Safe-Use Patterns for AI Assistants
Simple patterns for using Copilot, ChatGPT, Gemini and Claude safely in a social housing context — for staff, managers and IT teams.
A good safe-use pattern works like a marked route through a busy station. It does not explain the whole railway. It tells people where they can go, where they need permission, and where they should not step.
Safe-use patterns
These uses are generally safe with normal care and attention.
| Pattern | Safe example | Conditions |
|---|---|---|
| Draft from public or non-sensitive material | Ask AI to draft a first version of an internal note about a published Regulator of Social Housing update. | Check accuracy, links and tone before sharing. |
| Summarise non-personal documents | Summarise a public consultation, policy paper or internal guidance that contains no tenant or staff personal data. | Keep the original source. Do not rely on the summary alone. |
| Improve plain English | Rewrite a service update so it is clearer for residents. | A human must check meaning, accessibility and tone. |
| Plan a workshop or meeting | Ask for an agenda, discussion questions or facilitation plan. | Remove names, live cases and confidential details. |
| Analyse anonymised themes | Ask AI to group anonymised complaint themes or survey comments. | Anonymisation must be checked before upload. Validate themes manually. |
| Build a first checklist | Ask for a draft checklist for reviewing a process. | Treat it as a starting point. Service owners must adapt it. |
| Create prompts for low-risk tasks | Ask AI to improve a prompt for summarising public documents. | Do not include sensitive business context or personal data. |
Use with extra care
These uses may be helpful but need approval, review or a DPIA screening first.
| Use | Why it needs care | Minimum control |
|---|---|---|
| Drafting complaint responses | May affect tenant trust, accuracy and fairness. | Human review by someone who understands the case. |
| Summarising tenant records | May expose personal or sensitive data. | Approved tool, data protection review and access controls. |
| Analysing arrears, vulnerability or tenancy data | May affect people in high-impact situations. | Governance review, fairness check and clear human decision-making. |
| Using AI with supplier systems | Data may move through third parties. | Supplier assurance and contract review. |
| Creating official records | AI may omit, invent or distort facts. | Human verification before saving to the record. |
| Allocating AI licenses and API keys | Software license fees (e.g., Copilot) scale rapidly, and unmonitored API usage can cause unexpected billing spikes. | Establish business-case sign-off for licenses, set strict token usage caps and budget alerts on API consoles, and track ROI. |
Do not do this
- Do not paste tenant names, addresses, case notes, arrears history, complaint files, safeguarding or medical information into public AI tools.
- Do not ask AI to make final decisions about repairs priority, allocations, complaints, vulnerability, benefits or access to services.
- Do not upload confidential contracts, board papers, staff records or supplier data unless the tool is approved for that data.
- Do not copy AI text into resident communication without checking facts, tone and policy position.
- Do not let AI create or update official records without human verification.
- Do not use AI to bypass access controls, procurement rules, data protection review or professional judgement.
A simple prompt pattern for low-risk work
For higher-risk work, add: "Before answering, list what information would be unsafe to include and what a human should verify."
Staff rule of thumb
If the information would worry you on a train seat, in a public email, or in a supplier demo, do not put it into an unapproved AI assistant.
If the output could affect a resident, check it like a decision.
If the task involves personal data, sensitive records, service access or a vulnerable resident, pause and ask for the approved route.