Most KM or CCaaS or CMS vendors have an AI story. Most of them involve the same underlying product they've always had, with a chatbot bolted on the side.
The pitch sounds a bit like, your agents can now ask questions using natural language and get instant answers. The demo shows a smiling agent typing how do I process a refund and a simple set of steps appears. The content behind the demo is the same five-page PDF the team wrote in 2022, parsed through a retrieval model, and the demo works because the demo question exactly matches the demo document.
In production it doesn't exactly work the same way. Real content is more like a wall of text with mixed topics and rationale tangled throughout. The AI might return the right page but the agent still has to decipher it mid-call. Or worse, it returns a confident summary that drops the compliance clause, and nobody notices until complaints come rolling in.
I've been building and implementing knowledge platforms for regulated industries for fifteen years, and things have remained depressingly consistent. Teams treat AI as the solution and content as plumbing. Then they're surprised when AI inherits every problem the content always had.
The frame I have now, that even fits on a bumper sticker is:
Structured knowledge is the substrate. Adaptive delivery is the surface. AI is the connective tissue, not the solution.
The rest of this piece is the long version of that argument. If you're buying a KM platform, an AI chatbot, building one, or trying to work out why your current one feels broken, those three nouns are the three questions worth asking.
Part 1: The substrate
Think of structured content as the sand under a paved path. Nobody admires it. Nobody takes a selfie with it, and contestants on The Block aren't winning a room because of it. People only want to see the pavers on top, and the people gracefully walking across them. But the sand is what makes everything else possible. Lay it badly and every paver shifts, no matter how beautifully they were set.
Structured knowledge is the content layer underneath everything else, and most organisations don't have it or understand its importance.
When I say structured, I don't mean has headings. I mean atomised and classified. Each piece of writing has one purpose and knows what kind of thing it is. A procedure is only a procedure. A concept is only a concept. A decision table is only a decision table. The step is the unit of delivery, not the document.
A refund procedure in this shape could look like:
- Access CRM.
- Navigate to customer account.
- Open Billing tab.
The Billing screen displays. - Click most recent invoice.
- Is the refund greater than $500?
- If no, continue to step 6.
- If yes, follow Refund approval thresholds
- Continue to step 6.
- Click Refund.
The refund processes.
Each step is a single instruction. No filler. No rationale embedded in the action. The italicised line under step 3 is a result confirmation. The agent sees the Billing screen and knows they're in the right place. They don't have to guess whether they did the step correctly. They recognise it. Step 5 lists the common branch first, most refunds are under the threshold, so the typical agent reads one short clause and keeps going.
The decision itself, whether this refund needs approval, belongs in the procedure, because the agent has to make it. The why behind the $500 threshold (why that number, which policy sets it, who signs off) doesn't. That lives in a separate Concept called Refund approval thresholds, linked from step 5 for anyone who needs the detail. The agent handling a $47 refund takes the No branch and moves on.
This is harder to write than it looks. The authors I work with have spent years getting fluent in it, and new starters still catch themselves slipping back into prose. The instinct is to explain, to protect the reader from confusion, to put the whole picture and every little detail on the page. Structured writing asks you to trust the structure and resist the explanation.
Why does any of this matter for AI? Because AI doesn't fix bad content. It amplifies it.
Retrieval models work by chunking text and matching chunks against a query. A 400-word paragraph with four actions and three explanations returns 400 words of noise around whatever the agent actually needed. A six-word atomic step tagged step 3 of 6, refund procedure returns a precise unit the agent can act on. Same model, different substrate, different result.
Most of the teams spending the most on AI tooling and getting the least from it are running top tier models over the same documents written as walls of text they've had since 2022. The models can't help them. Nothing can, until the substrate is fixed.
Part 2: The surface
If structured content is the sand, delivery is the pavers laid on top. The pavers have to carry everyone who walks across them: a courier running at full pace, a kid on a scooter, someone with a walking frame. Same path, very different loads. The flex happens on the surface, but only because the base beneath is sound enough to support it.
Adaptive delivery is assembly, not retrieval. Same content, different response shape, chosen at the moment of the question based on what the agent needs right now.
How do I complete a refund? doesn't have one correct answer. It has several, depending on who's asking and why. A new starter on day three needs the whole procedure with context. A ten-year veteran needs the one step they've forgotten. An agent mid-call with an impatient customer needs the next action and nothing else. A trainer in a workshop needs the policy behind the procedure. Same content, same question, very different responses.
I've found five patterns cover most of what a well-designed delivery layer needs to do.
The lookup answer. What's the refund approval threshold? One sentence from a Reference document. The hard part is correctly classifying this as a lookup, so the system doesn't silently drag the agent into a full procedure they didn't ask for.
The headline-then-procedure answer. How do I complete a refund? First response is one line, Processed in CRM Billing. Six steps, about a minute. Then a choice: walk me through it, show me everything, just the next step. The headline is a contract. It tells the agent how big the thing is they're about to do, so they can decide whether to start now or put the customer on hold first.
The guided walkthrough. Once the agent commits, the current step is the focus with its result confirmation underneath. Previous steps stay visible but faded as a trail of what's been done; the next step or two sit faintly ahead as a preview. This is one of three common patterns, some systems reveal one step at a time for strict focus, others show the full list with just the current step highlighted. The faded-context middle ground works best for most call-centre work. The agent keeps focus on what they're doing now without losing the shape of the whole flow.
The whisper. The agent didn't ask anything. The system received a trigger from the call that a refund might apply. The procedure appears in a side panel. Nobody interrupts the conversation. The agent engages when/if they want to. This is the pattern that pays back the structuring work the most, because it only works if the system has something coherent to surface. A six-step atomised procedure can be whispered. A seven-page PDF full of prose and screenshots cannot.
The clarifying question. Ambiguous asks, go through a follow-up question. Full refund, partial refund, or goodwill gesture? Three options. The agent picks one and the system takes them to the right flow. The clarifying question is the first row of a decision table, surfaced as dialog. The content type literally generates the interaction.
Underneath all five is one simple design principle, recognition, not search. The agent on a call doesn't have the bandwidth to search. They have the bandwidth to recognise. Every piece of searching the system can do for them is bandwidth returned to the customer, which is the actual goal. Agents aren't paid to be good at prompting and tools. They're paid to be present with the people they're talking to.
Part 3: The connective tissue
If structured content is the sand and adaptive delivery is the pavers, AI is the walker. It carries intent across the surface, from one side to the other. When the base is level and the pavers are laid well, the walk is effortless and the walker barely notices the surface at all. When the base is bad, every step is a trip hazard, and the walker is called clumsy even though the problem was always the path. That is what people mean when they say the AI hallucinated. The AI didn't hallucinate alone. It walked on a path nobody laid properly. It infers intent. It picks the pattern. It assembles the response. It judges when to interrupt and when to wait. It does the work that used to require either a well-trained agent who learned the content over months, or a well-designed interface that got you most of the way but couldn't adapt.
The confusion I keep seeing is treating AI as the silver bullet rather than the connector. We have AI now becomes a pitch and a purchase decision, as if the AI itself solves everything on its own. It doesn't tell me whether the AI is doing anything useful. The useful questions that should be asked are:
- Is the AI determining intent well enough to pick the right delivery pattern for each query?
- Is it assembling responses from classified atomic content, or paraphrasing from paragraphs?
- Is it citing its source so the agent can verify, especially on compliance-sensitive content?
- When it doesn't know, does it ask a clarifying question, confidently hallucinate, or refuse to answer?
- Is it respecting the layers the authors built, surfacing hovers and sub-processes at the moment of need?
An AI that does all of this isn’t a silver bullet, it’s connective tissue, and it only works because the substrate and surface it connects are sound. An AI that paraphrases unstructured PDFs is a liability dressed up as a solution.
Now I know I'm going to be really unpopular here, but AI should not be generating facts. It should not be writing new procedures. It should not be summarising a 500-word regulatory passage into something the agent relies on instead of the original. For a KM platform in a regulated industry, the AI's job is to find, surface, assemble, and ask. The source of truth stays in the structured content, authored by humans who are accountable for it. The AI is a very capable assistant to that content, not a replacement for it.
This is where my view diverges most from the market. A lot of the current AI pitch assumes that if the model is good enough, you can throw any unstructured document at it and the quality problem goes away. That's not true in any industry, and it's especially not true in regulated ones, where the cost of a confidently wrong answer gets measured in fines, lawsuits, and bad outcomes for customers.
The quality of your AI output can only ever be as good as the content it’s reading from.
What this means in practice
For buyers. When you're evaluating KM platforms, the AI demo is the last thing worth looking at, not the first. Ask to see how the vendor's authoring environment enforces structure. Ask what happens when an author tries to write a 400-word paragraph into a procedure field. Ask how the system distinguishes a procedure from a concept, and how the delivery layer behaves differently for each. If the vendor can't answer those questions without pivoting back to the AI story, you're buying plumbing dressed up as intelligence.
For builders. The structured authoring work isn't glamorous and it isn't quick. The research says it's an 18-month program for an established team to get fluent, and it's right. What's changed since the last time anyone wrote that playbook is the author productivity ceiling. LLMs are now genuinely good at the specific transformation that used to take authors days per document. Paragraph prose with embedded rationale, converted into atomic typed content with the rationale moved to its own concept. SMEs can review AI-drafted conversions in a fraction of the time it took them to author from scratch. The economic argument against structuring content at scale, which was always we can't afford the author headcount, no longer holds.
For positioning. The vendors who win the next five years in regulated-industries won't be the ones with the best and most expensive AI. They'll be the ones whose content substrate makes AI genuinely useful. That substrate has to be built. It can't be bought as a feature, and it can't be bolted on after the fact. AI depends on the content. The content depends on the authoring discipline. All of it depends on the operating model your business runs on.
Nobody admires the sand. The people who lay it well don't get famous. The people who skip it don't get caught until the surface starts to shift. That's the political problem with substrate work in general, and it's the reason it's hard to fund internally.
Which brings us back to my new bumper sticker.
Structured knowledge is the substrate. Adaptive delivery is the surface. AI is the connective tissue, not the solution.
Three nouns, in order of what they do. Get the substrate right and the surface becomes possible. Get the surface right and the AI finally has somewhere useful to live. Get those two right and you don't need to oversell the AI, because everything is doing what it is supposed to do. Get the right information in front of the right person at the right moment, without making them pay for it in cognitive load.
Everything else is marketing.
