What a custom AI agent costs a business

A custom AI agent costs a business mostly in integration, data preparation and guardrails, not in model fees. Inference, the per-use charge paid to the AI provider, is usually the smallest line in the budget, but it is the only one that grows with usage, so it has to be measured before launch rather than discovered on the first invoice. This guide is for the person who has to approve the budget, not the person writing the code.

There is a lot of noise around agents right now. The aim here is to separate the parts of an agent project that are expensive from the parts that only sound expensive, and to show how to keep the risk proportionate.

Assistant, agent, automation: what are you actually buying?

"AI agent" covers very different things. Three levels make the budget conversation clearer.

Each step down the table raises the potential value and the cost of keeping it safe. A wrong answer from an assistant is an annoyance. A wrong write from an acting agent is a bad record in your system of record.

Use cases that tend to pay off

Agents that earn their keep in small and mid-sized businesses share three traits: a repetitive task, information scattered across several tools, and an output a person can check quickly.

• Account briefings: pull emails, contracts and open tickets into one summary before a call, with links to the sources.
• Inbound triage: read a request, classify it, extract the key facts and draft a reply for a person to approve.
• Answers from internal documentation: policies, product sheets, terms of service, with the exact passage cited.
• Pre-filled business documents: a quote, a visit report or a project update that someone reviews before it goes out.

The disappointing projects usually expect the agent to make a judgement no one can verify: a commercial commitment, a compliance sign-off, a diagnosis. Our guide to adding AI features to a business application covers what should stay under human control.

Analysts share the caution. In June 2025, Gartner predicted that over 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. It is a forecast, not a measurement, but it names the three things to settle before you start.

Where the money goes

1. Scoping the use case

Pick one task, define what a good result looks like and collect real examples to use as a test set. This is short work, and it is what stops you from building a general-purpose agent nobody uses. A prototype run on your own examples lets you judge before committing to a build; see prototype vs MVP vs proof of concept.

2. Data and retrieval

An agent is only as good as what it reads. Sources need sorting (outdated versions, contradictory duplicates), splitting, indexing and keeping in sync automatically. Nobody trains a model for this: the agent retrieves the relevant passages for each request and hands them to the model. The work is mostly documentation hygiene and plumbing, and it is routinely underestimated.

3. Integration with your systems

Usually the heaviest line. The agent must connect to your CRM, ERP, help desk or line-of-business software, with permissions that mirror the user's: a sales rep must not be able to read, through the agent, a file they could not open themselves. If a system has no API, one sometimes has to be built.

4. Guardrails

• Least privilege: the agent only gets the actions it needs.
• Human approval for anything irreversible or binding.
• Logging: who asked what, which sources were read, which action ran.
• Ongoing evaluation: the test set is replayed whenever a prompt or a model changes.
• Spending caps per user and per day.

These answer documented risks. The OWASP Top 10 for LLM Applications (2025) ranks prompt injection first, where text the agent reads (an email, a web page) carries instructions that hijack it, and also lists excessive agency, where an agent holds more permissions or actions than it needs.

5. Inference

Providers charge per token processed, with one price for input and a higher one for output. The spread between models is wide. Published prices checked on 3 November 2026, in US dollars per million tokens (input / output):

Sources: Anthropic pricing, OpenAI pricing, Gemini API pricing. The table shows the range, not a ranking: quality depends on the task, and only a test on your own examples settles it.

What is specific to agents is that one task means several model calls. The agent plans, calls a tool, reads the result, calls another tool and then writes, and each step resends part of the growing context. A back-of-the-envelope estimate based on a single question will understate real consumption.

A worked example, with assumptions you should replace with your own measurements: a task that uses 30,000 input tokens and 3,000 output tokens in total costs $0.09 on a model priced at $2 / $10 per million, and $0.45 on one priced at $10 / $50. At 2,000 tasks a month, that is $180 against $900. Two levers lower the bill, both documented on the pricing pages: batch processing, billed at half price by all three providers above for work that can wait, and caching of repeated context.

6. Running it

An agent in production is monitored like any application: correction rate, response time, spend, tool errors. Models also move. Providers retire old versions and change prices: Google's pricing page announces that from 1 December 2026, requests sent to Gemini 2.5 Pro will be routed to another model and billed at that model's rate, and that Gemini 4 Argon's prices double on the same date. Budget this follow-up as part of normal maintenance, alongside the build itself; our breakdown of custom business application cost covers the running costs that come after launch.

Limits to accept before you sign

• Agents are confidently wrong at times. Cited sources and human approval remain essential for anything that commits the business.
• Quality follows the documents. A stronger model will not fix an outdated procedure.
• Personal data stays regulated. Where is it processed, how long is it kept, does the provider reuse it? Our GDPR checklist for business applications lists the questions to ask.
• The EU AI Act may apply, depending on use and on where your users are: an agent that talks to customers must at least disclose that it is an AI. See the EU AI Act for companies building software.
• Provider lock-in is reduced by design: a common model access layer, prompts and test sets you own, and the ability to swap models without rebuilding the application.

How to scope the project

1. One task, one user group, one metric, such as the time it takes to prepare an account briefing, before and after.
2. A test set of twenty to fifty real cases, with the expected answer.
3. A read-only prototype, with no actions in your systems, run on those cases.
4. A consumption measurement on the test set, multiplied by expected volume, for each candidate model.
5. Actions switched on one at a time, each with human approval and a log.
6. A decision point at each stage: continue, adjust or stop.

This staged path costs less than an ambitious agent delivered in one go, because it lets you stop early on a use case that does not hold up. The contract should mirror it: a defined scope per stage, and rights over the code and prompts set out case by case in the contract.