BenjaminSchleier
Blog

AI

What AI agents in business really take off your plate

Benjamin Schleier · November 19, 2026 · 4 min read

An AI agent can read information from emails and documents, check data against other systems and prepare a case for the next step. On paper that sounds simple. In practice you first have to settle where the data comes from, how errors get noticed and who makes the final call.

1

The agent does the groundwork

A typical agent reads emails and PDFs, checks the details against the CRM and prepares a reply or a quote. That is the work that comes before the actual decision. It eats time, and nobody on the team enjoys it. Good candidates share a few traits. The task comes up often and runs roughly the same way each time, the input is text, and there is a reliable source the agent can check against. Two examples I like are transferring orders from a PDF into your system and answering questions about delivery status. In both cases your team gets a finished draft and decides whether it goes out as written.

2

Tasks that are a poor fit

Keep agents away from anything that hinges on judgment and relationships. A complaint from a major customer gets answered by a person, and the price negotiation is led by one too. Rare tasks are a poor fit as well, because the effort to set up and maintain the agent never pays for itself. Then there are processes nobody in the company can describe. If three colleagues handle the same inquiry three different ways, an agent mostly automates the mess. Agree on one way first, and write it down.

The same goes for your data. A CRM full of duplicates and stale contacts makes the agent confidently wrong, so clean it up before you start. For every task, ask whether a simple rule would do. If every invoice above a certain amount goes to management, you do not need a language model to route it. An agent earns its keep where the input is messy and every sender phrases things a little differently. For everything else, ordinary software has been around for decades, and it is cheaper and more predictable.

3

When a human needs to sign off

A language model sounds convincing even when the facts are wrong. So nothing should go out unchecked if it reaches a customer, moves money or makes a commitment. The agent presents a draft, and a person approves it. How strict that approval needs to be is something you decide task by task. Ask yourself what the most expensive possible mistake would be. For a delivery status update, that is an irritated follow-up question. A quote with the wrong price can mean an order you lose money on.

Every correction is useful material. Collect what your team changes in the drafts, and after a few weeks you will see where the agent keeps getting it wrong. For simple tasks where little can go wrong, you can later scale the check back to samples. Quotes, sensitive data and anything involving serious money keep the approval step. Your company carries the responsibility for the result anyway, no matter who wrote the text.

4

Start with one manageable task

Pick a single task to begin with. It should come up often enough that you have enough cases for a verdict after a few weeks. Sit down with the people who handle it today and collect twenty real cases together, including the five that annoy everyone. The difficult cases show you whether the agent is any good. It nearly always manages the easy ones.

5

Test the prototype carefully

Run the prototype on those cases and put its drafts next to what your team actually wrote. Count how many could have gone out unchanged, how many needed a correction and how many would have done real damage. The third number decides. Make sure you are allowed to use the test data for this purpose, especially if it contains customer data. Only with these numbers in hand is it worth talking about a larger rollout.

6

What changes for your team

People who used to type in inquiries will be checking drafts from now on. On paper that is less work. To some people it feels like a threat. Talk about it openly, and bring the people who do the task today into the project from the start. They know the exceptions that appear in no process description, and in the end they decide whether the agent gets used or quietly worked around. Settle one more thing before launch, namely who looks after the agent once it is live. Price lists change and products get added. An agent nobody maintains will still be answering with today's knowledge six months from now.

  • AI agents
  • Automation
  • CRM
  • Sales

Let's talk.

Write me two sentences about what you need. Or pick a time for a 20-minute video call.

What is it about?

I use your details only to reply to you. More in the privacy policy.