UK Department for Transport Deploys Google Cloud AI for Consultation Analysis

TL;DR: The DfT is using a Gemini-based tool built with Google Cloud and the Alan Turing Institute to analyse over 100,000 free-text responses per consultation, reducing work that previously took months to a few hours and saving up to £4 million annually.

The UK’s Department for Transport runs roughly 55 public consultations every year, each capable of generating more than 100,000 free-text submissions. Classifying those responses manually has historically stretched across several months, making the department’s 12-week publication commitment difficult to meet.

Context and Background

The Consultation Analysis Tool (CAT) was built on Google’s Vertex AI platform using Gemini models, with input from the Alan Turing Institute. DfT reports the system reaches up to 90% accuracy across its evaluation metrics and has already been used to process responses on the Integrated National Transport Strategy and revisions to driving test booking rules.

Beyond consultations, DfT is running two further AI deployments: a Connectivity Tool for urban planners (built on Cloud Run, Cloud CDN and Firestore) and an AI Correspondence Drafter that uses Vertex AI Search to retrieve policy context before generating first-draft responses to public inquiries. A “human-in-the-loop” check applies across all three tools, with policy experts validating outputs for accuracy and bias.

For UK public-sector AI procurement, the DfT case matters for two reasons. First, it is one of the few deployments to publish both a quantified efficiency saving (£4 million annually) and a baseline accuracy figure, giving other departments a reference point when scoping similar projects. Second, the three tools together suggest a pattern: pair retrieval-augmented generation with Vertex AI for policy-grounded drafting, reserve foundation models for bulk classification, and retain human sign-off on decisions.

Looking Forward

DfT’s approach contrasts with earlier UK public-sector AI pilots that focused on chatbot-style citizen interfaces. Consultation analysis is a narrower, more defensible use case — the outputs feed into existing policy workflows rather than directly engaging the public — and that framing is likely to shape how other departments pitch AI projects under the government’s ongoing AI adoption push.