What never goes in an AI tool
Every business needs this list. Five categories, specific to you, written down. It takes about twenty minutes and it's the highest-value twenty minutes available.
The five categories
- Identifiable records about the people you serve. Names, records, case detail, anything that would identify someone to a person who knows them.
- Credentials and access. Passwords, keys, codes, account numbers — including inside screenshots, which is how it usually happens.
- Material held under someone else's terms. Client confidentiality clauses are often tighter than privacy law and rarely read before the tool arrives.
- Work where a wrong output is a regulated harm. Clinical decisions, advice, tax positions, safety sign-off, award interpretation.
- Anything you couldn't check. If nobody in the business can verify the output, it shouldn't be produced this way.
Write yours like this
Take the five above and make each one concrete for your business, using your own words for the things you hold. “No participant progress notes with names in them” is a usable rule. “Protect sensitive data” is not.
De-identifying without wrecking the document
- Roles instead of names: the tenant, the participant, the referrer.
- Ranges instead of dates and ages where precision isn't doing work.
- Drop the geography when the geography narrows it to one person.
- Keep the clinical, technical or factual detail — that's the part making the output useful.
The test
Would someone who knows your business be able to tell who this is? If yes, it isn't de-identified, regardless of whether the name is in there.
What to do when it's already gone in
It happens, usually on a personal free account, usually with good intentions. The useful response is to find out the extent, check whether your notifiable breach obligations are engaged, fix the account and tier problem, and write the rule. Treating it as a disciplinary matter first reliably produces incomplete information.
The two columns
Usually fine, once the rule names it
- Tenders, capability statements, scopes of works
- Policies, procedures, position descriptions
- Induction and training material
- Standard letters explaining a process, with no client in them
- Job ads and internal summaries
- Reports built from properly de-identified inputs
Never, regardless of tier
- Client, patient or participant identifying information
- Anything under a confidentiality obligation you can't vary
- Credentials, keys, banking or payroll detail
- Documents you haven't got the right to share
- Anything where nobody will check the output
- Regulated judgement: clinical, legal, financial, safety sign-off
Rewrite the left column using your own document names before you hand this to anyone. A list that says "participant plans" gets followed; a list that says "personal information" gets nodded at.
The question people ask next
What about the six months before we had a rule?
Establish which account, roughly when, and what kind of document. Fix the setting. Write the rule. Move on. Treating it as misconduct guarantees you never hear about the next one, and unreported use is the thing you're trying to eliminate — not honesty about it. There's more on how to run that conversation in this article.
If you'd rather the awkward questions came from outside the business, that's what a policy and audit engagement does.
Start an enquiryThe questions we get asked about this
Does the paid tier change this list?
It changes the left column's edges slightly. It does not change the right column at all.
Who should own this list?
One person, named on the document, with the authority to answer the cases it doesn't cover. A list without an owner stops being current within a month.
See what this looks like for your business. Nine questions, about four minutes, and a plan ordered for your own work. No account, and the full plan is shown before any email is asked for. Build your plan.
Start an enquiry