Contact centre directors need short, tested literacy on four ideas: depth, certification, the decision record and who stays accountable. They also need one thing most boards miss: metrics that reward careful oversight of AI agents rather than raw speed. Without that, managers learn to clear held actions quickly, and the board's own reporting quietly undoes the controls.
Which "agent" is the board talking about?
Contact centres already call their people agents. Add AI agents and every board paper becomes ambiguous. Settle the language first: people are advisers or colleagues, software is the AI agent. It sounds trivial. It is the difference between a board approving "more agent autonomy" for staff and approving it for software that can issue refunds.
What must the board understand?
- Depth. Each AI agent capability has a depth: 1 read, 2 draft a reply a person sends, 3 send routine replies or make routine changes within bounds, 4 refund or take hard-to-undo actions, 5 coordinate other agents.
- Certification. Depth is earned on the centre's own cases, not switched on. Certificates lapse after 90 days, and a run that clears no depth drops the agent to depth 1 the same day, which can change queue volumes overnight.
- The record. Every verdict is filed in an append-only record. The centre can reconstruct any conversation, what the agent proposed and who approved it.
- Accountability. A manager who releases a held refund owns it. That manager should hold a current credential.
- Disclosure. Under Article 50 of the EU AI Act, people must be told when they are interacting with an AI system. Centres serving EU customers need scripts and agent behaviour to match.
What may AI agents do, and what stays with staff?
| Contact | AI agent | Stays with a person |
|---|---|---|
| Order status or balance enquiry | Answers within bounds at depth 3 | Escalations |
| Complaint | Drafts a reply at depth 2 | Sending it and owning the outcome |
| Goodwill credit within limits | Executes at depth 3 | Anything above the limit |
| Refund to card or account | Prepares and holds at depth 4, if certified | Releasing it |
| Signs of vulnerability or distress | Hands over immediately | The whole conversation |
Why do the board's metrics need to change?
Contact centre boards are used to handle time, first-contact resolution and cost per contact. Applied naively to AI agents, those measures punish the controls. A held refund adds minutes; a refusal sends a contact to a person. If managers are judged on clearing queues, they will release holds without reading them.
Add measures that show oversight is working: held actions released by uncertified people (should be none), the share of holds rejected or changed after review, and decisions sampled from the record and checked by quality assurance. A board that asks for these signals that care is the job.
How should managers be qualified before they approve?
Managers who release held actions should prove competence on real work, not in a classroom. A two-week practicum for contact centre managers, logged with a practicum log template, gives the board hard evidence. Whether managers need a certificate to approve AI agent actions is a policy choice the board should make explicitly. A centre can tie approval rights to a current credential, depth by depth.
Then plan for change. Each time an AI agent re-certifies, its depth can rise or fall, so managers should prepare for each 90-day recertification rather than discover the change mid-shift. Pryme Intelligence Academy's Operator track is one route: a written exam plus a practicum of real review decisions, signed off by an assessor. Compare the fintech board guide, and see the full model in AI literacy and certification for people who work with agents.
Questions people ask
Should an outsourced contact centre's board answer to the client on this?
Yes, where the contract requires it. A client may reasonably ask for evidence that the people approving AI agent actions on their customers are certified, and the record makes that checkable.
Can AI agents handle vulnerable customers?
They should hand over as soon as there are signs of vulnerability. Training should cover how staff pick up those conversations without making the customer repeat themselves.
How much time does board training take?
A few hours of literacy is enough for directors who do not approve actions. Managers who do approve need a full certification route with practice on real cases.