10 questions before your first AI agent project
By Philipp Kant 4 min read
An AI agent project introduces a program that completes a recurring task in your systems on its own and hands uncertain cases to people. Whether it holds is usually decided before the first line of code: by ten questions about process, data, access, approvals, and success. Here they are, with the answers we want to see before we build.
1. Which process, exactly?
Not “customer service with AI” but “answer emails asking for delivery dates”. A process with a clear start (an email arrives) and a clear end (reply sent, order created). The narrower, the better the first project.
2. How often does it happen, and how long does it take?
Count for one week: how many cases, how many minutes per case? Ten orders a month do not justify an agent. Forty a day do. Without this number there is nothing to compare against later.
3. Who does it today?
The person who handles the process today knows where the special cases are: the customer who always orders by fax, the item number that has been wrong in the system for years. Without them, the agent is built for the normal case, and the normal case is rare.
4. Where does the input come from?
Inbox, portal, upload, phone note? And in which form: PDF, Excel, free text, photo? An agent handles varied inputs well. It has to be able to reach them, though.
5. Which systems does it read, which does it write?
Usually one to three: ERP, PIM, CRM. For each: is there an interface or an MCP server? May the agent write there, or only read? The routes are compared in MCP or API.
6. Is the data good enough?
An agent checking orders against an item master with gaps finds false errors. Check a sample first. Often the first project is data quality itself, see 9 product data errors AI agents catch.
7. When does a person decide?
Set the rule upfront, not by the model’s feel: on every posting, above an amount, on complaints, on uncertain recognition. This rule belongs in code. How: What is an agent harness?.
8. What may go wrong?
A misfiled newsletter costs nothing. A wrong credit note costs money, a wrong reply to a customer costs trust. Start with a process whose mistakes are cheap, and let the agent make suggestions there before it acts.
9. What does a run cost, what does it save?
A run costs model calls, usually cents. It saves a person’s minutes. Calculate with the numbers from question 2: cases times minutes times hourly rate, against running costs and upkeep. Where no agent is needed, a workflow is enough, see AI agent, RPA, or workflow.
10. How do we measure success?
Three numbers, before and after:
- Lead time: minutes from arrival to done.
- Manual work: a person’s minutes per case.
- Acceptance rate: how many of the agent’s suggestions go through unchanged?
If the acceptance rate rises over weeks, the agent can handle simple cases on its own. If it stays low, context is missing or the data is not good enough.
The ten questions as a checklist
| # | Question | Good answer |
|---|---|---|
| 1 | Which process? | One sentence, with start and end |
| 2 | How often, how long? | Counted, for a week |
| 3 | Who does it today? | A name, and the person is involved |
| 4 | Where does input come from? | Known channels and formats |
| 5 | Which systems? | With interfaces, read and write rights settled |
| 6 | Data good enough? | Sample checked |
| 7 | When does a person decide? | Set as a rule |
| 8 | What may go wrong? | Mistakes are cheap, or there is an approval |
| 9 | Cost and benefit? | Calculated with real numbers |
| 10 | How do we measure? | Three numbers, before and after |
FAQ
How long does a first agent project take? For a narrowly cut process with an existing interface, usually a few weeks until it runs with approvals. It takes longer when the data or the interface has to be created first.
Do we need an AI strategy before we start? A concrete process gets you further than a strategy. After the first project you know what your data and systems can do, and the next decision gets easier.
Which process makes a good first one? One that is frequent, whose mistakes are cheap, and whose input varies. Twelve candidates are in 12 jobs for AI agents in ERP, PIM, and CRM.
The free KI-Potenzial-Check gives a first estimate for your processes. We are happy to work through the ten questions with you: write to us, fixed price after that.
Ask an AI about this post
