Uğur Deniz İlhan
  1. Home
  2. Insights
  3. AI Agent Manager
  4. Where should a company start with AI agents?
AI Agent Manager · Guide · 16 November 2026 · 7 min read

Where should a company start with AI agents?

The fate of a first agent project is decided by the task that is chosen, not by the tool. This guide covers how to pick the first task, the four decisions to make beforehand and a sample first-month timeline.

AI Agent ManagerFirst agent, first task
Short answerStart with a single task: one that repeats often, has clear rules, can be undone if it goes wrong and has a measurable result. Name the agent’s owner first and write the brief. In the first period the agent only makes suggestions; its authority widens as results earn trust. Choosing the tool comes after these decisions.

Why should the first decision not be the tool?

Most agent projects begin with a tool demo. The tool is chosen, then a task is found to suit it. A company that proceeds in this order builds a project around what the tool can do, not around what it needs.

The right order is the reverse. The task is chosen first, then its owner, its limits and its success measure are written down. When the tool is chosen against that sheet, it also becomes obvious which features are actually required.

How is the right first task chosen?

The first task should meet four criteria together. If one is missing, the task is left for the second or third project.

CriterionQuestion to askA fitting example
RepetitionIs this task done on a regular basis?The performance report prepared every week.
RulesCan the way the task is done be written down?Sorting incoming requests under fixed headings.
ReversibilityIf a mistake is made, is it easy to correct?A draft a person reads before it is published.
MeasurabilityCan a number say whether the task was done well?Preparation time and the number of lines that needed correcting.

The task that takes the most time, or draws the most complaints, is often the wrong candidate for a first project. Such tasks are usually hard because their rules are unclear, and an agent gets stuck on the same ambiguity.

Which four decisions should be made before starting?

  1. Owner. Who is accountable for what the agent produces? Not a team name; one person’s name.
  2. Brief. What does the agent do, what does it not do, where does it stop and ask? A text of no more than a page is enough. The agent brief guide explains how to write it.
  3. Permissions. Which data may it read, in which systems may it make changes? At the start it gets read access only.
  4. Success measure. Written before the agent starts work. How long does the task take today, how often does it go wrong? Without that number, improvement cannot be shown either.

How does the first month unfold?

The timeline below is an example; as the task grows the steps take longer, but their order does not change.

  1. Week one: the task and the baseline. The task is chosen, today’s time and error picture is recorded, the owner is named.
  2. Week two: the brief and shadow running. The agent does the task alongside a person; its output is not used, only compared with what the person produced.
  3. Week three: suggestion mode. The agent’s output starts being used, but a person approves every piece. Each correction goes back into the brief.
  4. Week four: evaluation. Results are compared with the baseline. The decision is one of three: widen the authority, continue at the same level, or stop.

What are the most common mistakes?

  • Starting many tasks at once. It becomes impossible to tell which one is failing and why.
  • Starting with the riskiest task. The first error also ends trust in the project.
  • Starting without measuring. The result is described as “seems good”; budget and support are cut at the next meeting.
  • Leaving the agent without an owner. The person who set it up moves on, and the agent keeps running as a system nobody looks at.
  • Removing human approval in a hurry. Approval is relaxed when evidence builds up, not when patience runs out. More in the article on human approval.

When is it time for a second task?

When the first task runs steadily under human oversight and meets its measure. At that point the company holds a working agent and, alongside it, a reusable method: choosing the task, the brief, the permissions and the measurement. The second task is set up far faster with that method.

The role that owns this method inside a company is known as the AI Agent Manager. To see where your organisation stands today, take the eight-question agent readiness test.

Our recommendation

A first agent project can start this week; it does not require buying a tool.

  1. List the tasks the team repeats every week and choose the one task that meets all four criteria.
  2. Record for two weeks how many hours that task takes today and how many corrections it needs.
  3. Write down the name of the person accountable for the result; that person writes the brief.
  4. Hold tool conversations after these steps are done. Show the vendor the brief and ask how the tool would do that task.

Frequently asked questions

Which tool should be chosen for the first agent?

The tool is the last decision. The task, the owner, the brief and the success measure come first; the tool is chosen against those four. When the order is reversed, the company starts looking for a project that suits the tool rather than the work.

How long does a first agent project take?

It depends on the task and the company. For a narrow task with clear rules, a first evaluation is usually possible within a few weeks. What stretches the timeline is not the technology but a vague brief and an unmeasured starting point.

Can a small team start working with agents?

Yes. What is needed is not a large team but a named owner and one well-chosen task. In small teams the chain of decisions is short, so starting is often easier.

What should be done if the first agent fails?

Look at the cause first: was the task wrong, was the brief vague, or was the measure missing? In most cases the problem sits in one of those three, not in the agent. Fix it and retry on the same task, or move to a narrower one.

Read next

All insights

For a conversation, message on WhatsApp.

A short introduction is enough: your company and what you would like to discuss.

Open WhatsApp

+90 530 242 14 55