For your first AI pilot, pick one task that comes back every week or month, has a clear end point, and takes at least an hour of manual work each time. Choosing where to begin is often harder than choosing the technology. A strategic plan is tempting, but one concrete task shows you much sooner what an AI agent can do for your business. Below you can read how to choose that task, which tasks to leave for later, and how the pilot runs.
Technology is not the only reason an AI project can fail. Scope can get in the way too. A project as broad as 'we are going to automate everything' can stall because nobody knows where it begins or when it is done. With one bounded task, you see within a few weeks what the agent does well and where it still needs guidance.
Many businesses are still at that starting point. According to Eurostat, about one in five EU businesses with at least ten employees used AI in 2025; among small businesses, the figure was 17 percent. Businesses that considered AI but did not use it most often cited a lack of expertise as the reason.
A good workflow for a first AI pilot has three characteristics:
It comes back every week or month
It has a fairly clear scope: you know when it is finished
It takes at least an hour of manual work each time
Think of the weekly report you compile from three sources, the product descriptions you keep editing, or the document review that returns every month.
Unsure between a few tasks? List them side by side. For each one, note how often it comes back, how long it takes, when it is finished, and whether someone checks the result before it goes out. Here is what that looks like for three example tasks:
| Example task | How often | Time each time | Clear end point? | Checked before it goes out? |
|---|---|---|---|---|
Weekly sales report from three sources | Every week | 1 to 2 hours | Yes | Yes |
Answering every incoming customer email | Every day | Varies | No | No |
Monthly contract check | Every month | About 3 hours | Yes | Yes |
In this example, the weekly report is the best first pilot: it comes back often, you know when it is done, and someone checks it before it is sent. Customer email comes back even more often, but every message is different and a reply reaches the customer right away. Before the pilot starts, write down how long the chosen task takes today. That is the baseline you compare the results against.
Also check whether the data the task needs is already in one place. If someone now has to pull it from five folders and two mailboxes, tidy that up first.
Want to build a shortlist first? Use the checklist and examples in automating recurring tasks.
A task with many exceptions, sensitive personal data, or no clear owner is not a good first AI pilot. The same goes for work where a single mistake reaches a customer right away, such as an invoice that goes out automatically. Start instead with work where someone on your team reviews the result before it leaves the building. That way you learn what the agent does well while any mistake stays in-house. Add sensitive tasks later, with clear agreements on access and review.
Once you have chosen, the pilot follows the four steps of Voltti's approach:
| Step | What you do | Result |
|---|---|---|
1. Introduction | A short conversation: is Hermes a good fit for this task? | You know whether AI could add value here |
2. Define the workflow | Define the workflow: which steps are fixed, and which need flexibility? | The agent has clear instructions and limits |
3. Install | Install Hermes on the system and set up projects and access | The agent is ready to use |
4. Start using it | The team starts using it and evaluates the results | Your team works with it and can assess the output |
We work this way at Voltti too. Every content change on this website starts as a ticket that the agent picks up and ends with a verification report in that ticket. People and the agent work from the same list: the agent does the work, we review the completed work and adjust where needed. It works because each task has clear limits: one page or one article, with a fixed end point and a check you can read back afterward.
If the pilot does not work, you find out early. A pilot is an evaluation period, not a commitment to roll out. If the chosen workflow is unsuitable, we discuss what did not work and whether another task would be a better fit. A bounded test gives you something concrete to evaluate before investing in a larger project.
Need examples? AI automation for your business lists five common candidates for a first pilot, and automating quotes with an AI agent shows one task worked out step by step. Once you have picked a task, connecting AI to your CRM and CMS explains which systems the agent needs and how to limit permissions.
From the first conversation to the evaluation, a pilot usually takes 2 to 4 weeks. During that time you go through the four steps in this article: the introduction, defining the workflow, the installation, and your team's first use. How fast it goes depends mostly on the chosen workflow and the systems the agent needs to reach. The time your team can set aside to review results also plays a role.
We discuss what did not work and whether another task is a better fit. A pilot is an evaluation, not a commitment to roll out. A pilot that does not succeed still gives you something: you know which data was missing, where the exceptions were, and which agreements you need. You take that knowledge to the next candidate, so you can judge it better.
We advise against it. Start with one task, so you can see what the agent does well and where it needs guidance. What you learn can help with the next task, but you still need to assess its data, risks, and exceptions separately. Once the first workflow works, add the next one step by step.
A task with many exceptions, sensitive personal data, or no clear owner. Work where a mistake reaches a customer right away, such as an invoice that goes out automatically, is also better left for later. Start with a task where someone on your team reviews the result before it goes out. That way a mistake stays in-house while you learn how the agent works.
Mostly clarity about the task. Briefly describe what needs to happen, collect a few examples of good work, and note which systems the agent needs. Also agree on who reviews the result and when you consider the pilot a success. Prepare your AI agent pilot gives you a seven-question briefing to put this on paper in advance.
