Getting started with AI agents at your company: the first decisions

Pick the process before the tool

An AI agent takes over a routine. So the starting point is the routine, not the technology. The good news is that almost every company has obvious candidates: tasks that repeat every day, take hours of the team’s time and follow a reasonable pattern.

If a task repeats, it can be automated. The question is which one to start with.

To choose the first one, look for these signs:

• Volume. It happens many times a week. Automating something that happens twice a month rarely pays back the initial effort.
• Pattern. Most cases follow a similar path, even if there are exceptions.
• Visible pain. The team complains, the customer waits, or someone stays late because of it.
• Manageable risk. An agent’s mistake can be caught and fixed before it does serious damage.
• Accessible data. The information lives in systems you can connect to, not just in one person’s head.

Examples that tend to work well as a first agent: answering frequent customer questions on WhatsApp, reading invoices that arrive by email and entering them into the system, chasing late payments, confirming appointments, tracking deadlines. These are just examples. The best candidate is whatever comes up when you ask, “what does my team do every week that nobody likes doing?”.

Avoid starting with the company’s most strategic process or the one with the most exceptions. The first agent is also how the team learns to work with agents. A successful start on something smaller paves the way for the next ones.

Give it a named owner

Every agent needs an owner inside the company. That person doesn’t need to know how to code. They need to know the process and have the authority to decide how it should work.

The owner does things nobody from outside can:

• Explains the rules of the process and, above all, the exceptions
• Decides what the agent can do alone and what needs approval
• Reviews results at the start and says whether they’re right
• Flags when a rule changes, so the agent doesn’t keep following the old one
• Answers for the agent inside the company, the way they would for a team member

When nobody has that role, the project is orphaned. Each department thinks it’s the other’s responsibility, questions go unanswered and the agent goes into production with half-finished rules. It’s one of the most common reasons projects never make it past the pilot.

In practice, choose someone who already looks after the process today, whether that’s the department manager or the most experienced person on the task, and block a few hours a week on their calendar during the build.

Measure how it works today

Without a measure of the before, nobody can say whether the agent improved anything. And without that answer, it’s hard to justify the second agent.

The measurement doesn’t need to be sophisticated. One or two weeks of notes already give you a good baseline. What’s worth measuring depends on the process, but it’s usually on this list:

• How many cases come in per day or per week
• How much time the team spends on each case, on average
• How long the customer or the next team waits for an answer
• How many cases have errors, rework or complaints
• How many cases fall outside the norm and need someone experienced

Pick one main metric that already matters to the business, like response time, cost per case or cash collected on time. That will be the project’s measure of success. The others help you understand what’s going on.

A good side effect: measuring the current process often reveals problems that have nothing to do with AI, like a step nobody remembers the reason for. Fix it before automating. Automating a broken process only makes it fail faster.

Only then, the tool

With the process, the owner and the baseline defined, choosing the technology gets much easier, because now you have concrete criteria. The right questions become:

• Can the agent read from and write to the systems this process uses?
• Can you clearly define what it does alone and what waits for approval?
• Is there a record of what it saw, decided and did in each case?
• Who keeps the agent running when a system changes or a rule is updated?
• Does the data stay where it needs to, in line with the privacy laws that apply to you, such as GDPR?

Notice that none of these questions is about which AI model is the smartest. Models change every few months. What determines whether the agent works is its connection to your systems, well-written rules and someone looking after it.

The first 90 days

With the decisions made, a common path for the first agent looks like this:

• Understand. Follow the process closely, write down rules and exceptions, agree on the agent’s limits with the owner.
• Build. Connect the agent to the systems, with access only to what it needs, and test it with real, recent cases.
• Run in parallel. The agent works alongside the team without acting on its own. Every difference between what it would do and what the team did gets reviewed.
• Recommend. The agent prepares the work and a person approves it before it goes out.
• Execute. In the cases where it has proven it gets things right, the agent starts acting on its own. The rest stays under approval.

How long each phase takes varies a lot by process. What matters is not skipping steps. Each one builds the team’s trust and shows where the agent still needs to improve.

Common mistakes when you’re getting started

• Trying to automate five processes at once, with none of them ending up good
• Leaving the project to IT alone, with nobody from operations involved
• Hiding the project from the team that does the work today, instead of explaining how responsibilities will change
• Waiting for perfection before running in parallel, and so never running it
• Not agreeing on what happens when the agent doesn’t know, and finding out with a real customer

All these mistakes share the same root: treating the agent as a software purchase rather than as someone new joining the operation. When you think of it as a new team member, the right questions almost ask themselves: what will it do, who teaches it, who supervises it, how do we know it’s doing well.