has no login to attach a seat to. You can see the market working this out on a single page: Salesforce sells Agentforce per user, per conversation, and through prepaid Flex Credits side by side, with pre-purchase, pre-commit, and pay-as-you-go buying options.

If you are still settling what counts as an agent, start with what an AI agent is. If you are here because a quote landed on your desk, read on.

The six AI agent pricing models at a glance

Every price below is a list price from the vendor's own pricing page, as of September 2026. Negotiated enterprise prices differ, and vendors change these pages often, so check the source before you rely on a figure.

The rest of this section takes each model in turn: how it works, where it fits, and what to watch.

The six pricing models, explained

1. Per seat: you pay for each person with access

Seat pricing charges a fixed monthly fee for each named user, usually on an annual term. As of September 2026, Microsoft lists Microsoft 365 Copilot at $30.00 per user per month, paid yearly, including Copilot Chat and access to build in Copilot Studio. Salesforce lists Agentforce add-ons at $125 per user per month for unmetered Agentforce usage by employees, Agentforce Industries add-ons at $150 per user per month, and Agentforce 1 Editions from $550 per user per month, which include 2.5 million Flex Credits per org per year.

Where it fits: employee-facing use, where usage roughly tracks the number of people. What to watch: a seat pays for access, not for work done, and the inclusion has edges. Microsoft's billing rates page says employee-facing agent usage by Microsoft 365 Copilot licensed users carries no charge within fair usage limits, but agent flows started by other triggers consume Copilot Credits at the standard rate, and Computer-Using Agents are not included. The autonomous work is where the second meter starts.

2. Per interaction: conversations, actions, executions, or credits

Interaction pricing charges per unit of activity. The unit is what varies. Salesforce offers Agentforce at $2 per conversation, or through Flex Credits at $500 per 100,000 credits, where an Agentforce action costs 20 Flex Credits and an Agentforce Voice action costs 30. At list price that is $0.10 per standard action. Salesforce also sells an Agentforce User License at $5 per user per month that requires Flex Credits, so seat and credit meters run together.

Microsoft Copilot Studio sells prepaid packs at $200.00 per pack per month for 25,000 Copilot Credits, with savings of up to 20% for up-front purchase, and a pay-as-you-go option billed monthly through an Azure subscription. Credits are consumed at different rates per feature, according to Microsoft's billing rates: 1 credit for a classic (authored) answer, 2 for a generative answer, 5 for an agent action, and 10 for tenant graph grounding.

Workflow tools use the same logic with a different unit. n8n prices on monthly workflow executions, regardless of complexity: as shown on its pricing page when we checked in September 2026, Starter is 20€ per month billed annually for 2.5K executions, Pro is 50€ per month for 10K, Business is 667€ per month for 40K executions and is self-hosted, and Enterprise is contact sales.

Where it fits: variable, measurable work, such as customer conversations or event-triggered tasks. What to watch: two vendors' units are rarely comparable, and the design of the agent, not only the traffic, decides how many units a task burns. The worked example below shows how far that can move a bill. Ask what happens when prepaid capacity is exhausted: in Copilot Studio, custom agents are disabled at 125% of prepaid capacity unless pay-as-you-go is set up, and new agent flow runs are blocked once prepaid capacity is fully consumed.

3. Consumption: compute, memory, calls, and tokens

Consumption pricing is what you meet when you build agents on cloud infrastructure. Amazon Bedrock AgentCore lists Runtime at $0.0895 per vCPU-hour and $0.00945 per GB-hour, charged only for active consumption, so I/O wait with no background process incurs no CPU charge. Gateway API invocations cost $0.005 per 1,000, short-term memory $0.25 per 1,000 new events, and identity $0.010 per 1,000 token or API key requests when used outside Runtime or Gateway. Underlying model inference through Bedrock is billed separately.

Tokens sit underneath most meters even when they are not the headline unit. In Copilot Studio, an agent that uses a reasoning model pays the feature rate plus a premium rate of 10 Copilot Credits per 1,000 tokens, per Microsoft's billing documentation.

Where it fits: engineering teams building and running their own agents, who want to pay for exactly what runs. What to watch: the price sheet is a set of small meters, and your total is the product of all of them. The infrastructure is metered; the engineers who design, secure, and maintain what runs on it are not on the page at all.

4. Per agent or bot: you pay for each deployed worker

Per-agent pricing charges a flat fee for each deployed bot or digital worker, the unit robotic process automation has used for years. As of September 2026, Microsoft Power Automate lists Power Automate Process at $150.00 per bot per month and Power Automate Hosted Process at $215.00 per bot per month, both paid yearly, alongside Power Automate Premium at $15.00 per user per month. UiPath lists a Basic Automation Cloud plan starting at $25 per month, lists its Standard and Enterprise plans as contact sales, and says users and robots are available for purchase across all three tiers.

Where it fits: steady, always-on workloads, and finance teams that want a number that does not move with traffic. What to watch: a per-bot price tells you nothing about how much one bot can do, and the process design, build, and upkeep are separate work. For how that upkeep differs between scripted bots and reasoning systems, see RPA vs AI agents.

5. Outcome-based: you pay for a defined result

Outcome pricing charges only when the agent delivers a defined result. Intercom prices its Fin AI Agent at $0.99 per outcome as of September 2026. It counts an outcome when a customer confirms the issue is resolved, when the customer does not ask for more help after Fin responds, or when Fin completes a workflow it calls a Procedure, including handoffs, and it charges once per conversation even if several questions are answered. Fin needs no seats when used with an existing helpdesk; on Intercom's own plans, seats run $29, $85, or $132 per seat per month.

Where it fits: work with a countable, attributable result, which is why customer support leads here. What to watch: the definition is the contract. Ask how the vendor distinguishes a customer who was helped from one who stopped replying, and whether you can audit the count. Outcomes are harder to define for cross-system back-office work, where several systems and people touch one result.

6. Platform plus implementation: a scoped engagement

The last model is the one without a number on the website. The vendor scopes an engagement that bundles platform access with configuration, engineering, and support, and quotes it after discovery. UiPath Standard and Enterprise and n8n Enterprise are listed as contact sales, and Salesforce notes on its pricing page to contact sales for details.

Where it fits: regulated, cross-system, or on-premises work, where scope drives cost more than traffic does. What to watch: you cannot benchmark the price from the outside, so the burden moves to the scope document. Insist that it names what is included, what success looks like, and what the price becomes after the pilot. Mission Control prices this way; the structure is set out below.

Hybrid pricing is the norm

Few vendors run one meter. Salesforce's $5 user license requires Flex Credits. Intercom combines seats with per-outcome charges. Copilot Studio includes some employee usage with a Microsoft 365 Copilot license and meters the rest in credits. When you compare quotes, write down every meter a vendor uses, not only the headline one.

A worked example: how design moves the bill

Unit prices are easy to compare and nearly useless on their own. What decides the bill is how many units a task consumes, and that depends on how the agent is built. Two examples, worked from published rates. The arithmetic is ours, at list price, before discounts.

Same traffic, different design: Copilot Studio

Microsoft's billing documentation includes a customer support example: an agent serving 900 customers a day, each run using four classic answers and two generative answers, consumes 7,200 Copilot Credits a day. Over 30 days that is 216,000 credits, or nine prepaid packs at $200 each: $1,800 a month.

Now change the design, not the traffic. If the same agent answered all six questions generatively and grounded each one in the tenant graph, each run would cost 72 credits (six answers at 2 credits plus six groundings at 10). That is 64,800 credits a day, 1,944,000 over 30 days, or 78 packs: $15,600 a month. Same customers, same questions, nine times the credits.

Same vendor, different meter: Agentforce

At Salesforce list prices, a conversation costs $2 on the per-conversation meter, and a standard action costs $0.10 on Flex Credits (20 credits at $500 per 100,000). The two meters cost the same when a conversation averages 20 actions. Below that, Flex Credits cost less per conversation; above it, the per-conversation meter does. You cannot pick the cheaper meter until you know how many actions your real conversations take, which is a question for a pilot, not a pricing page.

The lesson carries to every vendor: ask for a consumption estimate on your own workload, designed the way you will actually run it, and repeat it at three and ten times today's volume.

AI agent pricing examples, as of September 2026

The same figures in one place. All are list prices from each vendor's own pricing page, accessed September 2026, in the currency and terms shown there.

Total cost of ownership: the lines beyond the license

A price sheet shows the meter. It does not show what it costs to get a system into production and keep it there. In a survey of 372 organizations by Benchmarkit and Mavvrik, reported by CFO Dive in September 2025, 85% of companies missed their AI cost forecasts by 10% or more, and 24% missed by more than 50%. The top drivers of unexpected cost were data (acquisition, storage, cleaning, and security) and hidden costs from historical business practices and legacy systems. Neither appears on a pricing page.

Build the total cost of ownership from these lines:

Costs that sit outside the meter

Two published examples show why the license and the bill are different numbers. Amazon states on its AgentCore pricing page that model inference through Bedrock is billed separately from AgentCore. Microsoft's Copilot Studio billing rates exclude bring-your-own-model configurations, including Azure Foundry models, which are billed separately. Neither is hidden; both are easy to miss when a spreadsheet only has a column for the platform.

Implementation, integration, and maintenance

The build is the line that most often surprises. A consumption price assumes someone has already designed the agent, connected it to your systems, and secured it. A per-bot price assumes someone has already mapped and scripted the process. Price that work honestly, including the hours your own experts spend teaching and reviewing. If you are weighing a build on open-source agent frameworks, count the governance you will have to build and maintain yourself; the AI agent security checklist lists the controls to cost out.

Maintenance is the line that grows. Tools change, models are upgraded, interfaces move, and every change needs retesting. If the work you are pricing is currently done by an outside provider, compare against that full run-rate too, as set out in synthetic workers vs managed service providers.

Deployment model changes who pays

Where the system runs moves costs between you and the vendor. A vendor-hosted service folds infrastructure into the meter. A self-hosted deployment, such as n8n Business, moves hosting, patching, and monitoring to your team. On AgentCore, the instance option bills the EC2 instance cost plus a management fee of 12% of the On-Demand rate. If your data falls under a regime that keeps it inside your perimeter, settle that before you compare prices, because it rules some pricing models out entirely.

Questions to ask before you sign

Take these into every vendor call. A vendor that answers them plainly, in writing, is easier to budget for than a cheaper one that does not. For the capability side of the evaluation, pair this list with how to evaluate AI agents for the enterprise.

About the meter

• What exactly is the billing unit, and where is its definition written down?
• How many units does one of our typical tasks consume, designed the way we will actually run it?
• Which features consume at higher rates (grounding, reasoning models, voice, computer use)?
• What does the estimate look like at three and ten times our current volume?

About limits and overage

• What happens when prepaid capacity runs out: does work stop, queue, or bill at an overage rate?
• Can we cap spend per agent or per environment, and who is notified before a cap is hit?
• For outcome pricing: how is an outcome counted, and can we audit the count?

About everything outside the meter

• Is model inference included, or billed separately? Can we bring our own model, and how is that priced?
• What implementation work is included, by whom, for how long, and what do our people have to do?
• Which of our systems are covered by included integrations, and which are custom work?
• Which security controls ship in the product, and which would we have to build?
• Who pays for fixes when a connected system or a model changes?

About the contract

• What is the price after the pilot or first term, and on what unit?
• Which commitments are annual, and what happens to unused capacity?
• If we leave, what do we keep: configurations, logs, and the procedures the system learned?

How Mission Control prices synthetic workers

Here is how the vendor writing this guide prices, so you can put the same questions to us. Mission Control builds synthetic workers, which run on the Swarm platform. Mission Control does not publish list prices. Engagements are scoped per pilot, so pricing is a conversation with sales, and in the six-model framework above it sits in model 6, platform plus implementation.

The structure is fixed even though the price is scoped. As published on the synthetic workers page, every engagement follows the same structure: 12 weeks, three phases.

• Weeks 1-3: Scoping and configuration. Select the use case, ingest your documents and standards, configure the worker.
• Weeks 4-8: Single-user learning loop. Your designated expert works with the synthetic worker. It learns, converges, improves.
• Weeks 9-11: Cross-user validation. A second user starts from a higher baseline. Knowledge transfers without degradation.
• Week 12: Analysis, documentation, and go-forward recommendations.

The pilot includes platform access, worker configuration, all engineering support, and a forward-deployed team embedded with your organization for the duration. That answers several of the TCO questions above by design: implementation and engineering are inside the engagement rather than a separate services line, and the expert's time is planned into weeks 4 to 8 rather than discovered later.

Now the limits, stated the way you should expect any vendor to state them. Because there is no public list price, you cannot benchmark Mission Control from its website; you get a number after scoping. A 12-week engagement with an embedded team is a heavier commitment than buying a credit pack and starting tomorrow. And the questions about post-pilot pricing, units, and inference belong in your scoping conversation with us, in writing, exactly as they would with anyone else.

Before you sign: a one-page pricing sheet

AI agent pricing looks like a choice between meters. In practice it is a forecast of how your work will run, and the forecast is only as good as the workload you test it on. Before you sign anything, fill in one page for each vendor:

• Unit and definition: the billing unit, quoted from the vendor's own documentation.
• Your consumption: units per task for your real workload, designed the way you will run it.
• Stress test: the monthly bill at today's volume, three times, and ten times, including overage rates.
• When capacity runs out: whether work stops, queues, or bills on.
• Outside the meter: inference, infrastructure, integration, security, and maintenance, each with an owner and an estimate.
• After the first term: the price and unit once the pilot or discount ends.
• Exit: what you keep if you leave.

The vendor whose page you can fill in completely is the one you can budget for. If you want to fill it in for Mission Control against your own work, talk to our team.

References

• Salesforce, "Agentforce Pricing" (accessed September 2026)
• Microsoft, "Microsoft Copilot Studio" pricing and pricing FAQ (accessed September 2026)
• Microsoft Learn, "Billing rates and management - Microsoft Copilot Studio" (accessed September 2026)
• Microsoft, "Power Automate pricing" (accessed September 2026)
• UiPath, "Plans and pricing" (accessed September 2026)
• n8n, "Plans and Pricing" (accessed September 2026)
• Amazon Web Services, "Amazon Bedrock AgentCore pricing" (accessed September 2026)
• Intercom, "Pricing" (accessed September 2026)
• CFO Dive, "Most firms miss AI cost forecasts, survey finds" (September 2025), reporting a Benchmarkit and Mavvrik survey
• Mission Control pilot structure: /synthetic-workers/

Related reading

• How to Evaluate AI Agents for the Enterprise
• AI Agent Security Checklist: 12 Controls for Agentic AI
• What Is Agentic Process Automation?
• UiPath Alternative for Enterprise Work
• Power Automate Alternative
• Synthetic Workers vs Managed Service Providers
• All posts on the blog