Artificial intelligence

The starting point

An artificial intelligence initiative begins with a specific decision: which task needs support and who remains accountable for its outcome. We examine the full path of that task, from the arrival of a document or enquiry to the answer a person receives. That is where useful opportunities emerge: routing requests, finding prior records, identifying exceptions and preparing evidence so a team can act sooner. A useful assistant must recognise when it lacks context and hand the case over; a sensitive decision is never delegated merely because a model can produce a plausible sentence.

Source material deserves as much care as model selection. Before connecting semantic search, we organise permissions, provenance, validity dates and document formats. An outdated policy, an undated instruction or a duplicate file can undermine any answer. For each use case we define test sets covering routine questions, ambiguous situations and cases where the system should abstain. Evaluations compare accuracy, coverage, citations and response time before a new source is included or an instruction is changed.

The interface communicates uncertainty and keeps a route back to the source. When a staff member corrects an answer, the reason is recorded and used to review the workflow; every correction does not automatically become training data. Operational logs avoid retaining unnecessary sensitive information, while conversations containing personal data follow the organisation’s retention and access rules. A pilot begins with a small group and repeatable tasks, overseen by people who can pause it when the results are no longer reliable.

Rollout proceeds through clear thresholds: internal assistance first, supervised operations next, and limited automation only when evidence supports it. A public enquiry channel also needs review of tone, accessibility, language and handover time. In healthcare, education and financial services, human review belongs in the design from the beginning. This turns a prototype into a tool that can be audited, maintained and improved while keeping the people affected by its output in view.

Tools and method

• Document inventory with version control and an owner for each source.
• Text extraction and hybrid search to retrieve verifiable passages.
• Vector indexes and permission filters applied before context reaches a model.
• Reference questions, abstention tests and human review for evaluation.
• An API with rate limits, request traces and a change log.
• Separate dashboards for latency, cost and quality without exposing personal data.

Use cases

• An internal policy assistant citing the current approved version.
• Incoming enquiry classification and routing to the right team.
• Supervised field extraction from forms and documents.
• Detection of unsupported answers and handover to a person.

Frequently asked questions