When an AI-assisted report contains a wrong or exposed client detail, pause further distribution, contain access, preserve the necessary evidence and appoint an incident owner. Verify the facts against authorised records, correct the recipient’s copy and assess any privacy or safeguarding implications. Treat notification as a decision with legal conditions and deadlines, rather than assuming every AI mistake is a reportable data breach.
What could go wrong in an ordinary nonprofit report?
This is a hypothetical scenario. A social worker uses an AI assistant to shorten case notes for a donor report. The notes include a client’s name and personal circumstances. The assistant inserts a plausible but false appointment date, and the draft is emailed before anyone checks it against the source record.
Two problems now need attention: the donor received an inaccurate statement, and personal details may have reached people or systems that were not authorised to receive them. Fixing the invented date alone does not settle the second problem.
The team should establish which information was uploaded, what the tool retained, who received the report and whether an individual could be identified. A name is not the only identifier. A distinctive combination of location, age and circumstances may identify someone too.
Which kinds of AI error require different responses?
| Error | Example | First response |
|---|---|---|
| Invented fact | A draft supplies an appointment date absent from the record. | Verify against the authorised source and correct the claim wherever distributed. |
| Mixed client cases | Details from two people appear in one case summary. | Separate the records, identify affected recipients and assess confidentiality consequences. |
| Oversharing | A donor report includes personal details it did not need. | Restrict further access and ask the privacy or safeguarding lead to assess the disclosure. |
| Misjudged urgency | An AI summary makes a serious concern sound routine. | Return to the original record and escalate through the established safeguarding process. |
These categories can overlap. An inaccurate report can also create a confidentiality problem. Conversely, an invented figure with no personal information may need an urgent correction without being a personal-data breach.
What should the team do once the report has left?
Use this five-step protocol as an operational starting point. Apply your existing incident procedure and involve the appropriate privacy, safeguarding or legal lead when the facts require it.
| Step | Action | Who owns it |
|---|---|---|
| 1. Pause and contain | Stop further sends and the affected workflow. Restrict shared-document access where possible. Ask recipients to pause use of the report. | Workflow owner with the incident lead. |
| 2. Record and scope | Log discovery time, tool, inputs, output, recipients and known access. Preserve only necessary evidence in approved restricted storage. | Incident lead with IT or the tool administrator. |
| 3. Verify the facts | Compare the output with authorised original records. Identify incorrect claims, affected people and possible safeguarding consequences. | Responsible programme professional and relevant specialist. |
| 4. Correct and assess notification | Send a clear correction through an authorised channel. Assess privacy obligations and notification deadlines separately. | Incident lead, privacy lead and accountable manager. |
| 5. Recover and prevent recurrence | Confirm recipients have the corrected version, repair the workflow and test its review step before restarting. | Workflow owner and the manager authorising restart. |
Do not copy sensitive material into another unapproved chatbot to investigate it. Do not circulate screenshots more widely than necessary. Preserve the evidence needed for the incident record while reducing further exposure.
The correction should say which version or claim is affected and what replaces it. It should not repeat confidential details to recipients who never needed them. Record unanswered questions as unanswered; an AI-generated explanation is not evidence of what happened.
When do GDPR duties become relevant?
First establish whether the GDPR applies to the organisation and processing. Under Article 4(2), processing includes collecting, storing, using and disclosing personal data. Uploading identifiable client notes can therefore be processing, even when the task is “just summarisation”.
Article 33 requires controller notification to the competent authority without undue delay, where feasible within 72 hours of awareness, unless risk to rights and freedoms is unlikely. Explain delays, document the breach and establish the controller/processor roles; processors notify controllers without undue delay.
Article 34 separately requires communication to affected people without undue delay when high risk is likely, subject to stated exceptions. Ask the privacy lead to assess applicability and deadlines promptly, alongside the donor-report correction.
Which five safeguards should be in place before the next report?
- [ ] Approved tools and data: document permitted uses and data categories using Gate 3.
- [ ] Protected identities: keep client identifiers and sensitive case details out of unapproved tools; use approved minimised inputs for permitted tasks.
- [ ] Named human approval: assign someone to check factual claims and confidentiality before external distribution. Apply the Signature Test.
- [ ] Recurring practice: train staff using fictional examples and repeat the exercise when the workflow or tool changes.
- [ ] Clear reporting route: make it easy to report a mistake quickly, with an incident owner and backup who know where the record belongs.
Put those rules in the one-page AI policy, and keep the detailed incident procedure where staff can find it. A rule that names nobody cannot reliably trigger a response.
My view
The useful question after an AI error is what happened and what now needs protecting. Starting with blame can discourage the next person from reporting early. That does not remove accountability: managers still need to assign the correction, notification assessment and restart decision.
Trust can take years to build and be damaged quickly. A specific, timely correction and a documented response give recipients something firmer than a general reassurance that the AI made a mistake.
What to do next
Before an incident, rehearse the hypothetical donor-report case with the programme lead, tool owner and privacy or safeguarding contact. Ask who can stop the send, who checks the original record and who decides on notification. Write the answers into the existing procedure.
During a live incident, use that procedure immediately. Record the time of awareness and involve the appropriate lead promptly. Restart only once the affected work has been corrected and the revised approval step has been tested.
FAQ
Is every AI mistake a personal-data breach?
No. An invented non-personal statistic can require correction without being a personal-data breach. If personal information was exposed, altered, lost or accessed improperly, the incident needs a separate privacy assessment.
Does entering client notes into an AI tool count as processing?
Uploading identifiable notes can involve storage, use or disclosure of personal data. Whether the GDPR applies and whether the processing is permitted depends on the organisation, purpose, tool and safeguards.
Should we tell the donor and delete the evidence immediately?
Correct the donor’s affected copy through the incident owner, while preserving necessary evidence in restricted approved storage. Avoid further disclosure. Retention and deletion decisions should follow the applicable incident procedure and privacy requirements.
Can we restart as soon as the wrong date is fixed?
Check the wider scope first: inputs, recipients, confidentiality and required notifications. The responsible manager should authorise restart after the correction and the revised review process have been verified.
Sources
- European Union: GDPR, Articles 4, 33 and 34
- NIST: Artificial Intelligence Risk Management Framework, Generative AI Profile (AI 600-1)
About the author
Milan Bogojevic works on practical AI adoption in nonprofits, with a focus on workflows, human accountability and organisational capability.
Last updated: 10 October 2026.