Why your AI assistant starts giving outdated answers

An assistant is only as current as its least-recently-updated page, and grounding is not a launch task: it is a maintenance schedule with owners and dates.

The launch was done properly. The sources were gathered, the answers were tested against real questions, and for weeks the assistant was accurate. Then a price changed and nobody told it. A service was quietly dropped, an office moved, a policy was rewritten. The assistant did not get worse. The world moved and it stayed, answering with total confidence about a company that no longer exists.

The honest framing is that grounding is not a launch task. It is a maintenance schedule. The assistant has no sense of time passing; it knows only what its sources say, so its knowledge is exactly as current as the least-recently-updated page among them. And in most companies that update has no owner. Everyone assumed someone else was telling the assistant the news, which in practice means nobody was.

The failure versions are recognizable. The assistant still enthusiastically describing a service the company stopped offering months ago, to a customer who then asks why the website disagrees. The two customers who received opposite answers because one asked before the update and one after, both holding screenshots, both certain. And the quietest damage: the team that loses trust over one stale answer and stops reading the assistant's correct answers too, which wastes every good response it will ever give.

This is where transparent AI earns its keep. An assistant that names what it read from is far easier to trust and far easier to correct: when an answer goes wrong, the wrong page announces itself, the fix lands in minutes, and the customer can see the basis instead of having to take the machine's word. A closed box gets audited as if every error were a lie. An open book gets fixed and trusted again.

The practice is to treat the assistant's sources the way you treat a price list. Every fact that changes operationally, offers, policies, addresses, hours, gets an owner and a review rhythm, however simple. Version the updates so you can look back and see what the assistant knew when it answered, which turns an argument about the past into a lookup. And give any customer who quotes a stale answer a fast correction path, so the damage is one conversation instead of a lost account.

We run an assistant in our own funnel and learned the rot lesson from our own sources, which is why ours says what it read from and says plainly that it is AI. Guidance, not guarantees: how often each source should be reviewed, and what correction path fits your customers, depends on how fast your business changes. Assign the updates a name and a date before your assistant starts testifying, sincerely, about a company that no longer exists.