Many first meetings about AI go the same way. Someone describes a vague wish, such as "we want to use AI in operations", and the next hour is spent asking basic questions that nobody can answer on the spot. The meeting ends with a promise to "send more information", and a week is gone.
This article is for a managing director or an operations lead who is about to have that first conversation, with an internal team, a supplier or a consultant. It describes a single page you can write in about an hour, and what each part of it is for. It is about preparation, not about choosing which process to automate. For that decision, see choosing the first AI agent process: risk, value, reversibility.
Why one page
A page forces choices. If you try to describe the whole company, you end up with a list of wishes. If you describe one process, the other side can ask useful questions and give a useful answer. One page also travels well: it can be forwarded, corrected and compared with a later version.
Keep it to one process. If you have three candidates, write three short pages and pick later.
The five items
1. The process, as it runs today
Write the steps in plain sentences: what starts it, who does what, where the information lives, and what the result is. Include the exceptions people handle by habit, such as "if the customer is from abroad, Anna checks the price by hand". Those exceptions are where most of the difficulty hides.
Describe what happens, not what should happen. A tidy diagram of the official process hides the work that actually takes the time.
2. The volume
State how often the process runs and how long one run takes. A rough count from one typical week is enough: "about 60 requests a week, 10 to 15 minutes each". Note if the volume swings, for example at month end.
Volume decides whether the effort is worth it. A task done ten times a month rarely justifies a project, and a task done hundreds of times a week often does. If you are unsure of the number, count for a single week and write down that it is a sample. To turn it into money, the cost of manual processes shows a simple way to do it.
3. The errors today
Write down what goes wrong now: wrong entries, missed deadlines, rework, complaints. Again, a rough count is fine. If nobody records errors, say so. That is useful information, because it tells the other side that success will be hard to measure unless you start counting.
This item protects you later. Without a baseline, any result can be described as an improvement, and no result can be shown to be one.
4. The owner
Name one person who knows the process and can decide about changes to it. Not a committee, and not the person who will "look at it when there is time". The owner answers questions in the meeting, approves the final description and later judges whether the result is acceptable.
If the process has no clear owner, write that. It is a common finding, and it should be solved before any tool is selected.
5. The sensitive data
List what data the process touches: customer names, contracts, prices, health or financial details, employee records. Mark what must not leave your systems or must not be sent to an external service. If you are unsure about a category, mark it as unclear and ask the person responsible for data protection before the meeting.
Knowing this early saves a lot of time. It narrows the possible solutions at once and avoids the awkward moment when an attractive proposal turns out to be impossible. For the later stage, handling customer data in AI agents goes into retention and access.
A template you can copy
| Item | Your answer |
|---|---|
| Process and its steps | |
| Exceptions people handle by habit | |
| Runs per week / minutes per run | |
| Errors in a typical month | |
| Owner (name, role) | |
| Sensitive data and what must stay inside | |
| What would count as success |
The last row is optional but helpful. One sentence is enough, such as "orders are entered the same day with fewer corrections".
How to do it in an hour
- Fifteen minutes: talk to the person who does the work and write the steps and exceptions in their words.
- Fifteen minutes: pull one typical week from whatever record exists, such as a mailbox, a shared sheet or a ticket list, and count.
- Fifteen minutes: collect the errors and complaints you remember, and ask the team for two more.
- Fifteen minutes: agree on the owner and mark the data. Then read the page once as a stranger would.
If you cannot fill an item in the hour, do not hide it. A blank line is the first honest result of the preparation, and it tells you what to settle before spending money.
What to expect from the meeting
With the page in hand, you can ask better questions: what would be the smallest useful trial, what would it need from us, and how would we know it works. A good partner will react to your numbers and your constraints. If the other side presents a solution without asking about any of the five items, treat that with caution. For a first trial of small scope, the first AI agent pilot playbook shows what a reasonable plan looks like.
Common mistakes
- Starting with the tool. Naming a product before describing the process narrows the discussion too early.
- Describing the ideal process. The real one has exceptions, and they decide the effort.
- Using precise numbers nobody trusts. A clearly marked sample is better.
- No owner. Nobody can approve the description or judge the result.
- Leaving data until later. It can end an idea in the last minute.
Checklist
- Pick one process
- Write its steps and exceptions as they run today
- Count runs and minutes for one typical week
- Note the errors, or note that nobody records them
- Name one owner
- List sensitive data and what must stay inside
- Write one sentence on what success would look like
Sources
- Our related articles linked above.
- The five-item page is a working practice, not a standard; adjust it to your context. No client results or market figures are claimed in this article.