Start with the morning briefing
Imagine asking for yesterday's corridor performance briefing. The useful output is a short account of what changed, with the figures and source records needed to inspect each finding. Producing that account involves several different kinds of work: checking data coverage, choosing a comparable baseline, calculating measures and drafting an explanation.
This is a practical place to explore an AI co-worker. The role has a recognisable trigger, repeatable steps and a deliverable that an engineer can assess. It also has a clear stopping point: a draft awaiting review.
Give each part of the workflow a clear role
In this proposed workflow, analytical code calculates the performance measures. A retrieval step finds relevant approved documents. A language model helps assemble the explanation. An agent coordinates these permitted steps, keeps track of their outputs and surfaces anything it could not resolve.
That separation helps the reviewer inspect a claim. A change in travel time should point to the calculation and period used. An explanation involving planned works should point to the relevant source. Missing information should remain visible in the draft.
Design the review into the work
The review screen should distinguish observations from possible explanations. It should show missing-data flags, the comparison period and any changes the co-worker made to the draft. The responsible person needs enough context to accept, revise or reject the work.
NIST's AI Risk Management Framework provides a useful reference for considering trustworthiness across the design, development, use and evaluation of AI systems. For a transport pilot, we would translate that into specific permissions, evidence requirements and tests for the chosen task.
Measure the pilot against the job
A sensible pilot might examine calculation consistency, source coverage, missed data issues, the amount of editing required and the time taken to produce an acceptable briefing. Test routine days alongside disrupted and incomplete-data days. Agree the acceptance criteria before judging the results. Include cases where the correct response is to pause a comparison because the available evidence is insufficient.
RoadInsite's proposed co-worker roles begin with analysis, knowledge retrieval and draft preparation. The opportunity is to organise those tasks into a useful working process, with the team's judgement and responsibilities built into it.