Short answer
Start with the workflow that is highest in volume, most repetitive, and lowest in judgement — not with the most exciting use case. In practice that is almost always one of four things: lead response, repeat customer questions, document processing, or assembling reports. Pick one, measure how long it currently takes, automate it end to end, and report the before and after honestly before starting the next.
The question behind the question
When a CEO asks where to start with AI, they are usually asking something more specific: how do I spend money on this without it becoming another initiative that quietly dies? That is a reasonable fear. The pattern is common enough to be predictable — a pilot is launched with enthusiasm, produces an impressive demo, and never reaches the people who would have to change how they work.
The failure is rarely technical. It is that nobody wrote down how the work currently happens before deciding to automate part of it.
A test you can apply this week
Score every candidate workflow on four things. You do not need software for this; a whiteboard is enough.
- Volume — how many times does this happen per week? Low volume rarely justifies engineering, no matter how annoying the task is.
- Repetition — is it the same shape every time, or does each case genuinely differ? Sameness is what makes automation reliable.
- Judgement — what share of it requires a person to weigh something? High-judgement work is where humans should stay.
- Data — does the information the system would need already exist somewhere it can reach? If it lives in someone's head, that is a different project.
High volume, high repetition, low judgement, good data availability. That intersection is where the first project belongs. It is usually unglamorous, which is precisely why it works.
Why the exciting use case is the wrong first project
The most interesting AI idea in most companies is also the one with the least structured data, the most judgement, and the highest visibility if it fails. Starting there means your organisation's first experience of AI is a disappointment, which makes the second project politically harder than the first.
Boring and finished beats ambitious and abandoned. The second project is easier to fund when the first one worked.
Measure before you change anything
This is the step almost everyone skips, and skipping it costs you the ability to prove anything later. Before the build, record how long the current process takes, how many cases it handles, and where it stalls. Without that baseline you will be arguing about whether it improved rather than by how much.
What a good answer looks like
A properly scoped first project has a named owner, a workflow map that the people doing the work agree is accurate, a written boundary describing what the system may and may not do unattended, and a metric agreed in advance. If any of those four is missing, the project is not ready — regardless of how good the technology is.
Last updated 2026-09-01 · Koderead
Related
Apply this to your own business
Eleven questions, scored in your browser, with the arithmetic shown.
Find my AI opportunities