Which tasks are suitable for AI? A framework for managers

"Can we use AI for this?" is one of the most common questions a manager now gets. The honest answer is nearly always "it depends", which is no help to someone who wants to get on with their work.

This guide gives you a way to turn "it depends" into a decision. It uses eight questions about the task, a simple traffic-light rating, and two rules that override everything else. Then it works through six real-looking tasks from Fernway, the fictional office equipment company used across this site.

The framework helps you decide whether AI is worth trying. It doesn't replace your organisation's policy. Your organisation's policy and approved tools still take priority.

The eight questions

For any task, ask these eight questions. Each one pushes the task towards AI or away from it.

The traffic-light rating

Rate each question for the task as one of three labels:

• ● LOWER RISK the task sits well on this question
• ▲ CAUTION manageable with the right tool, source or checker
• ■ HIGH RISK a real problem on this question

Then apply two override rules before you count anything.

If neither override applies, read the overall rating like this:

This is a thinking tool. It deliberately avoids a points total, because one bad rating can matter more than five good ones. A task that is repetitive, quick to check and full of useful source material is still wrong for AI if the information can't go into your tools.

Six worked examples from Fernway

Priya Shah, Head of Operations, has been asked by Erin Vaughan, the Managing Director, which tasks across the business are worth trying with AI. Fernway has Microsoft 365 Copilot Chat on work accounts, approved for internal business information but not for sensitive personal data. Priya takes six tasks people have asked about.

1. Drafting the weekly operations update

Leah Bennett writes a weekly update for the operations team: deliveries, stock issues, supplier problems, and what's coming next week. She works from her own notes and the delivery tracker. It takes her most of an hour.

• Repetition: ● LOWER RISK every week, same structure
• Data sensitivity: ● LOWER RISK internal only, fine in the approved tool
• Consequence of error: ● LOWER RISK internal readers, easy to correct
• Need for judgement: ● LOWER RISK mostly structure and wording
• Source material: ● LOWER RISK her notes and the tracker
• Reversibility: ● LOWER RISK a correction can go out the same day
• Time saved: ● LOWER RISK long to write, quick to check
• Human review: ● LOWER RISK Leah knows the facts and can check every line

Overall: ● LOWER RISK A good candidate. Leah drafts from her notes in Copilot Chat and checks every date, supplier name and figure against the tracker before sending.

2. Summarising supplier contracts

Dan Okafor in Finance wants a one-page summary of each of Fernway's twelve supplier contracts: renewal dates, notice periods, price review terms and exit clauses. The contracts are long and written in legal language.

• Repetition: ▲ CAUTION twelve contracts, then occasional updates
• Data sensitivity: ▲ CAUTION commercially confidential, approved work tool only, and check whether any contract restricts sharing its terms
• Consequence of error: ■ HIGH RISK a wrong notice period could mean missing a chance to exit or renegotiate
• Need for judgement: ▲ CAUTION extracting terms is mechanical, interpreting them is not
• Source material: ● LOWER RISK the contracts themselves
• Reversibility: ▲ CAUTION an error may not be noticed until a deadline has passed
• Time saved: ● LOWER RISK reading twelve contracts takes hours
• Human review: ▲ CAUTION Dan can check each term against the clause, if the summary cites it

Overall: ▲ CAUTION Possible, with conditions. Use the approved tool. Ask for each term with the clause number it came from, so Dan can check every date and notice period against the original. Anything that needs interpreting goes to whoever advises Fernway on contracts. The summary is a finding aid. The contract is still the authority.

3. Shortlisting CVs

Tom Elliott, Sales Lead, has 60 applications for a sales support role and asks whether AI could shortlist them.

• Repetition: ▲ CAUTION occasional, in bursts
• Data sensitivity: ■ HIGH RISK personal data about applicants, which may reveal sensitive details such as health or ethnicity
• Consequence of error: ■ HIGH RISK an unfair rejection affects a real person and may be discriminatory
• Need for judgement: ■ HIGH RISK weighing people against a role
• Source material: ▲ CAUTION CVs vary widely, and the tool may reward style over substance
• Reversibility: ■ HIGH RISK rejected applicants rarely come back
• Time saved: ● LOWER RISK reading 60 CVs takes time
• Human review: ■ HIGH RISK checking an AI shortlist properly means reading every CV anyway

Overall: ■ HIGH RISK Override 2 applies: this is a decision about people. Tom reads and shortlists the CVs himself, against criteria he writes down first. AI can help with the parts that don't involve applicants' data: drafting the job advert, suggesting interview questions that match the criteria, or writing a neutral template reply. Any use of AI in recruitment should be agreed with HR and your data protection contact first.

4. Answering internal policy questions

Priya owns Fernway's hybrid working policy and gets asked the same questions again and again: how many office days, how to request a change, what the anchor day is. She wonders whether an AI assistant could answer them.

• Repetition: ● LOWER RISK the same questions, many times
• Data sensitivity: ● LOWER RISK the policy is an internal document, fine in the approved tool
• Consequence of error: ▲ CAUTION a wrong answer about working arrangements could cause a real dispute
• Need for judgement: ▲ CAUTION most questions are factual, some are requests for an exception
• Source material: ● LOWER RISK the policy document itself
• Reversibility: ▲ CAUTION a wrong answer can be corrected, if someone notices
• Time saved: ● LOWER RISK frees Priya from repeated questions
• Human review: ▲ CAUTION nobody checks each answer live, so the testing has to happen up front

Overall: ▲ CAUTION A good fit for a shared assistant built on the policy, with conditions. It answers only from the attached policy and cites the section. It says clearly when the policy doesn't cover something. It sends any request for an exception to Priya. Test it against the real questions before anyone relies on it, and keep it updated when the policy changes. The Team Brain course covers this build step by step.

5. Writing performance review feedback

Priya has six annual reviews to write and wonders whether AI could draft the feedback.

• Repetition: ▲ CAUTION once or twice a year per person
• Data sensitivity: ■ HIGH RISK personal data about named employees, possibly including health or absence
• Consequence of error: ■ HIGH RISK affects pay, progression and trust
• Need for judgement: ■ HIGH RISK the whole point of a review is the manager's judgement
• Source material: ▲ CAUTION Priya's notes, objectives, examples
• Reversibility: ■ HIGH RISK a review goes on file and shapes a relationship
• Time saved: ▲ CAUTION writing is slow, but the thinking can't be delegated
• Human review: ▲ CAUTION Priya can check, but may be tempted to accept fluent wording that isn't quite what she means

Overall: ■ HIGH RISK Override 2 applies to the ratings and judgements, which stay entirely Priya's. A much smaller task can score better. Once Priya has written her own feedback, she can ask the approved tool to make her wording clearer or more specific, with names and personal details removed, and only if her organisation's policy allows it. The judgement, the examples and the rating stay hers.

6. Reconciling invoices

Grace Lynch, Finance Assistant, reconciles supplier invoices against purchase orders each month. She asks whether she could paste both lists into AI and ask it to find the mismatches.

• Repetition: ● LOWER RISK every month
• Data sensitivity: ▲ CAUTION commercial data, approved tool only, and never bank details
• Consequence of error: ■ HIGH RISK paying the wrong amount, or missing an overcharge
• Need for judgement: ● LOWER RISK matching is mechanical
• Source material: ● LOWER RISK two clear lists
• Reversibility: ▲ CAUTION payments can be recovered, with effort
• Time saved: ■ HIGH RISK every match has to be checked, which is the task itself
• Human review: ■ HIGH RISK checking a chatbot's matching line by line takes as long as doing it

Overall: ■ HIGH RISK AI chat tools can misread rows and columns and give totals that look right but aren't. Exact matching is a job for a spreadsheet formula or the finance system.

The six at a glance

Decisions about people need extra care

Two of the six examples involve decisions about people. They get their own override because the stakes are different.

A decision about someone's job, pay, progression or conduct affects their life. They are entitled to a fair process and, often, to an explanation of how the decision was made. AI tools can reflect biases in the material they learned from. They can also produce confident reasons that sound plausible but aren't how the decision was actually reached. A manager who relies on AI for these decisions may be unable to explain them honestly.

UK data protection law has specific rules about decisions made about people by automated means. The ICO's guidance on explaining decisions made with AI sets out what organisations should consider. The practical rule for a manager is simpler: a named person makes the decision, can explain the reasons in their own words, and has looked at the evidence themselves. Anything beyond admin support around people decisions should be agreed with HR and your data protection contact first. This guide is not legal advice.

Using the matrix with your team

Download the task suitability matrix. It opens in Excel or Google Sheets and includes the six Fernway examples above, marked as fictional, plus empty rows for your own tasks.

A good way to use it:

1. Ask each person to list three tasks they'd like help with.
2. Rate them together in a team meeting. The discussion is as useful as the result, because people see why some tasks score differently.
3. Pick one or two lower-rated tasks to try first. Time them by hand before you start, so you can tell whether AI really helps: see is AI actually saving time?
4. Revisit caution tasks once you have an approved tool, better source material or a clear checker.

I'm a manager deciding whether a team task is suitable for AI. Rate the task below against these eight questions, using Lower, Caution or High for each, with one sentence of reasoning: repetition, data sensitivity, consequence of error, need for judgement, available source material, reversibility, time saved once checking is counted, and whether a person can realistically review the output. Then say whether either override applies: (1) the information can't go into an approved tool, or (2) the task decides something about a person. Finally, suggest a smaller part of the task that would rate lower, if there is one. Where you would need more information to judge, say so.

The task: [describe the task, who does it, how often, what information it uses and where the output goes. Do not include real names or personal details.]

Treat the AI's rating as a second opinion to argue with. You know the task, the people and the consequences. It doesn't.

Common mistakes

Rating the tool instead of the task. "Copilot is approved, so everything's fine" skips the questions about consequence, judgement and review.

Looking only at time saved. A task that saves an hour and occasionally sends a wrong price to a customer may not be worth it.

Forgetting that tasks can be split. Most tasks rated high contain a smaller part that rates lower. Look for it.

Over-trusting a good first run. A task that worked well three times can still go wrong on the fourth. Keep the checker in place.