Work through these eight items in order before spending money on AI, yours or anyone's. The order matters: every expensive AI failure I have seen traces to skipping an early step, usually the measurement one. Bookmark this page. It is written to be used, not read once.
1. Identify the repeated work
Check for: a task that happens every week, takes real time, and has a clear definition of done. Client intake. Invoice entry. Meeting notes. The weekly numbers report.
Why this matters: AI pays off on repetition, not brilliance. A task done once is a chat question. A task done fifty times a year is a system. Repeated work is an operating problem before it is an AI problem, and naming the exact task is what turns an AI interest into an AI project.
2. Measure the time cost
Check for: a written estimate of hours per week the task takes, and who spends them. Not a guess from memory. Watch it happen once and write the number down.
Why this matters: this number is the denominator of every later decision. Two hours a week at $30 an hour is about $3,000 a year. That might justify a $450 setup. It might not justify a $4,500 one. Without the number you are choosing by feel, and feel is what vendors price against. Full context in the AI cost breakdown.
3. Check data privacy requirements
Check for: what the task touches. Client confidences, contracts, personnel matters, health or financial records, anything under an NDA or professional standard. Write down which categories exist.
Why this matters: this single item decides local versus cloud, which is the biggest cost fork in the whole project. Private categories do not go into a consumer chat window, full stop, which means a subscription alone can never automate this task. A local setup, private data on your own machine, can. Deciding this early prevents the classic failure: a paid subscription everyone is afraid to actually use.
4. Choose local versus cloud, on purpose
Check for: a written answer to the three-question rule. Does the task repeat weekly? Does it need private documents? Does an error cost real money or trust? Any yes to the second or third pushes toward local or hybrid.
Why this matters: "local versus cloud" is where vague AI plans die, because both options are defensible and neither is automatic. The honest framing: cloud subscriptions win on raw capability, local setups win on privacy, repeatability, and cost per use at volume. You can also land on hybrid, subscription for exploration and a local machine for the repeated private workflow, which is where most small businesses end up. The ChatGPT cost-benefit analysis has the full comparison.
5. Pilot with ten examples
Check for: ten real instances of the task, run through the candidate setup, checked by the person who normally does the work. Score them: acceptable as-is, acceptable with small edits, or wrong.
Why this matters: ten examples is small enough to cost an afternoon and large enough to expose real failure modes. Demos survive on one hand-picked example; systems survive boring, typical, messy ones. If the pilot scores below seven out of ten acceptable, the honest next step is fixing the setup or dropping the task, not widening the pilot until it looks good. For higher-stakes work, reliability can be measured properly rather than sampled. The public delegation-bench method, 495 trials with certified floors, is the industrial version of this same checklist item.
6. Set the owner-approval rule
Check for: a written line stating what the AI may prepare and what a human must approve before it counts. Customer emails, invoices, anything public, anything with money attached.
Why this matters: The owner approval rule is the difference between delegation and abdication. AI should prepare better work for review. It should not silently publish claims, send customer messages, or make taste decisions for an owner-led business. Deciding this before launch costs an hour. Deciding it after the first bad auto-send costs a customer.
7. Measure the result against the baseline
Check for: the same metric you captured in step 2, measured again after four weeks of real use. Hours spent. Turnaround time. Error rate the human checker found.
Why this matters: without a before number there is no after number, and the project becomes a matter of opinion. Opinions cut both ways: enthusiasts declare victory on vibes and skeptics declare failure on vibes, and neither can be argued with. A written baseline ends the argument in either direction. This is also the discipline that makes the next project cheaper to decide.
8. Decide: scale or stop
Check for: a scheduled decision, a week after measurement, with three possible outcomes: scale to the next workflow, keep as-is, or stop and write down why.
Why this matters: "stop" is a real outcome and a good one. A business that runs one automation well and refuses the second until the numbers justify it is ahead of a business running five nobody measures. The written reason for stopping is an asset. Prices and capabilities change every year, and this year's no is next year's pilot, but only if you remember why.
The whole checklist on one screen
- Name one task that repeats weekly with a clear definition of done.
- Write down the hours it costs and who pays them.
- List the private data categories it touches.
- Answer the three questions and choose local, cloud, or hybrid, in writing.
- Pilot on ten real examples and score them. Below seven of ten acceptable: fix or drop.
- Write the owner-approval line: what AI prepares, what a human approves.
- Re-measure after four weeks against the baseline.
- Schedule the decision: scale, keep, or stop, with the reason written down.
Total cost of the checklist itself: an afternoon and zero dollars. Most businesses that complete it honestly either save the money they were about to spend or spend it with a number attached to the expectation. Both are wins.
FAQ
Can I do this with an outside consultant?
Yes, and steps 1 through 3 should be done by you regardless, because nobody else knows your week. If you want the whole thing pressure-tested in one sitting, the PuenteWorks service paths start with a $175 Setup Hour that ends in the written recommendation this checklist produces.
What if I get stuck on local versus cloud?
Default to hybrid and stop worrying. Subscription for exploration, local for the repeated private workflow. It is the most common landing point, and the Local AI Setup walkthrough shows what the local half involves.
How long should the whole cycle take?
Estimate: a small business running this checklist end to end, first task only, takes four to six weeks including the four-week measurement window. Anything faster usually means step 7 got skipped.