Generative AI — ground it, gate it, govern it

AI in CC · ~7 minute read

Generative AI in the contact centre is real capability and real risk. The disciplined deployment sequence puts hallucination, vulnerable customers and regulatory exposure behind disciplines that actually scale.

The three Gs

Ground the model in authoritative knowledge — retrieval-augmented generation from the real KB, not the model’s training data. Gate the output — human-in-loop on customer-facing content where the stakes warrant. Govern the deployment — named ownership, disclosure where customers interact, logging, fallback paths.

Each is straightforward in principle; each is skipped in haste; together they are the difference between deployment that scales and deployment that produces incidents.

The disciplined deployment sequence

A useful sequence: internal-first (summaries, knowledge drafts, training); agent-facing drafts (agent edits and sends); customer-facing gated (with HITL on borderline); customer-facing autonomous (last, narrow scope, with calibrated fallback).

Most operations should sit at stages 1–2 for some time. Stages 3–4 require operating-model maturity many don’t yet have.

What customer-facing generative needs

Five disciplines before customer-facing deployment. Grounding in current authoritative knowledge. HITL where regulated content is involved. Disclosure so customers know they’re talking to AI. Logging so any decision is traceable. Fallback so any failure routes to a human path.

Skip any of them and a complaint becomes an incident.

Failure modes worth naming

Hallucination on knowledge gaps. Customer-facing deployment too early. HITL skipped under cost pressure. Vulnerable-customer pathway eroded. Prompt-injection exploited. PII leaked into prompts or logs.

Each is documented; each is preventable; each is mostly an operating-model problem, not a model problem.

The closing principle

Generative AI deployed without the three Gs becomes an incident waiting to be discovered. Ground, gate, govern; sequence the deployment; sit at stages 1–2 longer than the vendor recommends.