Should AI agent training be online or in person?

Both, split by what is being learned. Concepts such as depth, verdicts and provenance work well online and self-paced. Judgement does not: deciding held actions, calibrating with colleagues and rehearsing incidents need a live queue and other people present, whether in a room or on a call. The mistake is choosing one format for everything.

What works online?

Anything with a right answer that does not depend on the moment. The shared vocabulary is the clearest case: what each depth allows, what held, refused and withheld mean, how sealed, signed, declared and external facts differ, why certificates lapse. People can learn these at their own pace, in short blocks, and many will retain them better if they can listen as audio on a commute rather than sit through a session.

Practice labs also work online, as long as they run in a safe workspace with sample data. A reviewer can try approving, editing and declining agent drafts there without any risk to customers. Module tests and written exams belong online too: they are easier to schedule, mark and record.

What needs a live queue?

Judgement under real conditions. Three things in particular:

Practicum. A reviewer's real decisions on the live approval queue, checked by a lead. This is the closest thing to proof of competence, and it only exists where the real work is.

Calibration. Reviewers decide the same small set of held cases separately, then discuss where they differed. The discussion is the learning. It works in a room or on a video call, but it must be synchronous; a forum thread loses the argument that makes it useful.

Drills. Walking through what happens when an agent's certificate lapses and it drops to depth 1 the same day, or a fraud attempt reaches the queue. People need to rehearse the sequence together.

How should you combine them?

A useful rule: online for anything you could mark automatically, live for anything you would want to argue about.

When does in person beat remote?

When the work itself is in person. A warehouse returns team that handles physical goods alongside the agent's refund drafts learns faster at the returns bench than on a screen. A contact centre floor calibrates well in a huddle at the start of a shift. Teams that are already remote should not be brought in for training they can do on a call; the live element matters, the room does not.

In person also helps the first time a team meets the agent. A half-day where the team lead walks through real drafts on a shared screen, and people argue about them, sets expectations faster than any module.

An example: an online retailer

Take a hypothetical online retailer with a customer service team working mostly from home, a returns team in the warehouse and a fraud team at head office. Everyone takes the vocabulary online, in their own time, within two weeks. Customer service reviewers do role modules and labs online, then their practicum on the live refunds queue, with calibration on a weekly video call. The returns team calibrates in person at the bench, comparing held refund drafts with the items in front of them. The fraud team does a live drill on an account takeover pattern. Who should start first in that business is set out in who in an online retailer needs AI agent training first, and the sequencing across sites in how to roll out AI agent training across an online retailer.

Format affects cost as well. Online modules cost little to deliver; live calibration costs team time. The budget lines are in what AI agent training costs online retailers. Pryme Intelligence Academy delivers its lessons, labs and tests online and free, and its Operator practicum takes place on the learner's real queue.