Training people to oversee AI agents is a control, not a development benefit, so fund it where controls are funded: in the operational risk budget, alongside the agent itself. Learning and development budgets are discretionary and are cut first. A reviewer's competence to release a held payment cannot be discretionary, and it cannot wait for the next budget round.
What does the money actually buy?
Look at what happens without it. An agent prepares a supplier payment and holds it at depth 4 because the bank details changed through an emailed instruction, an external claim that cannot authorise a payment. A person must decide whether to release it. If that person has been trained to verify through a known channel, the fraud stops there. If not, a single misdirected payment of, say, 50,000 leaves the business.
The training bought a lower chance of that loss. That is what a control is: something that reduces the likelihood or impact of a risk event. Firewalls, segregation of duties and reconciliations are funded on that basis, and nobody asks whether they fit the year's development priorities. Reviewer competence belongs in the same category, because it does the same job at the point where the agent hands the decision back to a person.
Why does the L&D budget fail this kind of spending?
L&D budgets are built for a different purpose. A learning budget typically funds individual development, leadership programmes, professional qualifications and course allowances. It is allocated per head or per department, prioritised by managers, and treated as discretionary when costs are squeezed.
Agent oversight training does not fit that shape. It is not optional for the people who approve held actions. It has fixed renewal points: agent certificates lapse after 90 days, a run that clears no depth drops an agent to depth 1 the same day, and people certificates expire on their own cycle. And its cost scales with the agent's depth, not with headcount. Put it in a discretionary budget and the first lean quarter will defer recertification, leaving uncertified people approving depth 4 actions. Nobody intended that; the budget structure produced it.
Who should own the spend?
Whoever owns the risk. Each agent has a named owner who decides what depth to seek and answers for its outcomes. If that owner wants an agent to release payments at depth 4, they are accepting the risk that a held payment is wrongly released. The cost of making the reviewers competent is part of the cost of that depth.
When training is funded elsewhere, the owner is accountable for a control they cannot pay for. They can ask L&D for reviewer certification and be told the budget is committed. Funding from the risk or operational budget, ideally inside the agent's own business case, puts the decision and the cost in the same hands. What an AI agent training budget usually covers sets out the lines: vocabulary for everyone, assessed review for approvers, builder training, recertification, and time away from the queue.
What does the auditor see?
Controls leave evidence, and auditors test them. A reviewer's certificate, with what was assessed and when it expires, is control evidence in the same way a reconciliation sign-off is. So is a register showing that everyone who approved held actions in a period held a current certificate.
When training sits in L&D, its records sit in a learning system, organised by course completion. When it sits in risk, the records are organised by agent, depth and approval right, which is how an auditor will ask for them. A business can adopt the practice of tying approval rights to a current certificate, depth by depth, and the register then becomes both the training record and the access control.
What is the trade-off?
Two risks come with moving the money. The first is that risk teams turn training into a compliance exercise: long policy modules, annual attestations, a tick in a box. That would be worse than what L&D does today. The second is that L&D's expertise in designing and delivering learning is lost.
The answer is to split funding from delivery. Risk funds the training and sets the standard: who must be certified, for which agents and depths, and how often. L&D designs and runs it, using the process owners' real cases. The operational owner signs off the assessment. Each does what it is good at, and the money cannot quietly disappear.
There is also a budgeting cost. Risk budgets attract scrutiny, and a new line will be questioned. That is healthy. A line that can be defended as loss prevention will survive scrutiny better than one defended as development.
What should you do differently?
Put reviewer training in the agent's business case, costed against the depth sought. Use an AI agent training budget template to show the lines: initial training, assessment, certificate issue, recertification, and the reviewers' time. Forecast recertification from the expiry dates, because they are known in advance.
When you take it to the committee, frame it as a control with a loss it prevents, not a course with a benefit. How to get budget approval for AI agent training covers the argument in detail. And if the finance team asks who will prepare the numbers, an agent can help: whether an AI agent can prepare a budget depends on depth, but at depth 2 it can draft one from the register for a person to check and submit.
Much of the content need not cost much. Pryme Intelligence Academy's lessons, labs and tests are free, and professional certificates carry an issue fee, so the larger cost is reviewers' time, which the risk budget should recognise. The order of who to train first is in the guide to training teams for AI agents.
Questions people ask
Does this mean L&D has no role in agent training?
No. L&D should design and deliver it, track progress and keep the learning experience good. What changes is who funds it and who sets the standard, so the training cannot be cut as a discretionary item.
What if the business has no separate risk budget?
Put the training cost in the same budget line as the agent. If the agent is funded from operations, the training is too. The principle is that the cost of the control sits with the cost of the risk it controls.
Should general AI awareness training move to the risk budget too?
Not necessarily. Broad awareness for staff who never approve agent actions can stay in L&D. Move the training that protects specific approvals: reviewers, team leads, builders and agent owners.