Review of AI-generated communications

What it does

It applies to machine-written communications the standard already applied to human ones. Content is checked for accuracy, clarity, balance, required disclosures and tone before it reaches a customer. High-risk content, such as promotions and anything about a customer's rights or money, is reviewed in full or generated only from approved text; lower-risk content is sampled. The owner of the communication is accountable; compliance sets and tests the standard. It operates before release of any template or prompt and continuously on output.

How to implement it

• Classify communications by risk, and set the level of review for each class.
• Fix mandatory wording outside the model, so it cannot be altered or dropped.
• Restrict generation for high-risk content to approved sources and templates.
• Check outputs automatically for prohibited claims, missing disclosures and pressure.
• Sample outputs for human review, including personalised variants.
• Approve prompts and templates as you would the communications themselves.

Evidence it operates

Design: content standards and a review matrix by risk class. Operation: approval records for prompts and templates; automated check logs; sampling results with error rates; communications withdrawn or corrected.

Where it fits

Between generation and the customer. It is the primary control for misleading AI-generated communications, and supports unsuitable AI-generated advice or guidance and manipulative AI-driven design. It works with factuality and hallucination verification, which checks facts.

Limits

Real-time conversation cannot be reviewed before it is sent, so for chat the control falls back on constraints and sampling after the event. Sampling misses rare errors. Personalised output means no reviewer sees what a given customer saw. Reviewers of fluent text tend to pass it.