Define the job before choosing the tool
There is a substantial difference between asking for a generic meeting agenda and uploading a named employee’s grievance to generate a response. Define the task and decide whether personal information is necessary at all. Use invented material during early testing wherever it will answer the same question.
This gives a clearer procurement brief: the information needed, the intended output, who will use it and what decisions may follow. It also avoids a common pattern in which a convenient tool becomes the destination for progressively more sensitive records without anyone noticing that the original trial has changed into an operational system.
Minimise the information at source
The ICO’s AI guidance emphasises data minimisation and security throughout the system lifecycle. Removing a name may not anonymise a record if the remaining details still identify the person. Dates, role descriptions, unusual incidents and combinations of facts can make someone recognisable.
A useful practical rule is to ask whether a shorter extract or a fictionalised example would meet the need. Keep the original record in its approved system rather than creating unnecessary copies across tools. Do not assume that a prompt disappears when the chat window closes. Understand retention, logs and deletion controls before supplying information that would matter if disclosed.
Check the actual product and account terms
The supplier’s role may depend on how it uses inputs and outputs. Review the relevant agreement, security measures, training settings, retention options, sub-processors and international-transfer arrangements. A consumer account and an enterprise product from the same provider may not offer the same controls.
Our recommended approval record identifies the exact product, permitted uses and information types, who can access it and which settings must remain enabled. A blanket statement that an organisation uses only compliant AI is not an adequate substitute. Reassess when a tool adds connectors, memory, automated actions or wider access to business files.
Decide what people need to know
Personal information remains subject to data protection requirements when AI is involved. The business must identify an appropriate lawful basis and provide relevant transparency. Sensitive information may introduce additional conditions. Privacy wording should describe the actual use rather than hide it behind a vague reference to technology suppliers.
Screen for a DPIA where the proposed processing could create high risks. A tool summarising public product descriptions is different from one assessing staff performance or eligibility for a service. The ICO’s AI guidance is under review following legislative changes, so significant automated decision-making needs a current, specific assessment rather than reliance on an old checklist.
Human review must change the outcome when necessary
AI output can be inaccurate, incomplete or unfair. A person reviewing it needs enough context, competence and authority to reject it. Merely clicking approval because the result looks polished is not meaningful scrutiny. The ICO’s governance guidance highlights the importance of genuine intervention where human review is part of the process.
For ordinary drafting, a practical control is to verify names, dates, quotations and assertions against the original evidence before using the text. For decisions affecting people, define an escalation route and keep an appropriate record of the reasoning. Do not allow a draft-writing experiment to become an unexamined decision engine.
Create a usable staff policy, not a prohibition nobody follows
A short policy can distinguish approved tools, permitted information and uses requiring further review. Give staff a route to ask about a new task. Explain what to do if information is uploaded by mistake, rather than encouraging people to hide it.
Start with a bounded use case and test both quality and information handling. Our suggested review asks whether the tool genuinely saves time after checking, whether users follow the agreed limits and whether the outputs remain fit for purpose. Responsible adoption is not the absence of AI. It is being able to explain and control what the business is doing with people’s information.
Sources and further reading
Primary guidance behind this briefing. Requirements depend on the facts; check the current guidance before making a decision.
This is general guidance, not a legal opinion on your circumstances or a guarantee of compliance. An individual assessment may be needed, particularly where sensitive information, disputed rights or regulatory action is involved.