Training providers adopting AI need a way to decide which work goes to AI and which stays with people. Less of that work needs to stay human than most providers assume. The deciding factors are who owns the outcome, what happens if it goes wrong, and whether the person checking the work can spot an error.

This article covers what inspectors and regulators currently look for, why a human in the loop is not a safeguard on its own, a five-question decision tree for placing any task, and a simple test for choosing where to start.

Why less work needs to stay human than you might think

Many providers treat apprenticeship delivery as human work by default and see AI as useful only at the edges. That view is often a reflex rather than a principle. AELP’s State of the Skills Sector 2026 report found that only 51% of providers plan to invest in AI and automation over the next two to three years.

This is not an argument that people matter less. The human part of the work matters more, so it is worth being precise about what it is.

What Ofsted and Ofqual currently say about AI

Ofsted’s published statement says it does not look at AI as a stand-alone part of inspection and does not directly evaluate any AI tool. Any evaluation is about the provider’s decision-making, what it has considered, and the impact on learners. Ofsted’s approach to AI is still developing, so it is worth checking for updates.

The further education and skills operating guide for inspectors, in use from September 2026, tells inspectors they do not need to ask whether AI is used or search for it. Where they come across it, they record how leaders chose to implement it, its impact, how it is monitored, and the provider’s checks and balances to make sure it is accurate, safe and used in the best interests of learners and apprentices.

Ofqual’s position, restated in July 2026, is that using AI as the sole mechanism for awarding marks does not comply with its current regulations.

Neither regulator is saying providers should avoid AI. Both are asking who checked the work, and whether they were able to.

Why a human in the loop is not a safeguard on its own

A human in the loop only protects learners if that person can spot a problem and actually looks. In one widely reported case in 2025, learners on a funded apprenticeship programme quickly spotted AI-generated course materials, including references to US legislation and a voiceover that changed accent partway through.

Jisc’s AI in marking and feedback pilot found that formative assessment is where AI delivers the most value with the least risk. It also found that formative assessment is not a no-stakes use case. The same work run through AI more than once can produce slightly different results, so moderation and human oversight remain important.

Overseeing AI works best when it is treated like internal quality assurance (IQA) for a new kind of assessor. Three questions apply:

• Can they judge? The person checking is occupationally competent and can tell good work from bad.
• Do they check? Time is set aside and a sampling plan is followed. The check is real, not a click.
• Do they answer for it? A named role signs off the sample, not the system.

If the person checking could not have spotted the error, you do not have a human in the loop. You have a human on the hook.

A five-question decision tree for any task

Ask these questions in order and stop at the first one that places the task.

1. Does a rule require a named person to make this decision? Examples include marking that counts towards a qualification, safeguarding decisions, and decisions with a legal or similarly significant effect on a learner. If yes, a named person decides. AI may prepare the ground, or the task stays human only for now.
2. Is the human contact itself the point? Examples include a disclosure, a difficult conversation about whether a learner will complete, or a wellbeing check-in. If yes, a person leads and AI supports with preparation, notes, actions and follow-up. This list is shorter than most people expect.
3. If it goes wrong, is the harm serious or hard to undo? This covers harm to a learner, an employer, your funding or your reputation. If yes, AI drafts and a person decides, with every output checked before it goes anywhere.
4. Could the person checking spot a wrong answer? If not, fix the skills or the standard first. Large language models are not deterministic, so 10 good outputs do not guarantee the 11th.
5. Would you sample this if a new team member did it? Hold AI to the same standard you hold people to. Start with one cohort or programme and a larger sample, then reduce it as the tool earns trust, without ever dropping to zero.

The order is the safeguard. Starting from “this is repetitive, so let’s automate it” skips the harm question. “Human only” also means human only for now. Rules change and tools improve, so revisit these decisions several times a year.

How common tasks fall into four lanes

The decision tree places each task in one of four lanes.

• AI does it, people sample check. Answering learners’ out-of-hours questions fits here. Any sign of distress or a safeguarding concern goes back to a named person.
• AI drafts, a person decides. Checking off-the-job training hours against logged evidence fits here, because errors affect funding. So do selecting an IQA sample by risk and writing individualised feedback on learner work.
• A person leads, AI supports. Withdrawal conversations and progress reviews fit here, with AI helping on preparation, summaries, actions and follow-up.
• Human only for now. Safeguarding decisions stay here.

Choosing where to start with AI

Plot each candidate task on two axes. Readiness covers whether there is a clear standard, usable data, someone who can check, and a team willing and able to do it. Value covers how often the task happens, how much time it gives back, and whether you can measure the result.

• High readiness, high value. Pilot here. An IQA sample is a good example: sampling is already well defined, the data exists, and quality underpins reputation.
• High readiness, lower value. A first draft of a self-assessment report may be ready to try but happens too rarely to lead.
• Lower readiness, high value. The gaps are your to-do list. Spotting apprentices at risk of withdrawal is high value, but most providers do not yet have a standard to test AI against.

Tasks low on both can move as tools and regulation change. As a rule of thumb, start with high readiness and high value, in a low-harm lane.

What stays human

Three things stay with people. Judgement, because AI tools tend to agree they can do whatever is asked. Empathy, because relationships and pastoral care sit at the centre of apprenticeship delivery. And accountability, because “the AI told me to” is never an acceptable answer. The organisation is accountable for every output, whether a person or a tool produced it.

Watch the webinar below to learn more and download the decision tree.

AI task decision tree