Can you trust AI to talk to your customers? The case for gates and approvals

Why is the fear legitimate?

You have seen the screenshots. A chatbot invents a refund policy, quotes a price that does not exist, or cheerfully agrees to something the business never offered. When that happens, the customer does not blame the software. They blame you.

For a small business, one bad exchange costs real money and real reputation. So the fear is not technophobia — it is a rational response to systems that were built to sound confident rather than to be accountable.

The mistake is concluding that AI cannot be trusted at all. The right conclusion is that AI without governance cannot be trusted. Those are very different problems, and the second one has an answer.

What does grounding actually mean?

The first control is limiting what the AI is allowed to know. In ReQuest DESK, answers are grounded only in your own published knowledge base and service catalog — nothing scraped from the wider internet, nothing improvised.

Publishing is an explicit act. Drafts stay drafts until someone approves them, amendments are time-boxed, and coverage mapping shows you where the knowledge has gaps before a customer finds them. If the answer is not in your published material, the system does not guess — it hands the request to a person.

That single rule removes the most common failure mode. An AI that can only repeat what you have signed off cannot invent a policy you never wrote.

Who decides what the AI is allowed to do?

Grounding controls what the system says. Autonomy controls what it does. Every case in ReQuest DESK runs through an eight-phase ITIL-style chain — Accept, Identify & analyse, Summarise, Plan, Initiate, Resolve, Validate, Communicate — and each step can be assigned to the AI, to a human, or placed behind an approval gate.

You choose the mix per request type. Sensible defaults look like this:

Full autonomy exists, but it is an explicit switch you flip per step — never a default you discover later. The governance model is designed so the cautious setting is the starting point.

• Anything touching money — refunds, credits, pricing exceptions — sits behind an approval gate.
• Clinical or safety-sensitive steps route to a qualified human, always.
• Policy exceptions pause the chain until a named person signs off.
• Routine, well-documented requests — opening hours, booking changes, status checks — can run end to end.

What happens when something goes wrong?

Governance also means being able to reconstruct events afterwards. Every request is logged with an owner and timestamps, so each request has a named owner and a full audit trail from first contact to resolution.

Classification comes with confidence scores. When the system is unsure what a customer is asking for, that uncertainty is visible and the case routes to a person rather than proceeding on a guess. Committed resolution times mean a stalled case surfaces instead of quietly dying.

None of this makes errors impossible. It makes them visible, attributable and correctable — which is exactly the standard you would hold a new employee to.

Trust is a dial, not a switch

The practical path is incremental. Start with the AI drafting and a human approving every send. Watch the audit trail for a few weeks. Then loosen the gates on the request types it handles well, and keep the gates on the ones that matter most.

That is how you would train a person, and it frees your people for the work that actually needs them. The question was never whether AI can be trusted in the abstract — it is whether the system in front of you gives you the controls to earn that trust step by step.