Finding good AI use cases
How to spot where AI genuinely helps: the kinds of tasks AI is good and bad at, when a simple rule or a person is the better answer, discovering opportunities with users, and scoring and prioritising ideas by value, feasibility and risk.
π― By the end of this lesson you will be able to:
- Recognise tasks where AI fits well, and where rules, ordinary software or people are better
- Discover AI opportunities by studying real work, not starting from the technology
- Score and prioritise use cases by value, feasibility and risk
In this lesson
Start from the problem, not the technology
"We need to use AI" is not a use case. Good AI projects start from a real problem: a task that is slow, expensive, error-prone or impossible today. Then you ask whether AI is the right tool for it. Sometimes the answer is a simple rule, a better form or a process change, and that's a good outcome too.
Where AI fits well
| AI is usually good at⦠| Examples |
|---|---|
| Understanding messy language | Classifying emails, extracting fields from documents, detecting intent |
| Drafting and rewriting | First drafts of replies, summaries, reports, translations |
| Finding information | Answering questions from policies, manuals or past cases with sources |
| Spotting patterns at volume | Grouping feedback themes, flagging unusual cases for review |
| Assisting expert work | Suggesting test cases, reviewing code, checking documents against a checklist |
| Better without AI (or with a person) | Why |
|---|---|
| Exact calculations, totals, prices | Ordinary code is precise and cheaper |
| Clear, stable rules ("if amount > 500, needs approval") | A rule is predictable and easy to audit |
| Decisions needing accountability or empathy | People must own decisions about others' jobs, health or rights |
| Tasks where any error is unacceptable and can't be checked | AI makes mistakes; you need a way to catch them |
| Rare tasks done a few times a year | Not worth the build, evaluation and upkeep |
Discovering opportunities
- Shadow the work: watch people do the task. Where do they copy, search, re-type, summarise or wait?
- Map the process: list each step with time, volume and error rate. Look for high-volume, language-heavy steps.
- Ask about pain: "What do you spend time on that feels like it shouldn't need you?"
- Check the data: does the information the AI would need exist and can it be accessed? (Lesson 34)
- Look at existing tools: the feature might already exist in software you have.
Scoring and prioritising
| Dimension | Questions | Score 1β5 |
|---|---|---|
| Value | Time or cost saved, revenue, quality, customer impact; how many people and how often? | Higher is better |
| Feasibility | Is the data available? Can AI do it well (quick test)? How hard is integration? | Higher is better |
| Risk | Who is affected by errors? Personal data? Regulation? Reputation? | Lower is better |
| Measurability | Can we measure the baseline and the improvement? | Higher is better |
Comparing three ideas
A Β· Draft replies to delivery-status emails Β· value 4, feasibility 5, risk 2, measurable 5 B Β· Auto-approve expense claims Β· value 4, feasibility 3, risk 4, measurable 4 C Β· Summarise sales calls into CRM notes Β· value 3, feasibility 4, risk 2, measurable 3
Start with A: high value, data ready, low risk and easy to measure. C is a good second. B needs rules for most claims and human approval for exceptions, so redesign it before using AI.
The use case one-pager
- Problem and who has it, with current numbers (volume, time, errors).
- Proposed AI help and where people stay in control.
- Success metric and target.
- Data needed and whether it is accessible.
- Main risks and how errors would be caught.
- Smallest first version worth testing.
What this means for your role
Developer
Offer quick feasibility tests: an afternoon prompt experiment on 20 real examples often answers "can AI do this?" better than a long debate.
Tester / QA
Ask how errors would be caught for each idea; if nobody can say, the use case isn't ready.
Business Analyst / PO
Lead discovery: shadow work, map processes with numbers, score ideas and write the one-pagers.
Tech Lead / Architect
Assess feasibility honestly: data access, integration effort and whether a rule or existing tool would do the job.
Project / Delivery Manager
Use scoring to build a balanced portfolio: quick wins first, bigger bets once the team has proven the approach.
Scrum Master / Agile
Run short discovery workshops with the team and users, and keep the backlog of AI ideas visible and prioritised.
Try it: score three ideas
- Write down three tasks in your team that take a lot of time or cause errors.
- For each, decide whether AI, a rule, a process change or a person is the best answer.
- Score the AI candidates on value, feasibility, risk and measurability, and write a one-pager for the top one.
Key takeaways
- Start from a real problem; AI is one possible answer alongside rules, process changes and people.
- AI suits language-heavy, varied, high-volume tasks where errors can be checked or are low-cost.
- Discover opportunities by studying real work and data, and prefer assist-before-automate.
- Score ideas on value, feasibility, risk and measurability, and capture each in a one-pager.
Key terms
- Use case
- A specific task or problem where a solution, such as AI, is applied for a defined group of users.
- Human-in-the-loop
- A design where people review or approve AI outputs before they take effect.
- Feasibility
- How practical it is to build something, given data, technology, skills and effort.
Check your understanding
Answer all 5 questions. Score 3 or more to complete the lesson.
1. Which task is the best fit for AI?
2. Expense claims follow clear written rules. What is the best approach?
3. Where should AI use case discovery start?
4. What does "assist before automate" mean?
5. Which idea should usually be prioritised first?