Faster paperwork. Same damn queue.

Faster work can leave the queue untouched

An agency gives staff an AI assistant. Drafting gets faster. Meeting notes take less time. Users report that the tool is helpful. Yet an applicant still waits six weeks for an answer.

Those observations can all be true. A public service is a sequence of connected activities: receiving a request, checking evidence, resolving uncertainty, making a decision, and communicating it. Speeding up one task changes the service only if it changes the capacity or quality of the process that determines completion.

The question worth studying is when do AI-assisted task savings become shorter, more reliable service journeys? Our argument is that the unit of evaluation should be a correctly resolved case, including the work imposed on applicants and the effort of correction. Individual time savings are valuable evidence, but they do not answer that question by themselves.

Read the productivity evidence at the right level

The UK Government Digital Service's cross-government Copilot experiment, conducted from September to December 2024 and published in June 2025, reported an average saving of 26 minutes per user per day. It covered common office activities. That finding concerns staff work; it does not establish an equivalent reduction in the waiting time for a public service. ^[1]

HMRC's July 2026 evaluation provides another useful reference. Its Phase III report describes 3,000 randomly allocated licences, with a primary survey response from 1,364 users. Participants self-reported savings of around 2–3% of their working week. The report explains adjustments for non-use and survey-response differences, and describes the figures as a baseline estimate rather than a precise impact measure. It also reports a small task-based exercise. ^[2]

These results should not be used to claim that every hour of licensed access releases the same amount of capacity. The tool, staff role, task, adoption pattern, and measurement method matter. A person may spend time saved on better checking, more work, or relief from an overloaded schedule. Each can be valuable without automatically increasing completed cases.

The National Audit Office's 2024 review identified legacy systems, data access, and data quality as fundamental barriers to public-sector AI adoption. These constraints are relevant to service capacity because a faster draft cannot supply missing evidence or make incompatible records agree. The report is a dated institutional assessment, not a description of every agency's current systems. ^[3]

Follow the case through the bottleneck

Consider a fictional service with 100 incoming cases per day and an existing backlog. Intake can prepare 120 cases a day, evidence review can handle 80 review passes, and final processing can close 100 reviewed cases. Assume each case initially needs one review pass and ignore daily variation. Throughput is capped at 80 cases per day. The backlog grows by 20.

Now suppose AI doubles intake capacity to 240 cases per day, with no change in evidence review. The service still completes at most 80 cases per day. The extra intake capacity may improve another part of the experience, but it does not remove this bottleneck.

If the new intake process also causes 20% of cases to need one extra review pass, 80 review passes can support about 66.7 completed cases per day on average. This arithmetic assumes that the extra pass always resolves the issue and requires the same effort as the first. It does not model repeated returns or variation in case complexity.

These are deliberately invented numbers showing a mechanism. They are not measurements of a government service or a prediction that AI will increase errors. The calculation tells us what to observe: which step limits completion, whether rework changes, and where the saved capacity can actually be used.

Demand can change after the service improves

A more accessible service may attract people who previously gave up. Better translation, clearer eligibility information, or an easier application can increase arrivals. That may be a success even if the queue temporarily grows.

This makes waiting time alone an incomplete measure. A service could reduce its queue by making access harder. It could close cases quickly by asking applicants to start again. It could look productive by processing easy cases while complex ones age in the backlog.

We would track the population trying to use the service, including incomplete applications, repeat contacts, withdrawals, and unresolved cases. Results should distinguish ordinary cases from cases requiring extra support. A lower median wait is not enough if the longest waits worsen or the people most in need are disproportionately excluded.

Demand and capacity must be read together. If 100 cases arrive each day and only 80 leave, the queue cannot remain stable indefinitely. Making the first step faster does not alter that balance. It may make the mismatch more visible, which is useful if the agency is prepared to act on it.

A decision record is part of the service

An AI-generated summary can omit an exception, flatten conflicting evidence, or make an incomplete record appear settled. A reviewer needs access to the original evidence, the unresolved questions, and the relevant rules. A confident paragraph is not an adequate decision record.

Our proposed starting point would be assistance with evidence organisation and administrative preparation under an authorised process. The pilot would preserve the agency's decision authority, existing review requirements, and routes for challenge. Introducing automated decisions about individual entitlements would be a separate project with different requirements.

The record should show which material was supplied, what the tool produced, what a person changed, and where uncertainty remained. This can make review more effective and help diagnose recurring problems. It also creates work, storage, and governance costs that belong in the evaluation.

The objective is not to maximise the number of cases touching an AI system. It is to improve resolution while keeping the reasoning inspectable and the service accountable to the people using it.

FIG. 03 / THE HUMAN CONSTRAINT

The person waiting
is part of the system.

Test at the level where the effect happens

We would first map one service using dated event records: receipt, evidence request, response, review, decision, correction, and closure. Distinguish active work from time waiting for the applicant, another organisation, or internal capacity. Missing timestamps and inconsistent status definitions are findings to resolve before making strong productivity claims.

The first evaluation would run on historical or synthetic cases in a controlled environment, with access and data handling agreed by the agency. Independent reviewers would check whether the proposed assistance preserves material evidence, handles exceptions, and exposes uncertainty. A failed output would be investigated before live use.

For a live pilot, the study design would need to account for shared queues. Randomising individual staff can contaminate a comparison if everyone draws from the same backlog or shares completed summaries. Depending on the service, a planned rollout by team or period may be more appropriate. The design should document changes in staffing, demand, case mix, and policy alongside the intervention.

The primary outcome would be time to a correctly resolved case, with an agreed quality review. Secondary outcomes would include rework, unresolved-case age, repeat contacts, staff effort, applicant effort, and cost. Quality criteria and stopping rules should be settled before the trial begins. They should not be relaxed to make a faster system look successful.

When the answer should be a different investment

If most delay occurs while waiting for an external record, connecting that record may matter more than improving drafting. If review is the bottleneck, saved intake time is useful only if staff can be redeployed or the review burden can be reduced without losing necessary scrutiny. If rework drives the queue, better evidence capture may outperform a faster assistant.

AI may still improve staff experience or the quality of written explanations. Those gains should be reported in their own terms. Turning them into a claim about faster public services requires evidence that the whole journey changed.

For JDGMNT, the opportunity is to investigate how intelligence fits into the operating system of a service: its capacity, records, incentives, and responsibilities. The strongest project may combine a modest AI capability with a change to how work moves between teams. The result worth pursuing is a service people can complete, with fewer unresolved cases and a clear account of how decisions were reached.