Rolling out AI agents across your organisation: four stages, board decisions and the roles you need

An AI chat tool such as Microsoft Copilot or ChatGPT answers questions and writes text. An AI agent goes further: it carries out steps in the company’s systems itself. It checks an order’s status in the ERP, opens a service ticket or prepares an invoice for approval.

Because an agent acts inside your systems, someone has to decide what it is allowed to do, which data it uses and who is responsible for its mistakes. These are decisions about processes, risk and people, not about software. That is why the business leads an agent programme and the IT department supports it.

This article covers a company-wide rollout: from the first agent to several running side by side. We describe building a single agent step by step in a separate guide.

An AI agent programme is best run in stages, with a board decision after each one. If something goes wrong, you lose the cost of one stage, not the whole budget.

The timings in the table below are indicative for a company with 50–500 employees. They get longer if data is scattered, if many systems need to be connected, or if the company’s own people have little time available.

The first stage is not choosing a tool, but honestly assessing whether the company is ready. Answering the five questions below is enough.

Each “no” is a task to complete before or during the pilot, not a reason to give up. However, if a process follows fixed rules and does not require understanding language, ordinary automation in Power Automate, without AI, is often enough. It is cheaper, predictable and easier to maintain.

During preparation, hold short meetings with department managers in finance, sales, customer service, HR and IT. In each one, collect two or three processes that take the most time or cause the most errors. For each, note the number of cases a month, the time per case and who handles them today.

Then score the candidates on two axes: how big the benefit is and how easy the solution is to implement. For the pilot, choose a process that scores well on both. Common candidates are employee questions to HR or IT, initial handling of purchase invoices and answers to recurring customer questions.

To start with, avoid processes where an agent’s mistake has serious legal or financial consequences. One example is initial screening of job candidates, which will be subject to the AI Act requirements for high-risk systems from 2 December 2027. The questions in the table below help you quickly rule out weak candidates.

The pilot’s scope should be deliberately narrow: one process, one group of users, a few knowledge sources and one or two tasks the agent performs on its own. The aim is not to show off what AI can do. The aim is to check whether your company can run an agent day to day and measure the results.

Before you start, write down the measure of success and the threshold for stopping. For example: “the agent answers at least half of HR questions on its own, and fewer than 5% of answers in a weekly sample are wrong”. Base the exact figures on where you are today, but write them down before the pilot, not after it.

The schedule below is for a 12-week pilot. A shorter pilot usually means a smaller test group, not skipping a step. Our case study of a knowledge assistant for production engineers shows what a narrow first agent like this looks like in practice.

Some decisions cannot be handed over to a supplier or to the IT department, because they concern risk and money. They are best made separately for each process, not once for the whole company.

A company of 150 people does not need a separate AI department. A few roles are enough, and they can be filled by people who also have other duties. What matters is that each role has a named person and set-aside time.

Once several agents are running, a small coordinating team helps, for example two or three people working on it part-time. It looks after the shared rules, takes requests from departments and runs quarterly agent reviews.

After a successful pilot, it is tempting to launch several more agents straight away. Without shared rules, within a year nobody knows how many agents are running, who owns them and what data they can access. It is the same problem as Excel spreadsheets circulating around the company, only with access to systems.

The second agent works best in the same area as the first, because it can reuse existing connections. An agent that answers questions about invoices might be followed by an agent for approving purchase orders. The third could go to a new department to test whether the rules you have agreed work more widely.

In the Microsoft environment, oversight is handled in the Power Platform admin center. There you will find an inventory of all the company’s apps, flows and agents, along with data on who uses them. We have listed below the rules worth agreeing before the second agent.

An agent that nobody uses delivers no benefit, even if it works perfectly from a technical point of view. That is why getting people to use it needs to be planned as carefully as building it.

Every week, check what share of the target group used the agent and how many cases it handled without human help. Set an alert threshold before launch. If, after a few weeks, fewer people are using it than you expected, talk to the users before you start changing the technology.

The cost of an agent programme is the sum of licences, implementation work, system connections, data preparation, training and maintenance. The item most easily overlooked is your own people’s time. The process owner and the employees doing the testing work on the project at the expense of their normal duties.

In Copilot Studio, you pay for agent work with credits. A pack of 25,000 credits costs USD 200 a month and is shared by the whole organisation. Credits do not carry over to the next month, so the pack needs to match actual usage.

Employees with a Microsoft 365 Copilot licence can use some agent features in Teams, SharePoint and Microsoft Copilot without using up credits. We cover the budget items and the ROI formula in detail in our article on the cost of implementing AI.

Also add the obligations under the AI Act, the EU regulation on artificial intelligence, to the plan. The deadlines below apply to almost every company that rolls out agents.

Most of these mistakes have nothing to do with technology. It is worth going through the list before each decision to move to the next stage.