How Much Does It Cost to Build an AI Agent? (2026, in $ and £)

The cost of an AI agent depends on which of three routes you take. A platform agent, such as Copilot Studio or Agentforce, has a small upfront cost and bills per action, credit or outcome. A custom agent built on a model provider’s API has a build price plus running costs. Some businesses use a mix of both. With CodeLeap, a first custom AI agent (one agent, one named job, inside your existing tools, with human sign-off where it matters) starts from £5,000 (about $6,500). An agent rollout across several workflows starts from £12,000 (about $15,600).

A price range on its own tells you very little. This article shows the arithmetic: what the platforms charge, what a model costs to run for 1,000 tasks a month, and what usually sits outside the headline price. If you are still unsure what an agent is, start with what is an AI agent. For the narrower question of a customer-service bot, our UK chatbot cost guide covers it in depth.

AI agent cost at a glance

This table sets the three routes side by side. The figures are explained in the sections below.

Route	Upfront	Monthly	Per use
Platform agent	Little or none	Subscription or credit pack	Per action, credit, conversation or outcome
Custom agent on a model API	Build price (ours from £5,000, about $6,500)	Maintenance, tooling, hosting	Model tokens, a few cents or less per task in our example
Mix of both	Build price for the custom part	Both bills	Both meters

What platform agents cost

Platform agents are priced by use, and each vendor counts use differently. At the time of writing (November 2026), these are the published prices. They change, so check the vendor’s page before you budget.

• Microsoft Copilot Studio. $200 a month for a pack of 25,000 Copilot Credits (Copilot Studio pricing).
• Salesforce Agentforce. Flex Credits cost $500 per 100,000 credits, and an action uses 20 credits. That works out at about $0.10 per action. For customer-facing agents, Salesforce also lists $2 per conversation (Agentforce pricing).
• Intercom Fin. $0.99 per outcome (Intercom pricing).

The unit matters more than the headline. A credit, an action, a conversation and an outcome are four different things, and a single agent task can use several credits or several actions. Before you commit, ask the vendor how your specific job would be counted, then multiply by your monthly volume.

When a platform agent is the right call

A platform agent is usually the better buy in two situations. You already live in that platform, so the agent works on data it can already see. And the job fits the platform’s built-in actions, such as answering from your help content or updating records in the same system.

We would rather say that plainly than sell you a build you do not need. A custom agent earns its price when the job crosses several systems, needs rules the platform cannot express, or when per-use fees have grown into a line you notice.

What drives the price of a custom agent

Six things move the price of a custom agent. Scope each one before you ask for a quote.

Systems it must connect to

Every system the agent reads from or writes to needs a connection, permissions and error handling. One well-documented API is a small job. Three systems, one of which has no API, is a very different one. Our guide to API integration covers where these connections tend to go wrong.

Whether it answers or acts

An agent that only answers questions needs content and a model. An agent that acts (raises the refund, updates the record, sends the email) needs write access, rules for when it may act, and a plan for when the action goes wrong. Acting is where most of the cost sits. Our AI agent examples show the difference in practice.

Approval steps

A human approval step costs build time, because someone has to design what the reviewer sees and what happens when they reject the work. It is also the cheapest risk control you can buy, so we would rarely drop it for a first agent.

Testing against real cases

An agent has to be tested on real examples, including the awkward ones, before it touches live work. Collecting those cases, writing down what a correct result looks like and re-running them after every change takes real time. Skipping this is how agents end up confidently wrong.

Security review

An agent with access to your systems needs the same care as a new employee with the same access. Decide which permissions it gets, what it logs and who can change its instructions.

How messy the input data is

Agents read emails, PDFs, spreadsheets and free-text fields. Clean, structured input is cheap to handle. Inconsistent, scanned or half-missing input needs extra handling and more testing.

Running costs: a worked monthly bill

Model usage is the part people worry about first, and it is usually the smallest part. The example below is illustrative. Every input is our assumption, so replace them with your own.

• 1,000 tasks a month
• About 6,000 input tokens per task (instructions, the relevant documents and context)
• About 1,000 output tokens per task

That is 6 million input tokens and 1 million output tokens a month. The table applies each provider’s published per-million-token prices to those volumes.

Model	Price per million tokens (input / output)	Arithmetic	Monthly model cost
OpenAI gpt-6-luna	$0.10 / $0.50	6 × $0.10 + 1 × $0.50	$1.10
Gemini 3.8 Flash (now)	$0.75 / $3.75	6 × $0.75 + 1 × $3.75	$8.25
Gemini 3.8 Flash (from 1 Jan 2027)	$1.50 / $7.50	6 × $1.50 + 1 × $7.50	$16.50
Claude Sonnet 5.5	$2 / $10	6 × $2 + 1 × $10	$22
Claude Opus 5.5	$4 / $20	6 × $4 + 1 × $20	$44

Prices come from Anthropic, OpenAI and Google, as published at the time of writing, in US dollars. Google lists Gemini 3.8 Flash at the higher rate from 1 January 2027, so a budget that runs past that date should use the second Gemini row. The example uses the same token counts for every model, although different models can count the same text differently.

Two published discounts can lower the figures. Anthropic lists a cache-hit price of $0.10 per million tokens on Sonnet 5.5, so a long instruction block that repeats on every task can cost much less than the table assumes. Anthropic also lists a 50% discount on input and output through its Batch API, which suits work that does not need an immediate answer, such as overnight document processing.

The other line items

The model is one line on the bill. These are the others, with the prices we could verify.

• Vector database. If the agent searches your documents, it may use one. Pinecone’s Starter plan is free, and its Standard plan has a $50 a month minimum (Pinecone pricing).
• Tracing and monitoring. You need to see what the agent did and why. LangSmith’s Plus plan is $39 per seat per month (LangChain pricing).
• Hosting. The agent has to run somewhere. We have not priced this here, because it depends on where and how it runs.
• Maintenance and support. Ours starts from £400 (about $520) a month.

Add the pieces we could price for one workflow, using Sonnet 5.5 at $22, one LangSmith seat at $39, and a vector database at $0 on Pinecone Starter or the $50 Standard minimum. That comes to roughly $61 to $111 a month, before hosting. This is why we say running costs for one workflow are often tens to low hundreds of dollars a month.

Maintenance is the line that changes the picture. At about $520 a month it is several times the rest combined. Models change, your systems change, your documents change, and someone has to review a sample of real work and fix what drifts. Budget for changes, not just for tokens.

Building it in-house: the salary maths

An in-house build has no agency fee, but it is not free. The US Bureau of Labor Statistics reports a median annual wage of $135,980 for software developers (May 2025) (BLS Occupational Outlook Handbook).

That is wage only. It leaves out benefits, management time and the time your engineers spend learning agent tooling. As our own arithmetic, one engineer for a quarter of a year is roughly a quarter of that wage, about $34,000, before those extras.

In-house is the right choice when you have engineers with spare capacity, a reason to keep the system entirely inside your company, and a plan to maintain it. If one of those is missing, the quarter you spend building is a quarter you do not spend on your product. We are not comparing this with agency rates here, because a day rate says little about what the finished agent costs you.

Why agent projects get cancelled, and how to avoid paying twice

Gartner predicted on 25 June 2025 that over 40% of agentic AI projects will be canceled by the end of 2027. It gave three reasons: escalating costs, unclear business value and inadequate risk controls (Gartner press release).

The same release warns of “agent washing”. Gartner uses the term for vendors that rebrand AI assistants, robotic process automation (RPA) and chatbots without substantial agentic capabilities. It estimates that only about 130 of the thousands of agentic AI vendors are real. When a vendor says “agent”, ask what the product does on its own, and watch it do it on your data.

The three reasons for cancellation each have a practical answer.

• Costs escalate. Name one job and price it. A project that starts as “automate operations” has no edge to its budget.
• The value is unclear. Measure the job before the agent and after it. If you cannot say how many minutes or dollars the job costs today, you cannot say whether the agent helped.
• Risk controls are missing. Keep a human approval step wherever a mistake would be costly, and log what the agent does.

If you are not sure which job to pick, start with a readiness check. Our AI readiness assessment is a fixed £3,000 (about $3,900) and takes one week. It exists so you do not pay for a build before you know what to build.

Pounds or dollars?

We price in pounds. The dollar figures here use $1.30 to £1 and are rounded. The Bank of England spot rate was $1.3476 to £1 on 5 November 2026 (series XUDLUSS in the Bank of England database), so an invoice in pounds may convert to a slightly different dollar amount on the day you pay. At that rate, £5,000 was about $6,740. Vendor prices are shown in US dollars exactly as the vendors publish them, with no conversion and no tax added.

UK readers who want the SaaS chatbot picture in pounds can read our chatbot cost guide. If your agent will talk to other tools through a standard connection, our guide on how to build an MCP server explains one way to do it.

A sensible next step

Write down the one job you want an agent to do, how often it happens each month, and which systems it would need to touch. With those three answers, the choice between a platform agent, a custom build or a mix gets much easier, and the price stops being a range and becomes a number.

If you want to talk it through, book a call about AI agents and automation. A first agent is scoped and priced before work starts. If you would rather settle the question of what to automate first, the AI readiness assessment is one week at a fixed £3,000 (about $3,900).