Lesson 2: Mapping Team Work into Good and Bad Use Cases

Why Map Work Before Choosing Tools

In Level 1 you saw that AI adoption changes tasks. Whole jobs stay in place. The useful question is therefore not "Can AI do customer service?" Ask instead: "Which of the 40 things my team does each week could AI help with, and which should it never touch?"

Think of it like a builder's survey before a renovation. You check each wall to see whether it holds the building up before you knock anything down. This lesson gives you that survey for your team's work.

Running case: You lead a team of nine people in customer service and admin. They answer emails and calls, update records, process refunds, handle complaints and produce weekly reports.

Step 1: Build a Task Inventory

DO: List the tasks with your team, not for them. They know the hidden work, such as chasing missing documents or fixing duplicate records.

1. Run a 45-minute session, or ask everyone to log their tasks for one week.
2. Write each task as a verb plus an object, for example "Draft reply to delivery-delay email". "Emails" is too broad.
3. Record four facts for each task:
   • Frequency: how often it happens (daily, weekly, monthly).
   • Time: typical minutes per instance.
   • Data sensitivity: none, internal, personal data, or special category/financial data.
   • Customer impact: what happens to the customer if it goes wrong (minor annoyance, or money, access or wellbeing at stake).

Example rows:

Step 2: Score Each Task on Four Criteria

DO: Give each task a score from 1 (low) to 3 (high) on each criterion. Agree the scores as a group so they are consistent.

1. Value: time saved multiplied by frequency, plus any gain in quality or consistency. Sixty six-minute replies a day is high value. A monthly five-minute task is low.
2. Risk: how much harm an error would cause, combining data sensitivity and customer impact.
3. Error visibility: how easily a person can spot a mistake in the output. A wrong total in a short summary is easy to catch. A subtly unfair complaint decision is hard to catch.
4. Judgement needed: how much the task depends on discretion, empathy, context or accountability for a decision about a person.

Useful rule: when risk is high and errors are hard to spot, the task fails the test. A high value score does not change that.

Step 3: Place Tasks on a Use-Case Matrix

DO: Draw a grid with Value on one axis and Risk on the other. Then use error visibility and judgement to decide each task's colour.

• Good (green): high value, low risk, errors easy to spot, low judgement. Examples: summarising ticket themes, first drafts of standard replies that a person reviews, tidying meeting notes, reformatting reports.
• Risky (amber): worth doing only with a human check on every output. Examples: drafting replies that contain personal data, suggesting which category a ticket belongs in, drafting a response to an upset customer for an agent to rewrite.
• Unsuitable (red): AI should not lead. Examples: deciding complaint outcomes, refund or compensation decisions, contact with vulnerable customers, anything where the customer needs a person to be accountable.

Low-value, low-risk tasks are not urgent. Park them. Spend early effort on green tasks with high value.

Step 4: Explain Why High-Risk Tasks Stay Human-Led

DO: For every red task, write one or two sentences giving the reason. A reason you can explain builds trust. "Because I said so" does not.

Three reasons apply in most teams:

1. Decisions about people carry accountability. A complaint decision affects a customer's money and their trust in you, so a named person must own it. In UK data protection law, decisions made solely by automated means that have legal or similarly significant effects need safeguards. Since 5 February 2026 these have been set out in Articles 22A–22D of the UK GDPR, introduced by the Data (Use and Access) Act 2025. They include the right to obtain human intervention and to contest the decision. The ICO consulted on draft guidance from March to May 2026. Your company's policy will set the detail.
2. Vulnerability needs human attention. The FCA's guidance FG21/1 (February 2021) groups the drivers of vulnerability as health, life events, resilience and capability. Bereavement or a new caring responsibility can change what a customer needs from a call. Even outside financial services, this is a sound model: people in difficult circumstances need a person who listens and adapts.
3. Errors are hard to see. If a reviewer cannot easily tell that an output is wrong, a quick review does not protect anyone.

AI may still support red tasks at the edges. For example, it can retrieve the relevant policy for a complaint handler. The decision and the conversation stay with a person.

Template: Use-Case Shortlist

DO: Fill this in and share it with your team for comment.

• Green shortlist (top 3–5): task, value score, why errors are easy to spot.
• Amber list: task, the human check required, who does it.
• Red list (human-led): task, reason in plain words, any limited support role for AI.
• Open questions: tasks the team disagreed on, to revisit.

Your green shortlist becomes the input for choosing pilots in the next level.

Key Takeaways

• Map tasks, not jobs. Write each one as a verb plus an object, and build the list with your team.
• Record frequency, time, data sensitivity and customer impact for every task.
• Score value, risk, error visibility and judgement from 1 to 3.
• High risk combined with errors that are hard to spot means red, however much time AI could save.
• Complaint decisions and vulnerable-customer contact stay human-led. Write down why, so your team can see the reasoning.