The defence AI workforce problem is ownership as well as skills

Public US audits connect AI adoption with role definitions, software capacity and the people who maintain evidence after a pilot.

An AI pilot can succeed because a small team understands every detail. Scaling it exposes a different problem: who will own the data, review changes, support users and investigate failures once those specialists move on? Public-sector workforce planning needs to include those responsibilities. Hiring a model developer alone does not create a durable service.

The US Government Accountability Office’s [defence AI workforce audit](https://www.gao.gov/products/gao-24-105645), published in 2023, examined management of the emerging AI workforce. Its recommendation entries include later developments and should retain their dates. A [software-modernisation audit](https://www.gao.gov/products/gao-23-105611) discusses related planning gaps. The department’s [November 2023 AI-adoption announcement](https://www.defense.gov/News/News-Stories/Article/Article/3578219/dod-releases-ai-adoption-strategy/) sets strategic context. These sources establish public management concerns and policy direction, not the present staffing condition of every defence organisation.

A safe original illustration is a training-administration office using AI to sort course requests. A data specialist prepares the permitted input. An application engineer maintains the service. A training officer checks ambiguous classifications. A support owner handles errors and access questions. If those duties are assigned informally to the same enthusiastic person, the pilot may look cheaper than the service really is.

The skills question should follow the task. Reviewers need to recognise unsupported conclusions, stale records and missing context. Engineers need to know which changes require a new evaluation. Procurement staff need to ask for evidence and usable data rights. Managers need to understand why a high average score can still conceal harmful exceptions. The necessary training differs for each role.

A proposed exercise can use a fictional request with conflicting dates. Ask the assistant to classify it, then ask staff to explain the source of the classification and the correct next step. Observe whether they check the record or accept the answer because it is well written. This tests practical review competence without involving operational military information or real personnel data.

Workload needs honest accounting. If officers spend time repairing AI outputs, include that effort in the programme measure. If specialists must manually prepare every dataset, include their time too. A claimed automation saving that moves work to another team can create staffing pressure while appearing successful in a narrow dashboard.

Continuity requires a maintained record. The office can keep the task definition, accepted data sources, model version, evaluation cases and unresolved problems together. A departing employee should leave enough information for an authorised successor to operate the service. This is an original management example, not a statement that the audited departments already use that exact process.

Contractors add another responsibility boundary. The public body needs to know which decisions remain its own and which support functions the supplier performs. A service agreement should explain incident escalation and knowledge transfer. A supplier’s general AI training course may be useful, but it does not replace task-specific preparation or the buyer’s ability to interpret its own evidence.

There is also a capacity question. A reviewer who receives more exceptions than they can handle will begin approving mechanically or allow the queue to grow. A pilot can measure exception arrival, review time and unresolved cases. Managers can then decide whether the task needs a narrower scope, a better interface or additional staffing before extending it.

The benefit of workforce planning is reliable adoption that survives routine change. The failure is a collection of pilots dependent on individual enthusiasm, with no durable owner for corrections or evidence. Public audits are useful precisely because they examine management conditions around the technology. They should be read as dated findings with traceable follow-up, rather than a permanent verdict on an entire institution.

For an administrative team, the next step is a responsibility map grounded in the service. Identify who can approve a change, correct a record, suspend a feature and answer a user. Then test those duties with realistic but safe examples. AI capability becomes useful when enough people understand the workflow to keep it accountable after the demonstration ends.

A periodic handover rehearsal can expose gaps early. Ask a colleague outside the original pilot to operate the review queue from the maintained instructions. Record the questions they cannot answer. The resulting fixes improve resilience and reveal where a programme is relying on tacit knowledge rather than funded roles.