This checklist needs JavaScript to show a personalised result. You can still use the questions and suggested actions below. Your answers are not submitted.
What task are you considering? 1 I haven’t chosen a task2 I have several ideas but haven’t chosen one3 I have one task, but need to define the result4 I can describe the task and the result I need
List two or three recurring tasks. Pick one whose result you can check, then describe what goes in and what should come out. You can return to the rest of this checklist once you have a starting point.
What do you know about the work involved today? 1 I haven’t measured it2 I know roughly how often it happens3 I know its frequency and approximate time4 I have recorded the work, including corrections and follow-up
Watch a few examples from beginning to end. Note how often the task happens, the time spent and the corrections needed. That gives you something fair to compare with a trial.
Can you find the information the task needs? 1 I don’t yet know what it needs2 Some of it is hard to find3 Most is available, with a few gaps4 The information is accessible to the people doing the task
Take one recent example of the task and find everything needed to complete it. Identify missing documents or knowledge that only one person holds.
How reliable is that source information? 1 I haven’t checked it2 There are conflicting or outdated versions3 Most is current, with some exceptions to resolve4 I know which versions to use and how to handle gaps
Select the current source for each fact or rule. Mark old versions and record what should happen when information is missing. An AI tool cannot settle a policy decision for you.
How will you recognise an acceptable result? 1 I haven’t defined it2 I have a general idea3 I have an example, but some checks are unclear4 I have examples and specific checks for important facts
Take an example result and mark what must be accurate: names, dates, amounts, commitments or other important details. Decide which mistakes would make the result unusable.
Who will review the trial results? 1 No one is assigned2 Someone could help, but we haven’t agreed time3 A reviewer is agreed, but needs time or guidance4 Someone who knows the work has agreed time and clear checks
Agree who will check the trial output, what they will check and when they can do it. Keep output within the trial until the required checks are complete.
What have you discussed with the people doing the work? 1 We haven’t discussed it2 I have mentioned the idea3 We have discussed it, but questions remain4 We have discussed the task, concerns and how to give feedback
Explain the task you want to try and what is undecided. Ask which parts are difficult today, what could go wrong and what would make the trial useful to them.
Have you chosen examples for a trial? 1 I haven’t selected any2 I have one straightforward example3 I have several typical examples4 I have typical and awkward examples with known expected results
Include ordinary work and cases with missing details, unusual wording or exceptions. Work out the expected result before testing so that a convincing answer is not mistaken for a correct one.
What information can the proposed AI service use? 1 We haven’t checked the service or information2 We know some restrictions, but have open questions3 The service is checked, with a few permissions to resolve4 The intended information, accounts and permissions are agreed
Identify the exact service and plan. Check the information you intend to use, who can access it and how the service handles it. Resolve unclear permissions before adding that information; a made-up example may be enough for an initial demonstration.
What would justify continuing after the trial? 1 We haven’t decided2 We want it to save time, but have no comparison3 We have a benefit in mind, but costs or criteria are incomplete4 We have a comparison, estimated costs and a decision point
Write down the improvement you need, the costs and checking time you will count, and what would cause you to stop. Review the result against that agreement.