How to implement an AI agent in your company: a practical guide from choosing a process to the decision after the pilot

An AI agent is a program based on a language model – the same technology as ChatGPT – that does more than answer questions. It can also reach into company documents and systems on its own, for example reading an email, checking stock levels in the ERP and creating an entry in the sales system. It works within the limits you set for it.

This guide is about one agent for one process. If you are planning a programme that spans many departments, with roles and rules for the whole company, also read our article on rolling out agents across an organisation. Here we focus on how to run your first project from start to finish.

Throughout the article, we use one hypothetical example. An electrical supplies wholesaler employs 150 people and receives about 600 requests for quotation (RFQs) a month by email. A salesperson spends an average of 15 minutes on each: reading the request, looking up products in the ERP and retyping the items into the sales system.

The best first process is neither the most spectacular nor the most ambitious. It has a lot of similar cases, knowledge that is recorded in documents or systems, and an effect you can measure. An agent’s error should be easy for a person to catch before it does any harm.

Compare two or three candidates against the same criteria. In the table below, the wholesaler from our example assesses three processes. Handling RFQs wins: many cases, data in the ERP, and the salesperson reviews the quotation before sending it anyway.

Before anyone starts building, the person responsible for the process writes an agent charter. This is a short document that sets out what the agent does, what data it works with and how you will know it works. Without it, everyone involved in the project imagines a different agent.

In our example, the goal is to cut the time spent handling one RFQ from 15 to 5 minutes, with at least 90% of items correctly identified. The 90% threshold is set by the sales director, not the supplier, because the director knows how many corrections the salespeople will accept. The list below shows what the charter should contain.

The most important decision in the agent charter concerns its autonomy. The greater an action’s impact on customers, money or the company’s commitments, the more human control it needs. Autonomy is best increased gradually, when pilot results justify it.

The table below shows four levels that work well in practice. At the wholesaler, the agent starts at level two: it prepares a draft quotation in the sales system, but a salesperson always sends it.

An agent’s answers are only as good as the sources you give it. At the start, connect less but better: the current catalogue, price list and a few instructions rather than the entire company drive. Every source must have a person responsible for keeping it up to date.

Grant access to systems on the principle of least privilege. The wholesaler’s agent needs to read product data in the ERP and create draft quotations in the sales system. It does not need access to financial data or the right to delete records.

Also decide on whose behalf the agent acts. It can use the permissions of the employee working with it, or it can have its own account with narrow access. The second option makes control easier, because the system history shows what the agent did and what a person did.

The best test is real cases the company has already handled. Choose 100–200 requests from recent months, together with how the salespeople dealt with them. The agent processes them again, and you compare its output with the people’s decisions.

The test set must not contain only simple cases. Add requests that are incomplete, contain typos or include products outside the catalogue, and ones where the right response is to hand the case to a person. These are the cases that show whether the agent is ready.

Keep the same set for later. After every change to the instructions, knowledge sources or AI model, you run it again and check that nothing has got worse. The list below shows what to measure in the tests.

When the test results meet the threshold in the agent charter, run a pilot with a few people for a few weeks. In our example, the agent works for four salespeople who assess its draft quotations every day. Once a week, the person responsible reviews the errors and improves the instructions or sources.

The whole project, from choosing the process to the end of the pilot, usually takes a few months. Most of the time goes on preparing the knowledge and reaching agreement with departments, not on building the agent itself. Below is the calculation of the effect for the wholesaler, assuming the agent meets the goal in its charter.

An agent must leave a trail. Record what task it was given, which sources it used, what actions it took and when it handed a case to a person. Without this, you cannot explain an error or improve quality.

The AI Act imposes several obligations that also apply to a first agent. Since 2 February 2025, companies that use AI have had to take steps that support their employees’ knowledge of these tools (Article 4). From 2 August 2026, an agent that talks to customers must tell them it is an AI system. An agent that supports recruitment or employee assessment is a high-risk system, with obligations applying from 2 December 2027.

If the agent processes personal data, the GDPR applies. Check the data processing agreement (DPA) with the platform provider, where the data is processed and whether the provider uses your data to train models.

The cost has four parts: licences or consumption-based fees, building the connections to systems, your own people’s time, and ongoing care after launch. You can check licences in the price list; the other elements depend on the process and the state of your data.

In the Microsoft ecosystem, agents are usually built in Copilot Studio, and their work is paid for with credits: a pack for the whole organisation or by actual consumption. Employees who will use the agent in Microsoft Copilot may need a licence. According to the price list, Microsoft 365 Copilot Business costs USD 21 per user per month with annual billing.

The most often overlooked cost is the time of the person responsible for the agent after launch. Someone has to review errors every week, update sources and rerun the tests after changes. We set out a detailed cost breakdown and how to calculate the return in a separate guide on budgeting for AI implementation.

A pilot ends with a decision, not a report. Compare the results with the agent charter and choose one of three paths. Stopping a project honestly is also a good outcome, because it protects the budget from further spending.

Before extending the agent, decide who will maintain it day to day and how quickly they will respond to problems. An agent that nobody looks after starts giving out-of-date information within a few months, because price lists and products have changed.

First projects rarely fail because of a weak AI model. More often, they fail because decisions that should have been made before building the agent were never made. These are the mistakes we see most often: