Experimenting with Generative AI to Support Delivery Officers at DfE

Key Takeaways

• Copilot Studio enabled rapid prototype of Teams‑based AI assistant
• Agents saved experts time by surfacing relevant sections of long guidance
• AI struggled with queries requiring local context or judgment
• Multiple voice styles improved answer clarity for generalist officers
• Human‑in‑the‑loop remained essential for verification and decision‑making

Pulse Analysis

The UK government’s recent "test‑and‑learn" AI strategy encourages ministries to pilot generative tools before large‑scale adoption. In this spirit, the Department for Education (DfE) collaborated with digital consultancy dxw to explore whether an AI assistant could streamline the flow of policy guidance to regional delivery officers. By embedding the agent in Microsoft Teams—where civil servants already collaborate—the experiment leveraged existing workflows, reducing friction and providing a realistic gauge of productivity gains. Early results suggest that well‑targeted AI can free up specialist time, allowing them to focus on higher‑order policy interpretation rather than repetitive document hunting.

Technical choices shaped both the speed and limits of the trial. Using Copilot Studio’s web connector forced the team to manually curate source documents, as the tool cannot ingest URLs with query parameters or deep hierarchies. Nevertheless, the platform’s rapid prototyping capabilities allowed dxw to iterate quickly, employing the CRISPE framework to define the agent’s capacity, insight, statement, personality, and guardrails. This structured prompt engineering ensured the assistant referenced trusted DfE materials and maintained a consistent tone, while also revealing that a single authoritative voice can hinder comprehension for non‑expert users. Adjusting the agent’s personality and introducing multi‑voice responses proved pivotal for user acceptance.

The broader implication for public‑sector digital transformation is clear: generative AI is most effective when it augments, not replaces, human expertise. Agents excel at navigating voluminous, static guidance, yet they stumble on questions that demand local nuance or judgment. Consequently, a hybrid model—where AI surfaces relevant excerpts and humans validate and contextualize—offers the optimal balance of efficiency and accountability. As the DfE plans to scale successful patterns, other ministries can adopt a similar staged approach, prioritizing use cases with extensive documentation and confident user bases while embedding rigorous oversight mechanisms. This measured rollout could accelerate government service delivery without compromising the essential human element.