You can buy the platform, the agents and even a certificate that an agent performs well on your cases. You cannot buy your people's judgement about those cases. That judgement decides every held action, every refusal and every widening of depth, so it has to be built inside the business, on its own work, by its own people.
What can a business actually buy?
More than ever. A business can license a platform that enforces bounds at the moment of action. It can start from pre-trained agents built for common jobs such as invoice matching or customer correspondence. It can have each agent certified by replaying it against real cases from its own operation, with every answer scored against a floor.
None of that is trivial, and none of it should be built from scratch. The point of buying is to stop spending effort on parts that are the same everywhere. Bounds, records and certification mechanics are the same everywhere. Which supplier's bank detail change looks wrong is not.
What stays behind after the purchase?
Everything that depends on how your business decides. A governed agent stops at defined points and hands the case to a named person: the held payment, the drafted refund over the customer's limit, the letter waiting to be sent at depth 2. The quality of the whole system at those points is the quality of the person who picks up the case.
Consider a mid-sized firm that buys an agent to prepare customer replies in a regulated service. The agent drafts a reply to a customer whose account has been flagged internally for a money laundering concern. The draft is polite, accurate and complete. It also mentions that the account is under review. A reviewer who knows what tipping off is stops it. A reviewer who does not approves a well-written breach. No vendor can supply that reviewer. The knowledge is generic; the instinct to apply it under queue pressure is built only by practice on the firm's own cases.
Why can't vendor training fill the gap?
Because vendor training teaches the product, and the risk lives in the process. A vendor can show where the held queue is, what the buttons do and how to read a verdict. That is necessary. It is not enough, for three reasons.
First, the cases are yours. Training set on someone else's invoices teaches someone else's exceptions. Second, the thresholds are yours. Your customer limits, your red lines and your appetite decide what gets held, and a reviewer must know why each exists. Third, the habits are yours. A team that has always approved the backlog on Friday afternoon will approve the agent's backlog on Friday afternoon, unless someone changes the habit deliberately.
The same applies to adoption as a whole. Organisations move along an adoption curve at the pace their people can absorb change, not the pace at which licences are signed. Buying ahead of that pace produces shelfware or, worse, an agent running at a depth its reviewers cannot safely oversee.
What is the trade-off?
Time and the patience of whoever signed the contract. Building judgement takes weeks of real decisions, which feels slow beside a platform that can be connected in days. Some leaders will read that as a reason to switch the agent on and train later.
That is the wrong trade. An agent switched on before its reviewers are ready either produces losses through careless approvals or stalls because nobody trusts its drafts. Both cost more than the weeks saved. The better trade is to run the agent at depth 1 or 2 while people learn on its drafts, which is why the advice to train the reviewers before you buy the agent is less paradoxical than it sounds: reviewers who already understand verdicts and provenance shorten the shadow period once the agent arrives.
There is a second trade-off. Building training in-house means the process owners give time they would rather spend on their day jobs. That time is the investment. A business that will not give it should not expect its agents to move past drafting.
Is it one effort or a standing one?
Standing. The agent changes every time its certificate is renewed, and certificates lapse after 90 days. A run that clears no depth drops it to depth 1 the same day, which sends work back to people who may not have handled it for weeks. New capabilities arrive and old ones narrow.
So training cannot be a launch event. The question of whether AI agent training is a one-off or a programme has a short answer: a programme, with calibration sessions, updates when the agent changes, and recertification on a cycle. New joiners are the clearest test. They arrive into a team where the agent already does the routine work they would once have learned on, so a business needs a deliberate route for training new joiners when AI agents already run part of the work, or each new hire becomes the least-prepared reviewer in the queue.
What should you do differently?
Put the people plan in the business case, at the same weight as the licence. Before signing, name who will review the agent's held actions, how they will be trained on your own cases, how their competence will be assessed, and what happens when they leave. The guide to training teams for AI agents sets out what each group must learn and in what order.
Then change three habits:
- Budget training as part of the adoption, not after it. If the training line is empty, the business case is incomplete.
- Use the agent's own drafts as the curriculum. While it runs at low depth, every draft is a practice case with a known outcome.
- Make approval depend on competence. A business can adopt the practice of granting approval rights only to people with a current certificate, depth by depth, so the agent's authority never outruns the people overseeing it.
Where the Academy fits
Pryme Intelligence Academy supplies the structure but not your cases, and is built that way on purpose. Lessons, labs and tests are free, labs run in a free Build workspace, and module tests are set in each learner's own industry and role. The Operator certificate rests on a two-week practicum of the learner's real review decisions, so what is certified is judgement shown on your own work.
Questions people ask
Does buying pre-trained agents reduce the training our people need?
It reduces the time spent building the agent, not the time spent learning to oversee it. A pre-trained agent still meets your thresholds, your exceptions and your customers. Your reviewers still need to decide its held actions correctly.
Who should design the in-house training?
The people who own the process the agent works in, with learning and development running delivery. Process owners know which cases are hard and why each threshold exists. Learning teams know how to schedule, track and assess.
How do we know when people are trained enough?
When their decisions on real cases are consistently right and they never miss a red line. Attendance and quiz scores are early signs, not proof. Measure reviewers on a run of real decisions before the agent's depth widens.