For a support adviser, the practicum is ten working days on their own real queue: drafted replies, held refunds, handovers and refusals. Because volume is high, the reviewer checks every refund and handover and a fixed sample of replies before they go. The adviser passes with at least 95 decisions in every 100 right and no missed red line, such as sharing account details with an unverified caller.
What is the practicum, and who owns it?
Support advisers spend most of the day on the AI agent's drafts: reading, editing, sending. Some also approve refunds the agent held. A practicum checks that they do both well on live tickets, not just in a quiz.
The support team lead owns the practicum and either reviews or names a certified reviewer. The support manager sets the business's own limits and the red lines. The adviser owns each decision and leaves a short reason on anything they changed or escalated.
How do you build the queue?
Support volume is high. Thirty or more tickets a day is normal, so a fortnight produces several hundred decisions. That makes the bar easy to apply but the review hard to sustain, so split it:
- every held refund or credit the adviser decides;
- every handover: complaints, vulnerability signals, out-of-scope requests;
- every refusal the adviser handles, such as a request to change account details from an unverified contact;
- a fixed sample of drafted replies, for example one in four, chosen by the agent at random and set before the day starts.
Seed anonymised past tickets for anything thin. For example, a seeded ticket might be a polite request to update a delivery address from a sender whose email does not match the account. The agent's draft confirms the change. The adviser should notice the mismatch, hold the change and verify through the account's registered contact.
The steps
- Person (team lead): brief the adviser on the bar, red lines, sample rule and log.
- Agent, depth 1: route the adviser's tickets to the practicum queue and mark which replies are in the sample.
- Agent, depth 2: present each ticket with the draft, the help articles or account records it used, and the deciding bound.
- Person (adviser): decide: send, edit, hold, refund, decline or hand over.
- Person (reviewer): decide independently on every refund, handover and refusal, and on each sampled reply, before it goes.
- Person (reviewer): release refunds and sampled replies during the practicum.
- Agent, depth 1: log both decisions, the match and any red-line flag.
- Person (reviewer, then assessor): review the log at the end.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Drafted reply | Drafts with sources | 2 | Adviser edits; reviewer checks sample |
| Refund within limit | Prepares | 4 | Adviser decides; reviewer releases |
| Refund above limit | Holds | 4, held | Reviewer approves |
| Handover | Hands over with a summary | escalation | Adviser handles |
| Refusal | Refuses and routes | refused | Adviser explains; does not work round it |
| Log | Records decisions | 1 | Confirming |
How do you set and apply the bar?
A decision is correct if it matches the reviewer's or the reviewer agrees it was sound. A reply edited for tone but accurate is correct. A reply sent with a wrong delivery date, price or policy is not.
The red lines: sharing or changing account details for an unverified person; refunding outside policy or authority; missing a vulnerability handover; sending a reply that commits the business to something the policy does not allow. One miss fails the practicum.
Speed is not part of the bar during the practicum. An adviser who handles fewer tickets but gets them right is ready; their pace will come. One who keeps pace by sending drafts unread is not.
Advisers also earn credit for teaching the agent. When a draft is wrong because a help article is wrong, reporting the article is the right decision and should be logged as one.
Which controls matter, and what evidence do they leave?
- Fixed sample rule. Evidence: the sample list generated before each day.
- Reviewer release of refunds and sampled replies. Evidence: both names in the append-only record.
- Independent reviewer decisions. Evidence: logged before comparison.
- Daily debrief. Evidence: a line per mismatch.
- Full log. Evidence: the practicum log template for customer support agents.
What should you measure?
Correct decisions out of the total, split by refunds, handovers, refusals and sampled replies. Red-line misses (zero). Help article errors reported. And reopened tickets from the practicum fortnight, measured a week later.
Where does it usually go wrong?
The reviewer checks only replies that look wrong. The sample must be random and fixed in advance, or it measures nothing.
The fortnight falls in a sale or launch, when volume swamps review. Move it.
Advisers learn to send drafts unchanged because most are fine. Seed drafts with errors: a wrong delivery window, an outdated returns period, a refund promised before the item is back.
Team leads run their own practicum on a different set of decisions; compare running a practicum for support team leads and the practicum log template for support team leads.
Where the Academy fits
Pryme Intelligence Academy's Certified Pryme Operator practicum is two weeks of real review decisions, at least 95 in every 100 correct and no missed red line, signed off by a Pryme assessor. Advisers new to agents can start with the free Foundations badge, about 4 hours. More in the guide to AI literacy and certification for people who work with agents.
Questions people ask
Does a one-in-four sample really cover enough?
Over a high-volume fortnight it yields well over 100 checked replies, alongside every refund, handover and refusal. That is enough for the bar to mean something.
Should customers know an adviser is in a practicum?
No need. Every refund and every sampled reply is checked before it goes, and the agent's involvement is already disclosed.
What if the adviser handles live chat as well as email?
Include both. Chat decisions are faster and harder to review, so sample chats after the fact and debrief them the same day.