Knowledge as Infrastructure

One Article, a Thousand Wrong Answers

For most of the last decade, a stale knowledge-base article was a contained problem. An agent might read it, sense something was off, and quietly correct the customer in real time. Human judgement sat between bad content and the customer. That buffer is disappearing. Retrieval-augmented generation now lets AI copilots and customer-facing chatbots pull directly from the same repository, and they repeat what they find with total confidence, no hedging, no second-guessing. An error that once affected one call risks propagating across every conversation the AI touches that day.

This is the uncomfortable discovery many contact centres are making through 2024 and 2025 as they move AI pilots into production. The model is rarely the problem. The knowledge feeding it is. Gartner's widely cited July 2024 forecast — that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, largely due to poor data quality, inadequate risk controls, escalating costs or unclear business value — is not really a statement about artificial intelligence. It is a warning about organisational discipline around content.

The knowledge base is no longer just a support tool; it is becoming a production system. It needs the same rigour as the systems that move money.

Why Governance Became Urgent, Not Aspirational

Knowledge management has always had a governance layer in theory — style guides, review cycles, approval workflows. In practice, most organisations treated it loosely, because the cost of a stale article was small and localised. A customer might get a slightly outdated answer from search; an agent might catch the error before repeating it to a customer.

AI removes that margin for error. When the same article simultaneously trains a copilot's suggested responses, grounds a customer chatbot's answers and populates a self-service search result, its accuracy is no longer a quality-of-life issue. It becomes the determinant of whether three different channels tell the customer three different — or three wrong — things in the same afternoon. Ambiguous ownership, the classic knowledge-base failure mode, now surfaces as an AI hallucination that looks entirely authoritative to the person reading it.

What This Demands Operationally

Treating knowledge as infrastructure means applying infrastructure discipline: named owners for every article, scheduled review dates, version history, and — critically — a gate between content creation and AI ingestion. Some organisations are now running a staging step, where new or edited articles sit in a sandboxed retrieval index and get validated against sample queries before promotion to the live AI-facing corpus. This mirrors how software teams test code before deployment, and it reflects a broader truth: content is now code, in the sense that it executes, at scale, without human review at the point of delivery.

There is also a retirement problem. Most knowledge bases accumulate articles faster than they prune them. An AI system has no innate sense that a 2019 policy document has been superseded twice since; it will retrieve whatever is indexed and present it with the same confidence as this morning's update. Deprecation workflows — actively archiving, not just leaving to rot — become as important as creation workflows. Gartner's forecast is, in effect, a prediction of what happens when organisations skip this step: pilots that looked promising in a controlled demo collapse once real-world data quality, risk exposure and cost realities surface at scale.

The Behavioural Layer: Confidence Without Correction

There is a subtler risk here rooted in how people process confidently delivered information. Behavioural research on the affect heuristic shows that people use the fluency and certainty of a delivery as a proxy for its accuracy — a confidently worded wrong answer is often more persuasive than a hesitant correct one. AI agents deliver every answer, right or wrong, with identical fluency. Customers have no cue to distrust the wrong ones. This removes a check that human interaction used to provide almost for free: the hedge, the pause, the visible uncertainty that once signalled 'verify this before you rely on it'.

What Good Governance Looks Like

Organisations that avoid becoming part of Gartner's forecast statistic tend to share a few practical habits. They assign single, named owners to every article rather than leaving content 'owned by the team'. They set explicit review cadences tied to risk — pricing and compliance content reviewed far more frequently than general how-to guides. They build a validation gate before any updated article reaches the live AI retrieval index, treating content changes with the same caution as a code release. And they actively retire outdated material rather than letting it linger indefinitely in the index, waiting to be surfaced by an AI system with no sense of its own obsolescence. None of this is exotic. It is basic operational hygiene, applied to a layer of the business that quietly became load-bearing before most organisations noticed.