AI agents for experts: how to start without handing over the work

You hear two extreme views about AI agents. One says “it will do everything.” The other fears “I'll lose control.” For an expert who works alone or for a small firm, neither offers much direction.

There is a simpler way: leave the preparation part of a repeating task to an agent, and keep the decision with you.

I say this from my own practice. At INVAMAR, where I am a core founding developer, I set up AI agents in finance, human resources, operations and business development processes. Routine manual work went down and decisions got faster. The suggestions in this article are based on that experience.

First, a short definition: an agent is a software assistant that prepares a specific task based on instructions you write. It reads an email, gathers your notes and writes a draft. You decide whether to use the draft or not.

How do you choose the first task?

It is tempting to pick the biggest or most complex job for your first agent. Do not do that. Start with a small task with clear limits. A good first task passes these five tests:

• It repeats oftenA task you do several times a week, always with similar steps.
• Its input is clearAn email, a meeting note, a spreadsheet or a voice note.
• Its output is a draftA text or a list that you read and correct.
• You see mistakes firstYou catch a mistake in the draft, not your client.
• It really takes timeA five-minute task is not worth setting up an agent for.

These are typical tasks that pass the test:

• Logging incoming inquiries and preparing a first reply draft
• Turning notes you take in the field into LinkedIn post drafts
• Drafting a proposal from meeting notes
• Preparing short questions to ask your client when a project ends
• Putting together a one-page summary from public sources before a meeting

An example: writing a proposal after every first meeting takes you half a day. This task repeats often. Its input is your meeting notes, its output is a draft. You still set the scope and the price. It is a good candidate for a first agent.

If you have a small team, make the choice together with the person who does that task every day. They will work with the agent every day. They also know best where the work gets stuck.

The agent prepares, you decide

Working with an agent is a four-step loop:

1. 1The agent preparesIt reads the inquiry, gathers the notes and writes the draft.
2. 2You decideYou approve, correct or reject the draft.
3. 3The system carries it outThe reply you approved is sent, and the record goes into the spreadsheet.
4. 4You monitorYou update the instructions based on what you correct most often.

After the fourth step, the loop starts again. The instructions get a little better with each round.

The reason the decision stays with you is simple: agents can make mistakes. They can misread an inquiry, mix up a number or write in a tone you would never use. Your client, however, sees that email as coming from you.

Everything that goes to a client passes your approval first.

What should you not give the agent?

The easiest way is to split tasks into two groups. You can give the ones on the left to the agent. The ones on the right should stay with you, at least in the first months.

Give to the agent

• First reply drafts for incoming inquiries
• Summaries of meeting and field notes
• Proposal drafts (without numbers)
• Follow-up reminders
• A one-page summary before a meeting

Keep with you

• Pricing and scope decisions
• Contracts, payments and official correspondence
• Complaints and sensitive conversations
• Sensitive personal data such as health or salary
• Processes that are not yet clear

The last item is especially important: if you cannot describe how a task is done, the agent cannot do it either. Write the process down first, then set up the agent.

Before you start: data and privacy

Before you set up the agent, the answers to these questions should be clear and in writing:

• Which information will the agent access? Only what that task needs.
• Which tool does this information move from and to, and where is it stored?
• Do your contracts and your obligations under KVKK (Türkiye's personal data protection law) allow this? If you are not sure, consult a lawyer.
• Can you shut off any access whenever you want?
• Is there a log that lets you see later what the agent did?

If we work together, we answer these questions in writing before setup. There is no access you do not want.

A realistic first month

The goal of the first month is not a big transformation. It is getting one task to run reliably.

1. Week 1PreparationWrite the task down step by step. Collect 5 to 10 real examples from the past. Set the access rules.
2. Week 2Drafts onlyThe agent produces drafts. Compare each draft with the text you would have written.
3. Week 3CorrectionsNote which drafts you approved, corrected or rejected. Update the instructions accordingly.
4. Week 4DecisionAre the drafts useful, and did they save you time? Continue, adjust or stop.

This plan is not a promise. It is a framework. Some tasks settle in two weeks, some need a second month. Do not give the agent new work before you have seen the quality of its drafts with your own eyes.

Once the first task is running well, move on to a second one. If you set up several agents at once, it is hard to see which one is actually useful.

To see where you could start in your own work, take a look at the agents section. There I describe sample agents and how the work is split between the agent and you.

Short answers to common questions about AI automation: AI automation for experts and boutique firms.