The Department for Education plans to use generative artificial intelligence to determine which apprenticeship vacancies require human quality checks, replacing universal manual review with a risk-based system expected to process between 5,000 and 6,000 drafts each month.

The proposed application programming interface will inspect vacancies before they appear on the Recruit an Apprentice service, checking for potentially discriminatory requirements, missing or inconsistent information, geographic problems, duplicates, and spelling or grammar errors. Its result will determine whether a posting is sent to a reviewer or allowed to proceed automatically.

Details published through the government’s Algorithmic Transparency Recording Standard show that the system combines GPT-4o, conventional rules, statistical sampling, and a traffic-light risk classification. The department is also investigating GPT-4.1 and o4-mini through Microsoft’s Azure OpenAI service.

A red classification will trigger human review in every case, while 50% of amber vacancies and a randomly selected 1% of green vacancies will also be checked. Postings not referred under those rules will be approved without further intervention, giving the system a direct role in deciding which public-facing vacancies avoid manual scrutiny.

Manual review currently covers every vacancy

Approximately 40,000 postings are uploaded to Find an Apprenticeship each year, according to the department, and every one currently passes through a manual review that can take up to 24 hours. The Apprenticeship Service pays for each review even where a vacancy closely resembles an earlier posting or contains only minor mistakes.

The replacement system will run 11 checks against each submitted draft. Nine concern spelling across individual text fields, while the other two examine possible discrimination and missing or inconsistent content across the vacancy as a whole.

Rule-based components will handle known requirements, formatting, geography, and some missing information, while the large language model will interpret less structured language and context. Reviewer comments will also be converted into numerical representations and clustered to identify recurring themes that could inform later changes to the system.

The Department for Education estimates that the approach could reduce manual-review costs by around 40%, partly by avoiding repeated checks on cloned vacancies. Faster clearance would also shorten the delay between an employer submitting an opportunity and applicants being able to see it.

Those savings depend on the distribution of red, amber, and green results, as well as the accuracy with which the system assigns them. A conservative model could refer too many vacancies and preserve much of the existing workload, while an overly permissive one could allow unsuitable content to pass without examination.

Sampling becomes the main safeguard

The random review of low-risk postings is intended to detect problems the model has missed. Because 1% of green vacancies will still reach a human reviewer, the service can estimate whether apparently safe postings contain significant errors and adjust prompts, thresholds, or rules where patterns emerge.

A 1% sample will provide limited protection against rare but serious failures, particularly where harmful wording appears in forms that were poorly represented during development. The effectiveness of the safeguard will depend on how quickly missed cases are identified, investigated, and translated into changes across the wider system.

Amber vacancies receive substantially more scrutiny, with half sampled for review, although the published record describes amber mainly as a category for lower-risk issues such as minor grammar problems. The boundary between amber and red will therefore affect cost, publication speed, and the likelihood that ambiguous discriminatory language receives human attention.

Human reviewers will retain authority over vacancies routed to them, with the ability to approve a posting, edit it, or return it to the submitter. Postings cleared automatically will not receive the same safeguard, making model monitoring and false-negative measurement central to the service rather than supplementary technical work.

The transparency record says model reviews will take place when new systems become available, performance indicators change, or an underlying API approaches deprecation. That flexibility may improve performance, but replacing one model with another can alter classifications even where prompts and business rules remain unchanged.

Third-party delivery remains inside the department

Atos, Version 1, and Talent Consulting have contributed model development, technical architecture, software delivery, governance, and business analysis. Contractors will work through department-managed devices and cloud services, with no external sharing of operational data permitted.

Keeping access inside the departmental estate reduces some data-handling risks, although supplier dependence remains relevant because external companies have helped construct both the model workflow and its surrounding governance. The department will need sufficient internal capability to test changes, challenge performance claims, and maintain the service if contracts or technology suppliers change.

The published architecture also relies on Microsoft’s Azure OpenAI service, leaving the application exposed to model-retirement schedules, pricing changes, and modifications in the behaviour of externally supplied foundation models. The department acknowledges API deprecation as one reason that future model reviews may be necessary.

Although the system concerns vacancy quality rather than candidate selection, its checks can still affect employers and prospective apprentices. A missed discriminatory requirement could discourage or exclude applicants, while an incorrect high-risk classification could delay a legitimate vacancy and create more work for the employer.

The government’s decision to publish model, procurement, sampling, and workflow information provides a level of operational detail often missing from public-sector AI announcements. It also exposes the remaining evidence gaps: no production performance figures, false-negative rate, cost baseline, or launch date have yet been disclosed.

Replacing universal manual review with selective intervention is a measurable operational change rather than a demonstration project. Once the system enters service, its credibility will depend on whether the department publishes evidence showing how many vacancies were cleared automatically, how often sampled reviewers found missed problems, and whether the promised savings were achieved without lowering the quality of apprenticeship advertising.