Pick one champion per team that works with an agent, chosen for sound judgement in review rather than enthusiasm. Train them a level deeper than colleagues, protect a few hours of their week, and write a one-page charter. Then connect them through a monthly meeting, a shared channel and a direct line to each agent's owner.
Why do agents need champions at all?
Most questions about an agent arise mid-shift: why did it hold this, can I send that, is this refusal a fault? If the only person who can answer sits in another building, reviewers guess, or work around the agent. A champion in the team answers on the spot, from the record, and notices when the same question keeps coming up. That pattern is often the first sign that the agent's brief, or the training, needs to change.
What does an AI agent champion network do, and who owns it?
A champion network is a group of people across teams who help colleagues use a capability well. For governed agents, champions do four things: answer colleagues' questions about why the agent held or refused something; run or support calibration; collect patterns in corrections and pass them to the agent's owner; and spot early when people are working around the agent.
Champions are not the agent's owners, and they do not change its brief or bounds. They are not compliance officers either. The charter must say so, or champions drift into roles they have no authority for.
The network is usually owned by the head of operations or a transformation lead, with learning and development supporting. The shorter answer to this question sits in how do you build an internal champion network for AI agents.
What are the steps?
- Network owner: maps where agents run. One row per team, with the agent, its depth and the number of reviewers.
- Team leads: nominate one candidate each, using a written profile: decides well in sampling, explains clearly, respected by peers, willing to raise problems.
- Agent, depth 1: supplies evidence for selection. Sampling accuracy and calibration agreement for each candidate, drawn from the record.
- Network owner: confirms champions and agrees protected time with each team lead, in writing.
- Champions: complete deeper training: full reviewer certification, plus sessions on facilitation and on how agent changes are made and certified.
- Network owner: issues the charter and sets up the monthly meeting and channel.
- Champions: run their first calibration session, observed by the network owner or an experienced champion.
- Agent owners: hold a standing slot in the monthly meeting to hear patterns and explain changes.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Map deployments | Lists agents, depth, reviewer counts | 1 | Network owner |
| Nominate | Nothing | none | Team leads |
| Selection evidence | Summarises sampling and calibration results | 1 | Network owner judges |
| Confirm and protect time | Nothing | none | Network owner, team leads |
| Deeper training | Nothing | none | Champions |
| First calibration | Lists held cases for the pack | 1 | Champion facilitates |
| Monthly meeting | Summarises corrections by cause | 1 | Champions, agent owners |
What should the charter say?
Keep it to a page. The champion network charter template covers the full set; the essentials are:
- Purpose: help colleagues review agent work well and surface problems early.
- Duties: answer questions, run calibration, collect correction patterns, report workarounds.
- Not duties: changing the agent, approving exceptions on others' behalf, disciplinary matters.
- Time: the protected hours each week, agreed with the team lead.
- Escalation: to the agent owner for agent behaviour, to compliance or risk for anything that looks like a breach, to the network owner for anything else.
- Term: a year, renewable, so the role rotates and does not depend on one person.
Which controls keep the network healthy?
Selection on evidence. Evidence: the selection record with sampling and calibration data.
Protected time honoured. Evidence: a quarterly check of hours actually spent, from champions' own logs.
Certificates current. A champion whose reviewer certificate lapses steps back until it is renewed. Evidence: the register of certificates and expiry dates.
Speaking up is safe. Champions must be able to report that the agent, or a team practice, is wrong without consequence. That depends on a speak-up culture. Evidence: issues raised through champions and the response to each.
Agent changes go through owners. Evidence: each pattern champions report has a recorded decision by the agent owner.
What should you measure?
- Questions answered by champions versus escalated to agent owners. A healthy network handles most locally.
- Calibration sessions held per team per month.
- Correction patterns reported and acted on, and the time between report and decision.
- Workarounds found. A rising count early on is good news: champions are looking.
- Champion turnover, a warning sign if high.
Where does it usually go wrong?
Enthusiasts, not reviewers. The person most excited about agents is not always the person who decides best. Select on evidence.
No protected time. Champions with full queues stop championing within a month.
Champions as unofficial owners. Teams start asking champions to "just change the agent". The charter and the agent owner's slot in the monthly meeting prevent this.
Bad news filtered out. Champions report only successes. Ask every month: what did people work around?
Audit left out. Internal audit will eventually test how well agents are reviewed; involving them early helps. See how to train the internal audit team to review AI agent decisions. The full programme is in training teams for AI agents.
Some organisations require champions to hold a professional certificate, such as Pryme Intelligence Academy's Certified Pryme Operator, so their standing rests on something anyone can check at source.
Questions people ask
How many champions do we need?
One per team that reviews an agent, plus cover where teams are large or work shifts. Fewer than that, and some teams have nobody nearby to ask.
Should champions be team leads?
Usually not. Team leads already carry authority and workload. A respected peer often gets more honest questions.
What happens when a champion leaves?
The charter's one-year term means a successor should already be identified. Hand over the calibration library and the open patterns log.