Having an AI agent built: approach, cost and timeline

When should you have an AI agent built?

Having an AI agent built makes sense when a task has a lot of repetition, needs information from several sources, and produces an outcome a human can assess quickly. If the task is entirely predictable, an ordinary workflow or classic automation is often better. If you’re not yet sure what sets an agent apart from other AI applications, read what an AI agent is first. For inspiration by department, see the AI agent examples.

Once it’s clear that an agent is the right route, the first choice isn’t which vendor, but which form.

I set out that trade-off between routes, including the multi-year costs and how you get off a platform again, in build, buy or partner. For bespoke work, the page on AI development is also relevant.

The development process

A good project starts with a specification you draw up yourself, not the delivery partner. Leave that to the party that’s also building, and you get a specification that suits what that party would rather build.

Specify the agent. At minimum, set out:

• Tasks: which steps the agent carries out and what result it delivers.
• Boundaries: which systems and data it may read, what it may change, and what it never may.
• Human checks: at which points a human approves, who that is, and what happens if there’s doubt.
• A quality measure: a test set of real cases with the known correct outcome, so you can judge objectively.
• Logging and ownership: what’s recorded, who manages the agent, and who can shut it down.

Select a delivery partner. Send every candidate the same scenario, built on your own data, and judge the outcomes separately from the pitch. Ask about experience with similar integrations, their approach to testing and logging, ownership of code and instructions, and the terms on ending the contract. The full approach is covered in AI vendor selection in eight steps.

Build and test. Have the party build a limited version first that only prepares or proposes. During the test period, a human reviews every outcome. Only once it’s clear which mistakes occur do you increase its independence.

What does an AI agent cost?

What an AI agent costs varies sharply by route. Naming a fixed figure here would be misleading. The factors that determine the price are easy enough to name:

• Route: an agent inside an existing platform costs mainly licences and configuration; bespoke work costs build hours and then management.
• Integrations: every system the agent talks to takes build work, testing and upkeep.
• Write access and control: the more the agent may change, the more testing, logging and checkpoints are needed.
• Usage: language models are often billed per volume of text processed, so volume and document length drive the running cost.
• Management: monitoring, periodic tests, changes as processes shift, and the time of an internal owner.

The biggest risk factors are too broad a first scope, sources that aren’t in order, and dependence on a single vendor with no exit terms agreed. Ask every party for a multi-year cost estimate, including management and switching.

My own rates are on the pricing page. If you want quotes from several delivery partners independently assessed, a Vendor Review of 3-6 weeks fits, costing €9,800 - €15,400. If you need someone to oversee the project on an ongoing basis, that’s available as Fractional AI Lead.

Security, privacy and management

An agent that calls systems is a new user in your IT landscape, with the risks that come with it. Don’t give it more access than it needs, have it log in with its own account so its actions can be traced, and keep a confirmation step at every irreversible moment.

If the agent processes personal data, the ordinary UK GDPR rules apply: a purpose and a lawful basis, no more data than needed, a data processing agreement with the vendor, and in many cases a data protection impact assessment. If the agent makes decisions about people, stricter requirements come into play. That’s covered separately on this site.

Management isn’t an afterthought. Models and platforms change, and an agent that works well today can respond differently after an update. Agree with the delivery partner who monitors, how often the test set gets rerun, and who picks up an incident.

My role. I don’t build agents myself and I don’t sell implementations. I help you scope the task, draw up the specification, compare delivery partners, and oversee delivery, from your side of the table. If you don’t yet have a delivery partner, I can suggest parties from my partner network; you contract them directly, with no margin for me. More on agents in general is on the page about AI agents for business. Have a concrete idea? Book a consultation and you’ll hear what I’d do by the end of the call.