Short answer: pick generative AI use cases where the task is frequent, mostly language, easy to check, and cheap to get wrong. Score each candidate on value, feasibility and risk before anyone writes a prompt. Most companies start with the opposite: the most impressive idea, which is usually the hardest to verify.
The three questions that filter 80% of ideas
- Is it frequent? A task done 200 times a month repays the setup; a task done twice a year does not.
- Can a person check the output quickly? If checking takes as long as doing, you saved nothing.
- What happens when it is wrong? A bad first draft costs minutes. A wrong number in a contract or a medical note costs much more.
Score every candidate
| Criterion | 1 (low) | 5 (high) |
|---|---|---|
| Value | Saves minutes per month | Saves hours per week or unlocks revenue |
| Frequency | A few times a year | Many times a day |
| Ease of checking | Needs an expert and a long review | Anyone can confirm in seconds |
| Data readiness | Data scattered, sensitive, undocumented | Clean, accessible, allowed to use |
| Risk if wrong (inverted) | Legal, financial or safety impact | Easy to fix, internal only |
Multiply value by frequency, add the other three, and start with the top three. Keep the spreadsheet: it becomes your AI roadmap and your answer when someone asks “why not this other idea?”.
Good first use cases
| Area | Use case | Why it works |
|---|---|---|
| Operations | Summarize tickets and suggest the next action | Frequent, easy to check, low risk |
| Sales | Draft follow-up emails from call notes | The rep reviews before sending |
| Finance | Explain variances in a monthly report | Numbers come from the system; AI writes the narrative |
| HR | First draft of job descriptions and interview guides | Template-heavy, reviewed by a person |
| Engineering | Explain legacy code and write tests | Tests prove whether the output is right |
| Management | Turn meeting notes into decisions and owners | Participants can confirm immediately |
Avoid starting with a public-facing chatbot. It is the hardest case to verify, the most visible when it fails, and the one with the most data and legal questions.
From idea to pilot in four weeks
- Week 1: pick one use case, define what “good” looks like and collect 20 real examples.
- Week 2: build the simplest version (often a prompt template, not an application) and test it on the examples.
- Week 3: five real users, with a feedback button and a log of what they changed.
- Week 4: compare time and quality against the baseline and decide: stop, iterate or scale.
Related: how to measure the ROI and a template for your AI usage policy.