Sagi Eliyahu is CEO of KMS Lighthouse. Leading the company’s vision to disrupt the knowledge management market.
For most of my career, I’ve watched companies run on knowledge that disagrees with itself. The refund policy in the help center says 30 days, the agent script says 45 for loyalty members and the PDF a team lead circulated last spring says 30 unless the item was on sale. No one planned it that way. It built up one update at a time, usually for good reasons.
For years this mostly worked, not because the knowledge was good but because people are good at compensating for bad knowledge. An experienced agent knows which document is stale, knows to ask the person at the next desk and knows that when the system says one thing and the floor says another, the floor usually wins. That informal correction layer has quietly held customer service together for decades.
AI doesn’t have that layer. When you point a model at your content and ask it to answer customers, it has no way of knowing which version is the real one, so it picks one, phrases it confidently and moves on. When it picks wrong, the customer hears the mistake in the same calm, confident tone as a right answer.
The Problem Was Always There, AI Made It Public
The most cited example is still Air Canada. In 2024, a Canadian tribunal held the airline responsible after its website chatbot told a grieving passenger he could apply for a bereavement fare refund after his trip, even though the airline’s own policy page said otherwise. Air Canada argued that the correct information was available elsewhere on its site. The tribunal didn’t accept that, since the chatbot was part of the website and the website belonged to the company.
I see that case less as an AI failure and more as a knowledge management failure. The bot didn’t invent a policy out of nowhere. Somewhere in what the company had published, two versions of the truth lived side by side. A human agent might have caught it, but the machine had no reason to.
Contradictory knowledge used to be an internal annoyance that meant slower onboarding, inconsistent service and the occasional escalation. Once AI is answering customers directly, it becomes an external commitment, and whatever the system says, you may have to stand behind.
Why A Better Model Won’t Fix It
When an AI assistant gives a wrong answer, the instinct is to blame the model and go shopping for a better one. In my experience, that’s rarely the problem. The model is the part everyone sees, but underneath it sits the knowledge it draws from, and that layer decides the quality of every answer built on top of it.
Most organizations have invested heavily in the visible layer and very little in the one underneath. They’ve bought the bots and co-pilots, but far fewer have asked a simple question: If we put all our content in one room, would it agree with itself? Usually it wouldn’t. For years, knowledge management was treated as back-office housekeeping, something the support team worried about. AI has changed that, because knowledge is now the foundation every customer-facing answer is built on.
In a typical enterprise, knowledge is spread across wikis, shared drives, ticket notes, policy documents, training decks and email threads. Each piece was correct when someone wrote it, but few have an owner today and many never expire. Retrieval tools are good at finding relevant content, but they aren’t built to decide which of two documents is authoritative. That’s an editorial judgment that has to be made before the AI ever sees the question, and making it is exactly what knowledge management is for.
Fixing The Layer Underneath
The companies I see getting this right treat knowledge more like a product than a library, and they give knowledge management the same attention they give the AI itself. A few practices stand out.
One Answer Per Question
If the same policy is written in three places, two of them should be removed or point to the third. It’s easy to agree with and hard to do, because every team believes its version is the one that matters.
Clear Ownership
Every piece of customer-facing knowledge needs a named owner and a review date, so that when a policy changes, one person updates one source and everything downstream inherits the change.
The Same Answer For People And Machines
It’s tempting to build a cleaned-up knowledge set for the AI and leave agents with the old mess, but that only creates a fourth version. The person on the phone and the bot in the chat window should be reading from the same verified answer, because when they disagree, customers notice quickly.
Verify Before You Scale
Before expanding AI to a new topic or channel, audit the knowledge behind it. Finding a contradiction in a review is far cheaper than finding it in a legal ruling or a viral screenshot.
Treat Gaps As Signals
When the AI can’t find an answer, or agents keep escalating the same question, treat it as a map of where your knowledge is missing or broken.
The Edge Underneath
Right now everyone wants to know which model is smartest, and I understand why. But for most customer-facing work the models are converging, and your competitors can buy the same intelligence you can.
What they can’t buy is your knowledge and the discipline that keeps it organized, verified and consistent. Two companies running the same AI will deliver very different experiences depending on what sits beneath it: One will answer accurately across every channel, while the other will occasionally tell a customer something it never meant to promise.
Your people spent years working around messy knowledge, and they were generous about it. AI won’t extend you the same courtesy, so before you let it speak for your company, make sure your company is saying one thing.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?