Experiment
AI Decision Copilot
Make specialist knowledge and the most relevant evidence available at the moment a human has to exercise judgement.
01
Problem to address
Frontline and specialist staff spend significant effort finding context before they can apply professional judgement. The risk is not only delay; relevant evidence can be missed or inconsistently interpreted.
02
What the evidence tells us
Field officers assemble dossiers and work around incomplete data. Regulators search across systems and unofficial tools. Analysts spend time unblocking others. People also need help directing attention to the most relevant claims first.
03
The proposition to explore
The idea
Give staff an evidence-grounded copilot that surfaces the small number of facts, anomalies and gaps that matter for the decision in front of them.
04
Hypothesis
If a copilot can surface the small number of evidence items that matter for the current decision, with provenance and uncertainty, staff can spend more time on judgement and less on information assembly.
05
How it could work
The user opens a case or visit. The copilot uses the connected evidence thread to produce a briefing, highlights anomalies or missing evidence, and answers follow-up questions with citations. It does not make the regulatory decision.
06
Example: bring it to life
Field officer
Before visiting:
“What should I pay particular attention to?”
Copilot:
Parcel 42 has conflicting hedge classifications.
The SSSI intersection has low confidence.
Previous imagery is 31 months old.
Natural England visited this holding 14 months ago.
Here is the relevant evidence.
Regulator
“What is unusual about this producer?”
Copilot:
Registration says Small.
Companies House information and submitted tonnage suggest this warrants review.
Packaging tonnage is 34% below last year.
Two submissions conflict.
These four checks require human judgement.
07
AI opportunity
Retrieval, synthesis, anomaly explanation, prioritisation and conversational interrogation. The copilot should expose source, freshness and uncertainty, and keep the human as decision-maker.
08
Potential value
- Less search time
- More consistent use of evidence
- Earlier identification of important anomalies
- Specialist knowledge available closer to frontline work
09
Reduce uncertainty without overcommitting
Smallest meaningful experiment
Create a copilot over a closed, curated evidence pack for one workflow. Compare preparation time, missed issues and user trust against the current process.
10
Open questions and risks
What we still need to learn
- Bad foundations create confidently wrong summaries
- Need strict provenance and access controls
- Automation bias could shift judgement toward the AI