The Source of Truth Is Now a Liability

The support bot was doing its job perfectly. A customer, frustrated with a billing error, asked for the company’s refund policy. The bot, wired directly into the corporate Confluence wiki, instantly retrieved the official document and provided a crisp, confident summary. The policy it cited, however, had been deprecated six months ago. The link it served up led to an old page someone forgot to archive.

This isn't a hallucination. This is a new and more insidious kind of failure. The model didn't invent a fantasy; it faithfully reported the "truth" it was given. The source was wrong.

Companies are in a frantic race to hook Large Language Models into their internal organs: the SharePoint sites overflowing with conflicting PowerPoints, the Zendesk tickets full of one-off workarounds, the Google Drives that have become digital attics. We call these repositories the "source of truth." That was always a generous term. Now it's a dangerous one.

For years, the mess was contained. The damage from an outdated wiki page was limited by human friction. An employee searching for a process would see the old date, notice the new version helpfully linked by a colleague in the comments, or just ask someone in Slack. Knowledge propagated at the speed of a conversation, with built-in, social error correction.

We have now removed the brakes. By plugging an AI into this chaotic substrate, we have created an engine for confidently laundering outdated information. The AI has no context for organizational history. It cannot tell the difference between a project manager’s brainstorm draft from 2019 and the current, legally-vetted HR policy. It just sees text, vectorizes it, and serves it on command.

The consequences are not trivial. When a sales bot promises a client a feature that was cut from the roadmap two quarters ago, the company is on the hook. When an internal developer assistant recommends a deprecated security library, it introduces a vulnerability at machine speed. The liability for the error no longer rests in some abstract model failure, but squarely on the company's own institutional memory. The AI is a mirror, and it is showing us that our house is not in order.

The real work of this era has nothing to do with prompt engineering or choosing the right foundation model. It is the brutal, unglamorous work of curation. It is about establishing clear ownership for every single document. It is about ruthless archiving. The new power player in the enterprise isn't the AI researcher, but the technical writer, the librarian, the data governor—anyone with the authority to declare a piece of information dead.

We've spent a decade accumulating digital cruft, telling ourselves we'd clean it up later. Later is here. Because the source of truth is now a weapon, and it's pointed directly at the people who built it.