You do not need to switch systems to get rid of manual work. An agent is built on top of the systems you already have and does the job between them: it reads, compiles, reminds and writes drafts. Start with a task that repeats every week, give the agent access only to what it needs and let a person approve what goes out.
Where the manual work usually sits
Most mid-size companies already have several systems, and each one works for its purpose. The manual work appears where information has to move between them. You will recognise it if you call two people to find out how a project is going, if the same summary is made by hand every Friday, or if important decisions sit in a chat group only a few people read.
The information already exists. What is missing is someone who gathers it, and that is a job for an agent.
One example is a weekly summary of the status across several projects. Today one person collects the details from several places and writes them up. An agent can make the first draft, and the person corrects and approves it.
How to pick the first task
Choose a task that meets these five conditions:
- It comes back every week, so the effect shows quickly.
- It has an owner, a person who cares how it works.
- The result can be measured, for example time saved or the number of reminders no longer written by hand.
- The risk is low, no sensitive decisions move over to the agent.
- Someone wants it, start where the will is, and you get a first example that others want to copy.
Feel free to write down several candidates and begin with the one that is most regular and least risky. The others wait for the next step.
What the steps look like
- Discovery, a few conversations about how the work really runs. The result is a short plan and a price for the next step.
- Pilot, one real project and an agent that helps from the first day.
- Stop point, you review the result together and choose to continue, change course or stop.
- Widen, more projects and more features, at the pace people actually use them.
The price of every step is agreed before the step starts, as a fixed price or an hourly rate with a cap, and you can end at any stop point. You do not need to take more decisions than that at once.
Keep control
Decide in advance what the agent may see and what it may do. Let it leave drafts, lists and suggestions, and let a person approve what goes out to customers or suppliers. Sign a data processing agreement before anything starts, and begin with information that is not sensitive.
The information stays in your systems. You do not have to teach everyone a new program, and you do not have to move data from one place to another.
Common mistakes
Four mistakes come back in almost every first project:
- Starting with the hardest task, choose a small one that can be finished in a few weeks.
- Giving the agent access to everything, give it only what the task requires.
- Skipping the owner, without a person who owns the task you cannot tell whether the result is good.
- Deciding the measures too late, decide before the pilot what you will look at, and the stop point becomes a real decision.
Measure that it works
Decide before the pilot what you will look at. Three measures go a long way: how much time the task took before and now, how many drafts from the agent were used, and how confident your staff feel with it. Measure use and time saved, and leave the feature count out. Then the stop point becomes a real decision.
Common questions
Which systems work?
That is decided in discovery. We go through the systems you use and what can be read from them.
Do we need our own technical staff?
No, but someone on your side needs to be able to say what works and what does not.
How long does a pilot take?
Think weeks, not years. The length is set in discovery.
What does it cost?
The work is divided into steps. The price of each step is agreed in advance, as a fixed price or an hourly rate with a cap, and there is a stop point after every step. Read more in the guide on what an AI consultant costs.
How is personal data protected?
You sign a data processing agreement before the start, decide which data the agent may see and begin with information that is not sensitive.
Keep reading
- Agentic Operations How it works, step by step.
- Where to start with AI in a mid-size company One problem, one pilot project and a stop point after every step.
- What is an AI agent and what can it do in a company A plain explanation, five examples and what an agent should never do alone.