When a team asks us where to start with AI, they usually point at the most impressive thing they can picture: an agent that sells, a bot that answers every client. We nearly always end up somewhere duller.
The Friday report. The pile of CVs. The tender notices nobody has time to read. That is not a lack of ambition. It is where the hours go, and it is the only kind of work where you can tell within a month whether AI helped.
Before we build anything, we ask four questions about each job on the team's list. The job that passes all four goes first.
1. Does it come back on its own?
Good first jobs have a clock. Every Monday. Every new CV. Every batch of tender notices. The work arrives whether or not anyone remembers it, so a process can start when the work does.
One-off work is different. Rewriting a proposal or thinking through a plan is better done in a chat window, by the person who knows what they want. If you have to explain the job from scratch each time, it is not ready to hand over.
2. Does someone already check the result?
If nobody reads the report before it goes out today, nobody will read the AI's version either. Pick work where a person already looks at the result before it reaches a client, a candidate or a manager.
That person becomes the owner. Their check stays exactly where it was. What changes is that they check a finished draft instead of building it first.
3. Can you reach what it is built from?
Most repeat work is built from email, calendars, shared files and the odd spreadsheet export. That is reachable. Work built from what one person remembers is not, at least not yet.
When the knowledge sits in someone's head, write it down first. When a bank asked us where AI could help, 75 employees wrote down their work before anyone talked about tools. Writing it down was half the answer.
4. Can you count it in hours?
Multiply how often it happens by how long it takes. A telecom's HR team told us screening CVs took four hours a day. An agency told us client reports were pulled together by hand every Friday afternoon. Those numbers are rough, and that is fine: say they are estimates. What matters is that you have one before you start, so you can compare after.
If you cannot put any number on it, you will not be able to say whether the pilot worked, and it will drift.
What we tell teams to leave for later
- Decisions with nobody reviewing them. If a wrong answer goes straight to a client, start somewhere else.
- Work that is different every time. There is no process to learn, so there is nothing to hand over.
- Work only one person understands and will not share. That is a conversation to have first, not a build.
How to start
List the repeat work your team does in a normal week. Ask the four questions of each job. Pick the one that passes all four and costs the most hours. Hand it over for a month, with one measure agreed before you begin, and decide at the end whether it stays.
Once the first process runs without surprises, the second one is much easier to choose. You will know what your team is willing to check, and what good enough looks like.