The Office of the Public Guardian is piloting an AI system that categorises bank transactions during investigations into possible financial misuse, offering an unusually detailed view of how generative AI is beginning to enter sensitive UK public-sector casework.

The Investigations AI Assistant processes financial transaction data supplied during cases involving attorneys or deputies responsible for the affairs of people who may lack capacity to manage their own finances. Investigators previously categorised transactions manually in spreadsheets, identifying items such as care fees, utility bills, pension payments, and cash withdrawals before deciding which activity warranted closer examination.

Initial testing found that the time required to categorise transactions almost halved, according to an Algorithmic Transparency Recording Standard entry published on 7 August. The result is now being tested through a controlled live pilot, with the system intended to produce a first-pass categorisation rather than make investigative decisions.

Every output is reviewed by an investigator, who can correct classifications and examine transactions flagged as uncertain before anything progresses or is shared. The OPG says no decision is taken automatically, while normal supervisory, complaints, and Court of Protection processes remain in place around the underlying investigation.

A hybrid system rather than a chatbot

The technical design is more revealing than the familiar description of a government body using AI because the system combines several models with conventional workflow software rather than relying on one general-purpose model to produce an answer. GPT-4o, running through Azure AI Foundry, performs classification work, while OpenAI’s text-embedding-3-large model generates embeddings used inside the processing pipeline.

A separately trained LightGBM model then estimates whether an upstream transaction classification is likely to be correct. That confidence model runs locally inside an Azure Function App without external API calls, using a threshold of 0.91 to determine which transactions should be flagged for further attention.

On the OPG’s higher-quality validation data, around 23% of transactions were flagged for review, while approximately 98.6% of unflagged transactions were correctly classified. The confidence model was trained and evaluated on about 54,000 labelled outcomes assembled from anonymised historic OPG transactions, synthetic data, publicly available examples, and additional AI-generated material.

Those details make the project useful beyond the immediate investigation workflow because they show one way public-sector organisations are attempting to constrain generative models inside more conventional assurance mechanisms. The language model performs the flexible interpretation, a smaller machine-learning model estimates confidence, and human investigators retain responsibility for the final judgement.

Automation stops before the decision

The distinction between assisting analysis and deciding a case is particularly important because the underlying data can include sensitive personal information, bank transactions, account activity, and evidence relating to suspected financial abuse. Errors in categorisation could draw attention towards innocuous spending or make genuinely concerning activity less visible, so the design assumes that model output remains provisional.

Investigators are required to review all AI-produced classifications rather than accepting the system’s spreadsheet as completed analysis. Transactions that the confidence model considers uncertain receive an additional flag, but even those passing the threshold remain subject to investigator review before the case proceeds.

Operational processing is confined to Microsoft’s UK Azure region, with encrypted storage and network connections, role-based access controls, and identity-based authentication. Microsoft helped configure the Azure components and AI workflows under a framework agreement call-off, although the OPG says Microsoft has no access to operational OPG data.

The public record also exposes limitations that procurement announcements often omit. No third-party testing has been conducted, evaluation has so far been carried out internally by the OPG and the Ministry of Justice’s JusticeAI Unit, and the live pilot is intended partly to establish whether time savings and model performance persist under routine operating conditions.

Public-sector AI moves into workflow redesign

The OPG considered increasing staffing and overtime, improving templates and case management systems, using deterministic rules, and deploying robotic process automation before choosing an AI-assisted approach. Each alternative could remove part of the administrative burden, but the agency concluded that none handled the variability of financial descriptions while preserving the same scope for automated categorisation.

That comparison places the project beyond the use of general-purpose assistants for drafting or search. Here the technology is being inserted into a defined operational process with identifiable inputs, measurable outputs, sensitive data, named models, human review requirements, and a specific claim about time saved.

Whether the pilot ultimately delivers a productivity gain will depend on more than classification accuracy because investigators must still spend time checking outputs, correcting errors, responding to unusual cases, and maintaining sufficient scepticism to avoid automation bias. A system that accelerates straightforward transactions but encourages excessive confidence in ambiguous ones could move effort further down the investigation rather than remove it.

The transparency record nevertheless provides something still uncommon in public-sector AI deployment: enough technical and operational detail to examine what the system actually does. As government departments push AI deeper into administrative work, the more useful evidence will come from error rates, staff time, oversight costs, and case outcomes rather than the number of pilots announced.