AI tools for social work should not begin with the question, what can the model do? They should begin with a more responsible question: what could go wrong if this output is misunderstood, over-trusted or used without context?

That is not anti-innovation. It is the practical starting point for any tool that touches safeguarding, family life, professional records or vulnerable people. The goal is not to slow good technology down; it is to make sure useful tools can be trusted.

The Minimum Safety Bar

• The user must know when content is AI-assisted.
• Important outputs must remain editable and reviewable.
• The system must preserve audit trails.
• Sensitive data should stay local where possible.
• High-risk claims should be checked or escalated.
• No generated text should silently become a final professional record.

Safeguarding Is A Workflow Issue

Safeguarding is not only about whether a model gives a safe answer. It is about whether the workflow makes uncertainty visible, keeps the practitioner accountable and prevents a draft from being mistaken for a decision.

A good system should help a practitioner notice missing information, separate observation from analysis, and prepare for supervision. It should support professional curiosity. It should not compress complex family circumstances into a confident but thin summary.

Privacy By Design

Social care data can include special category data, information about children, family circumstances and third parties. For that reason, local-first processing should be considered before cloud processing, especially for drafts, anonymisation, transcription and document review. This is a practical privacy choice, not a slogan.

The right AI tool makes professional accountability easier to evidence. The wrong one makes accountability harder to trace.

The AI Social Worker is being built around that distinction: support the work, reduce the admin burden and keep safeguarding at the centre. That is the standard practitioners, families and organisations deserve.