We Audited 24 Public Help Centers for AI-Support Readiness

An AI agent cannot tell whether a policy answer is current, and neither can a customer. Yet at least 14 of the 24 help centers showed no visible update date on their policy articles at all. Only three dated their articles clearly enough to trust the answer was current. Two others were dated, to 2022.

When the source is undated, an AI agent serves a three-year-old refund policy with full confidence and no idea it is stale.

In 3 of the 24, two public pages gave conflicting answers to the same core money question, and we only counted cases we could confirm from two public sources. A productivity tool mid-migration listed two different refund windows depending on where you bought. A supplement brand stated 'all fees are non-refundable' on one page and offered a 30-day money-back guarantee on another. A consumer investing app stated its monthly fee as one amount on one page and a different amount on another.

None of these teams is careless. This is what happens when help content grows across teams and years. But it is exactly the kind of conflict an AI agent resolves by guessing, sometimes wrong, on money, in writing.

More than half, 13 of 24, answered a core policy with language that sounds like an answer but never states the rule: 'generally non-refundable', 'may be eligible', 'in some cases'. One telehealth brand publishes an article literally titled 'What is your refund policy?' that never states the refund policy.

A human reader shrugs and contacts support. An AI agent fills the gap with a plausible answer it invented.

Against a checklist of eight high-risk customer intents, 10 of 24 were missing a clear answer to at least one. Average coverage was 7.25 of 8. Two gaps were systematic: a dedicated answer to 'I was charged the wrong amount' was usually absent, and data deletion was buried in a legal privacy policy rather than the help center, invisible to an agent scoped to support content.

At least 9 of 24 help centers blocked automated retrieval through bot protection or JavaScript-only rendering that returns nothing to a plain request. The content is public to a human in a browser, but unreadable to a naive AI fetcher. The answers exist; the agent cannot load them.

If your AI rollout assumes it will just read the help center, confirm that it actually can.

To be fair, all 24 offered some path to a human. A few gate it behind paid plans or a 24-to-48-hour email queue, but the escalation door exists everywhere. The problem is not that teams do not care. It is that the knowledge an AI inherits is stale, conflicting, vague, or unreadable in ways nobody notices until the agent is live.