Never don't invest in the content base

Knowledge and help content has been the easiest thing to underfund in customer operations for as long as I've worked in it. It is a cost line with no obvious revenue attached. Its impact shows up as an absence — the contacts that didn't happen, the handle time that didn't grow. When budgets tighten, the content team is small enough to cut and quiet enough that nobody notices for two quarters.

I've made this argument in various forms for a decade and mostly lost it. What's changed is that the argument no longer needs me. The economics did it instead.

The substrate stopped being optional

An AI service system is a machine for turning what your organisation knows into answers. It is very good at the turning. It has no ability whatsoever to supply the knowing.

When help content was primarily consumed by humans, gaps were survivable. A customer who couldn't find an article contacted support, and an agent filled the gap from experience. The undocumented knowledge lived in people's heads and in team chat, and it worked, expensively, because a human sat between the gap and the customer.

Remove the human and the gap is load-bearing. The model reaches for something that isn't there. Depending on how it's built, it either declines to answer, which is honest and unsatisfying, or it assembles something plausible from adjacent material, which is much worse. Either way the failure traces directly back to an article somebody didn't write in 2019.

Content debt behaves exactly like technical debt, except almost nobody tracks it and there's rarely an owner.

Bad content is worse than no content

The uncomfortable discovery for most organisations turning on retrieval is not that they have too little content. It's that they have too much, and a meaningful share of it is wrong.

Every mature help centre carries sediment. Articles describing flows that were redesigned two releases ago. Regional policies that were harmonised but never merged. Three overlapping documents on the same topic, published years apart, each partly correct. Internal macros contradicting the public article they were derived from.

Humans navigate this without difficulty because they carry a recency filter the content doesn't. An experienced agent glances at a 2021 article and knows to ignore half of it. Retrieval has no such instinct. It surfaces the confidently-worded obsolete page, and the model, being cooperative, presents it as current.

So the first phase of most AI service programmes isn't building. It's an audit that nobody scoped, that takes longer than the model work, and that surfaces the accumulated evasions of years. Teams that funded content maintenance all along skip most of this. Teams that didn't pay it all at once, under deadline pressure, which is the most expensive possible time.

What good looks like now

The bar has moved, and it hasn't moved toward more polish. It's moved toward structure and truth.

• Write for a reader who needs the conditions. Human readers infer scope from where they found an article. A retrieval system does not. Content needs to state which product, market, plan and version it applies to, inside the text.
• Say what the boundary is. The most valuable sentence in many articles is the one stating what the process does not cover and what to do instead. It's also the sentence most often cut for brevity.
• Cover the ugly cases. The topics nobody wants to publish on — the workarounds, the known limitations, the awkward edge conditions — are exactly the ones customers escalate about. They're where the gap between your content base and your competitor's is widest.
• Delete aggressively. One accurate article beats four partially-correct ones, and the maintenance cost of the four is what stops anyone maintaining the one.
• Treat every escalation as a content signal. A question that had to reach a human is a fact about your knowledge base. Most organisations collect that fact and never route it anywhere.

The part I find genuinely satisfying

For most of my career, the people writing help content were the furthest thing from the strategy conversation. They were treated as a documentation function: necessary, downstream, invited afterwards.

That was always a misreading, and the technology has now made the misreading expensive. The quality of the content base sets the ceiling on the quality of every automated interaction on top of it. You can improve models, prompts, retrieval and orchestration indefinitely and you will still be bounded by whether the thing the customer needs to know exists somewhere in a form a machine can find and trust.

Which means content strategy is not a support function of the AI programme. It is the AI programme, or at least the half of it that determines whether the other half was worth doing.

My advice to anyone standing up this work: before you scope the platform, count how many people are responsible for the accuracy of what your company tells customers, and check whether that number went up or down in the last three years. That number predicts more about your outcome than your choice of model will.