By the end of this lesson you will be able to
- Build a shortlist of candidate processes from your actual week
- Score each one on fit, effort, and payoff
- Define a success test before starting, not after
The instinct is to pick the project that would be most impressive if it worked. That is almost always the wrong one. Pick the one you can finish, measure, and quietly rely on by the end of the month.
Finding candidates
Do not start from what AI can do. Start from where your week actually goes. The gap between those two is where most wasted AI spend lives.
List what happens every week
Not projects — recurring work. Quoting, invoicing, scheduling, chasing, reporting, answering the same five questions. Aim for fifteen to twenty lines. This hour is valuable whether or not you ever adopt AI.
Mark how long each takes
Rough hours per week, and who does it. Be honest about the small ones — a fifteen-minute task done daily is over an hour a week, which is more than most people estimate.
Mark which involve writing or reading text
This is the crude but effective filter. Text in, text out is where AI is strong. Anything that is really about a decision, a relationship, or physical work is not a first project.
Circle the ones that annoy people
Adoption is the hardest part of any AI project, and a process people already resent doing is one they will happily change. Enthusiasm is a resource — spend it where it is already available.
Scoring the shortlist
Three scores, one to five each. Do it quickly — this is for ranking, not for a business case.
| Score | Question | A 5 looks like |
|---|---|---|
| Fit | Is this transforming text you already have? | Summarising documents you supply |
| Effort | How little setup does it need? (5 = least) | Works in a chat window today |
| Payoff | How much time or error does it remove? | Hours a week, or a costly mistake |
Define the success test first
Write down, before you start, what would make you keep this and what would make you stop. Projects without a stopping condition do not fail — they linger, absorbing attention indefinitely, which is worse.
A good success test is specific and has a date:
- "By 30 September, quote follow-ups take under 10 minutes a week instead of 90, and none have gone out with a wrong figure."
- Not: "Improve efficiency in the quoting process."
- Include a quality bar, not just a time saving. Faster and worse is not a win.
- Include the stop condition: "If we are still checking every output line by line after four weeks, we stop."
Two versions of the same first project.
Chosen to impress
- "An AI assistant for the whole business"
- Needs integration with four systems
- Success is undefined
- Three months in, still being built
- Quietly abandoned; team now sceptical
Chosen to finish
- "Draft quote follow-ups"
- Works in a chat window on day one
- Under 10 min/week, zero wrong figures
- Running properly in two weeks
- Team asks what is next
The traps
- Automating an undefined process. If nobody can state the rules, AI will produce fast, confident inconsistency. Define it first — that is the project.
- Starting with the highest-stakes work. Legal, financial, and client-facing outputs need the most verification, which is the worst place to learn.
- Building before renting. Run it manually in a subscription tool for a few weeks. You will learn the real volume and the real failure rate, which is what any build estimate depends on.
- No named owner. A project everyone supports and nobody owns does not ship.
- Skipping the policy. Before staff put customer data anywhere, agree what may be pasted where. This takes an afternoon and prevents the one failure that is genuinely hard to undo.
The best first AI project is one your team stops noticing within a month, because it just became how the work gets done.
How long should a first project take?
Two to four weeks from decision to running. If the realistic estimate is longer, the scope is wrong for a first project — cut it down rather than extending the timeline.
Should we hire someone or use a consultant?
For a first project, neither. Use a subscription tool and someone internal who is already interested. Bring outside help when you are automating something specific to how your business works — that is where experience actually pays for itself.
What if the first project fails?
Then you learned something cheap, which is the point of picking a small one. Write down why. Usually it is an undefined process or a missing owner rather than anything about the technology, and both are fixable before the second attempt.
Key takeaways
- Start from where your week goes, not from what AI can do.
- Score fit, effort, and payoff — and weight effort higher than feels right.
- Write the success test, with a date and a quality bar, before you start.
- Rent before you build. Manual runs in a subscription tool tell you the real volume and failure rate.
- The best first project is boring, finishable, and owned by a named person.