How to use the worksheet
Fill in one worksheet per use case, with the person who does the task. Score each criterion from 1 to 5 using the descriptions, and write the evidence for each score in a sentence. A score without evidence is an opinion.
Then apply the hold-point rules. They override the totals: a high value score does not rescue a use whose errors would be costly and hard to spot.
The online scorer applies exactly the same rules.
Part A: Describe the use
| Field | What to write |
|---|---|
| Task | The one step of work the AI would do or help with |
| Done by | The role that does it today, and roughly how long it takes |
| Proposed tool | Including AI features inside software you already use |
| Information in | What the tool would be given, and its class (green, amber, red) |
| Output goes to | Who sees or acts on the result: staff, a customer, a regulator |
| Accountable owner | The person who will sign the decision |
Part B: Score six criteria
| Criterion | 1 | 3 | 5 |
|---|---|---|---|
| How much time or money does this task cost you now?Benefit: higher is better | A few minutes a month | A few hours a month | Many hours every week |
| How often is the task done?Benefit: higher is better | A few times a year | Weekly | Many times a day |
| Is the information the AI would need digital, current and in one place?Readiness: higher is better | Paper, scattered or out of date | Mostly digital, in a few places | Digital, current and in one place |
| How quickly can a competent person check the AI’s output?Readiness: higher is better | Hard to tell if it is wrong | A careful read of a few minutes | Seconds, and errors are obvious |
| If the output is wrong and nobody notices, what happens?Risk: higher is worse | A minor annoyance, easily fixed | A cost, a complaint or rework | Harm to a person, or legal or major financial loss |
| What information would go into the AI tool?Risk: higher is worse | Public or already published | Internal business information | Health, financial or other sensitive personal information |
Totals: benefit is value plus frequency, readiness is data readiness plus ease of checking, and risk is consequence plus sensitivity. Each runs from 2 to 10.
Part C: Apply the hold points
| Hold point | Triggered when | Result |
|---|---|---|
| No reliable human check | Consequence of error is 4 or 5 and ease of checking is 1 or 2 | Stop |
| Not worth the effort yet | Value plus frequency is 4 or less | Stop |
| Information rules needed first | Information sensitivity is 4 or 5 | Fix first |
| Information not ready | Data readiness is 1 or 2 | Fix first |
Part D: Decide and sign
- Stop: don’t use AI for this yet if any stop hold point is triggered.
- Fix first, then reassess if only fix-first hold points are triggered. Name the fix and who owns it.
- Go to a small pilot if no hold point is triggered. Rank “go” uses by benefit, then readiness.
For every “go”, write the conditions: acceptance criteria agreed before testing; a named person checks every output when consequence is 3 or more; a business account when sensitivity is 3 or more; and problems logged as non-conformance reports.
Decision · Signed by the accountable owner · Date
See it filled in for an invented business in the worked scenario.