Grounding AI So It Cites or Hands Off

The most dangerous trait in a support AI is fluency without grounding. A model that produces a confident, well-written, completely wrong answer is worse than no AI at all, because it manufactures trust it hasn't earned. In support, where answers touch refunds, account access, policy, and compliance, an ungrounded confident answer is a liability with a friendly tone. The design goal, then, is not maximum coverage. It's a system that, on any given question, either cites or hands off.

Grounding means an answer is tied to a real source — a knowledge base article, a policy document, the customer's actual order or account state — and that the tie is inspectable. When the AI responds, it should be able to show what it drew from. This isn't decorative. Citations are what let an agent reviewing the conversation verify the answer in seconds, what let the customer trust the response, and what make the difference between "the AI said so" and "here is the policy that says so." An answer that can't be traced to a source should be treated as a guess, and guesses don't belong in support.

The corollary is the hard part: when the AI cannot ground an answer, it must not answer anyway. This is where most systems quietly fail, because the model's default behavior is to produce something plausible. Plausible is the enemy. The system needs an explicit boundary — a confidence and grounding check that, when not met, stops the AI from improvising and routes the conversation to a human instead. Handing off is not the system failing. It is the system working exactly as designed on a question it shouldn't answer.

This reframes how to measure an AI agent. The instinct is to maximize the share of conversations the AI handles end to end. But a high handle-rate achieved by answering questions it shouldn't is a problem disguised as a metric. A better frame: of the conversations the AI chose to answer, how often was the answer grounded and correct — and of the ones it shouldn't have answered, how often did it correctly hand off? An AI that knows the edges of its competence is more valuable than one that's confidently everywhere.

Grounding and billing are tightly linked, and this is where the discipline pays off commercially. If you only earn revenue on verified resolutions, and verification requires grounding, then the incentive to ground is built into the economics rather than bolted on as a guideline. An ungrounded answer can't be a verified resolution, so it can't be billed. A handoff isn't billed either. The pricing model and the safety model reinforce each other: the system makes money precisely when it does the safe, grounded thing, and makes nothing when it bluffs or bails. That's the alignment you want.

Always-on guardrails sit alongside grounding. PII redaction, jailbreak resistance, and grounding checks shouldn't be optional features you remember to turn on; they should be defaults that run on every conversation, including the ones that get billed. A subtle but important detail: the guardrails have to run on the path that actually replies to the customer, not only on a sandbox or a copilot suggestion. It's easy to validate the safe version and ship the unsafe one. The thing that goes to the customer is the thing that has to be checked.

We'll be honest about the state of the art, including ours. Grounding detection is imperfect, retrieval can surface the wrong document, and a confident model can still slip a fabrication past a citation check. There is no version of this where you flip it on and walk away. The responsible posture during alpha is conservative thresholds, humans supervising the boundary, and a strong bias toward handing off when grounding is uncertain. The goal isn't an AI that never hands off. It's an AI you can trust precisely because it does — an AI whose silence on a question is itself a feature, because it means it knew the limit of what it could prove.