10 things to prepare before an AI automation project

A practical checklist: pick one process, measure it, define success, find the data, collect examples, set approvals, data protection and an owner.
3 min
09 Nov 2026
  • Automation
  • Guides
10 things to prepare before an AI automation project

Most AI projects that disappoint do not fail because of the technology. They fail because nobody agreed what “done” means, the data was not where everyone thought, or nobody owned the result after launch. A few hours of preparation fix most of that. Here is the checklist we go through with clients before writing any code.

1. Name one process, not “AI in general”

Pick one process that is frequent, repetitive and well understood, for example “answering delivery questions” or “entering supplier invoices into accounting”. Write it in one sentence. If it takes a paragraph, it is probably two projects.

2. Measure how it works today

How many times a month does it happen? How long does it take each time? How often are there mistakes? Even rough numbers are fine. Without a baseline you cannot tell later whether the project paid off.

3. Decide what success looks like

Choose one or two measurable targets, such as “60% of delivery questions answered without a person” or “invoice entry time cut from 6 minutes to 1”. Agree on them with the people who will use the result, not only with management.

4. Find where the data lives

List every system involved: CRM, help desk, ERP, spreadsheets, shared drives, email. For each one note whether it has an API, who can give access, and how clean the data is. This step often reveals the real size of the project.

5. Collect real examples

Gather 30–100 real cases with the correct outcome: questions with good answers, documents with the right extracted values, tickets with the right category. They become the test set that shows whether the system works before anyone relies on it.

6. Write down the rules and the exceptions

Ask the people who do the work today what they check and when they make an exception. Unwritten rules (“we never refund orders older than 60 days unless it is a wholesale client”) are exactly what an AI system will get wrong if nobody tells it.

7. Decide where a human approves

For each action, choose: fully automatic, prepared as a draft for approval, or human only. Anything that sends money, contacts a customer in a sensitive situation or is hard to undo should start with approval. Autonomy can be increased later, based on results.

8. Check data protection

Will personal or confidential data be processed? Which AI provider will see it, in which region, under which terms? Do you need a data processing agreement or a review by your security or legal team? It is much easier to decide this before the build than after.

9. Appoint an owner

One person on your side should own the result: review the logs, update content and rules, collect feedback from users and decide on changes. An AI system without an owner slowly gets worse as your business changes around it.

10. Plan a small first version

Define a first version that can go live in weeks, not months: one process, one group of users, the main data sources. Run it alongside the current process for a short period, compare results against your targets, then expand.

A one-page brief

The answers to these ten points fit on one page. That page is enough for a vendor to give you a realistic scope and a fixed price for the first version, and for your team to see what is about to change.

Not sure which kind of solution fits your process? Start with AI agent vs chatbot vs workflow automation. When your brief is ready, or if you want help writing it, book a free 30-minute call. See our AI agents and automation and AI integration services.

    Book a free 30-min AI call