A lot of organisations are going to be disappointed by AI over the next two years, and it will have very little to do with the models.
The pattern is consistent enough to predict. An organisation buys or builds something impressive. It works in the demo. It reaches a handful of real users, who ask real questions, and the answers are subtly wrong — not obviously broken, plausibly wrong. Trust drains in about a month and the initiative quietly becomes a pilot that never concluded.
What failed was underneath.
The rungs
Rung one — collection. Whether the data exists in a form anything can read. Structured records and unstructured documents; data sitting in systems and data moving through them. Most organisations have far less genuinely usable data than their systems inventory suggests, because a great deal of it is trapped in formats, spreadsheets, and processes that were never designed to be read by anything but a person.
Rung two — governance. Definitions, lineage, catalogue, ownership, quality, access. This is the rung that gets skipped, and it is the one that determines whether an answer can be trusted. More on why in a moment.
Rung three — descriptive and diagnostic. What happened, and why. Ordinary reporting, done properly.
Rung four — predictive and prescriptive. What is likely to happen, and what to do about it. Worth separating these two: prediction tells you a facility will be over capacity by March; prescription tells you which lever to pull. The second is often the more valuable and is frequently an optimisation problem rather than a machine-learning one — you do not always need a model to get there.
Rung five — operational AI. Systems that act, not just answer.
Why rung two is the one that matters
Rung two is unglamorous, has no demo, and is the hardest thing to get funded. It is also where trustworthiness comes from, and here is the mechanism.
When a person queries a database, they carry context in their head. They know that "completed" excludes the ones pending verification. They know the Northern region was redrawn in April. They know one of the source systems double-counts and they mentally correct for it. That context is real institutional knowledge, and it lives in people.
When an AI agent queries the same database, it has none of that. It has table names and column names. It will produce an answer — confidently, immediately, in fluent prose — that is wrong in exactly the ways the human would have corrected for.
The layer that encodes that context so a machine can read it is the semantic layer, and it is the subject of the next article. It is the difference between a system that answers and a system that answers correctly.
This is what people mean, or should mean, by no AI without information architecture. Not that governance is virtuous. That without it, the thing you build produces confident output nobody can defend, and it is the confidence that makes it dangerous.
Both kinds of collection
One distinction that gets missed at rung one, and it constrains everything above.
Almost every data programme is designed for the first row. Sources are inventoried, extracts are scheduled, a warehouse fills up. Then somebody asks a question about behaviour — why do citizens abandon this application halfway through, where exactly do they stop, does it differ by channel — and the answer does not exist. Not because the analysis is hard. Because the events were never captured as they happened.
You cannot reconstruct data in transit after the fact. If it was not recorded in the moment, it is gone.
This matters more as services move online, because the most valuable operational insight is often in the drop-off rather than the completion. Worth asking at design time, of any new service: what do we want to know about how people use this, and are we capturing it?
Where AI actually sits
A distinction worth keeping straight, because it changes what you need underneath.
Productivity AI helps individuals work faster — summarising, drafting, searching. It sits on the individual, needs relatively little organisational data foundation, and delivers real value quickly. Most organisations should just do this.
Operational AI puts AI inside a business process, where it reads organisational data, produces something the organisation acts on, and possibly acts itself. This needs every rung below it, because it is making claims about the business rather than helping one person write an email.
Conflating the two is a common and expensive mistake. Productivity AI going well is not evidence that operational AI will. They rest on completely different foundations — one on a licence, the other on rungs one and two.
What to actually do about it
The awkward part: you cannot sell rung two directly, and I would not advise trying. "Let us do a data governance engagement" is close to unsellable, and it should be — it asks an organisation to fund something with no visible output.
What works is the sequence the rest of this series describes. Land on something small and real. Produce structured data. Show what becomes visible. And when the client asks what it would take to rely on this — which they do, once they want to rely on it — that is when the governance conversation is not only possible but wanted. It arrives as "can we put guardrails around this", which is the same work under a name that makes sense to them.
The ladder is a description of dependency, not a project plan. Nobody funds rungs. They fund outcomes, and the rungs get built underneath because the outcome needs them.
How CloudNala can help
Where we are most often useful is in the honest assessment of which rung an organisation is actually standing on, which is frequently a rung or two below where the strategy assumes. That is a short piece of work and it changes the plan considerably — usually by replacing an AI programme that would have disappointed with a sequence that gets to the same place and holds.
Work with CloudNala
CloudNala helps organisations move from technology ambition to practical execution across cloud, AI, data, platform engineering and digital services.
Whether you are exploring AI, modernising your cloud environment, building a public-sector digital service, or turning an idea into a working MVP, we can help you shape the roadmap and deliver the next step.
Book an AI Readiness Workshop or write to us at consult@cloudnala.co.za