Public Sector27 August 20267 min read

No AI Without Information Architecture — Part 10 of 10

From one form to a data strategy

Nine articles, one motion. The sales sequence and the architectural dependency order turn out to be the same order — which is why they do not have to be traded off against each other.

#Public Sector#AI Strategy#Solution Architecture#AI Governance

This series started with a photograph of a spreadsheet and has worked its way down to the foundation underneath an AI programme. What makes the whole thing cohere is something that took us a while to notice:

The order you have to sell this in is the same order you have to build it in.

That is unusual. Normally commercial sequencing and technical sequencing pull against each other — the client wants the visible thing, the architecture wants the foundation. Here they align, and it is the most useful thing in this series.

The motion, in the order it actually runsLANDFind a small visible problemDigitise one bounded formPROVEStructured data starts arrivingA prototype shows the valueSPONSORA business owner wants itDiscovery is now fundableFOUNDGovernance, semantics, ownershipThen the platform designCLIMBDescriptive to prescriptiveAI only once the base is trustedPRESENT FROM STEP ONE, EVEN WHEN IT IS NOT THE SALES MESSAGEGovernance and ownershipAccountable human oversightThe application is the first layer of the opportunity, not the finished value
Read left to right it is a sales sequence. Read as architecture it is the same dependency order the platform needs anyway — which is why the two do not have to be traded off against each other.

The motion

Land. Find a small, visible operational problem. Digitise one bounded process or form. Not the platform — the form.

Prove. Structured data starts arriving. Build something from it that shows the organisation a view it did not have — using synthetic data first if real data is still months away.

Sponsor. Someone who owns the pain now wants more. That is when discovery becomes fundable, and it needs to be the right someone.

Found. Now the conversation about governance, semantics and ownership is not only possible but wanted, because they want to rely on what they are seeing. Then the platform design.

Climb. Descriptive to diagnostic to predictive to prescriptive. AI only once the base underneath is trustworthy — because there is no reliable AI without information architecture.

Read left to right it is a sales sequence. Read as architecture it is the dependency order the platform needs anyway. You cannot govern data you have not collected. You cannot build trustworthy analytics without definitions. You cannot deploy operational AI on top of numbers nobody agrees on.

The principle underneath

If there is one line to keep from all of this:

The application is the first layer of the opportunity, not the finished value proposition.

An application is where most engagements start and where most of them stop. It gets delivered, it works, everyone is satisfied, and the data it generates accumulates unused — because turning that data into something the organisation acts on was never anybody's scope.

Every digitised process is a collection point. Every collection point is the start of a view of an operation nobody currently has. Recognising that at the point of building the application, rather than two years later, is most of what separates a supplier from a partner.

The things that stay true after a good demo

The risk in a series like this is that the motion reads as a formula. It is not, and these are the cautions that keep it honest.

Six things that stay true after a good demoA prototype is not production architectureA connection to data is not a guarantee of secure or truthful answersA lakehouse does not create governance by existingNatural-language analytics still needs access control, auditing and cost limitsConsequential public decisions need an accountable human in the loopMonetising data never means selling personal or citizen dataGovernance is present from step one, even when it is not the opening message
Each of these is easy to agree with in principle and easy to skip in the fortnight after a client says yes. They are written down for the same reason guardrails are written down anywhere: so that agreeing to them is not the same as remembering them.

Each is easy to agree with and easy to skip in the fortnight after a client says yes — which is exactly when the pressure to skip them is highest.

Two are worth expanding.

Consequential public decisions need an accountable human in the loop. When these systems inform budget allocation, service prioritisation or anything touching citizens, the requirement is not that the AI be accurate. It is that a person is accountable for the decision and can explain the basis for it. An organisation that cannot say who decided something has a governance failure regardless of how good the model was.

Monetising data never means selling personal or citizen data. "Data monetisation" is a phrase that gets used loosely and it is genuinely dangerous in a public-sector context. There is real value in aggregate, anonymised insight, and in an institution using its own data to run better. There is a bright line between that and treating information collected from citizens under a statutory function as a commercial asset. Anyone doing this work in South Africa should be able to state where that line is, unprompted, and should be suspicious of any conversation that gets vague about it.

What is genuinely different now

Two things have changed, and they cut in opposite directions.

The demo got cheap. Synthetic data and modern tooling mean a convincing prototype takes an afternoon instead of a quarter. That is a real shift in how these conversations can be opened, and it is the single most practical change in this series.

The foundation got more important, not less. When output was a dashboard a human read, a definitional inconsistency produced an argument in a meeting. When output is an agent answering in confident prose at machine speed, the same inconsistency produces a wrong answer that nobody catches, delivered fluently, to someone who acts on it. The semantic layer has gone from good practice to control.

So the gap between what can be demonstrated and what can be operated has widened. Anyone can now show something impressive. Far fewer can put it into an institution and have it hold. That gap is where the actual work is.

Where to start on Monday

If any of this is recognisable in your organisation:

Name three decisions you would make better with data you do not currently have. Specific decisions, with the person who makes each one.

Find the process that would produce the data for the first of them, and ask what it would take to digitise just that.

Write down ten definitions — the terms that appear most in your executive reporting, with their edge cases and authoritative source. It takes an afternoon and it is the beginning of your semantic layer.

Ask the hosting question before any architecture is drawn, and get it in front of legal and risk rather than resolving it technically.

None of those is a programme. All of them are things a competent team can do inside a month, and together they put you meaningfully further up the ladder than most organisations that have been talking about an AI strategy for a year.

How CloudNala can help

We work with organisations at both ends of this: the ones who have an application generating data nobody uses, and the ones who want AI and have not yet been told what has to be true underneath it. The useful early work is small and unglamorous — naming the decisions, choosing the first process, writing the definitions, settling the hosting question. Done at the start it is inexpensive. Done after a stalled AI programme, considerably less so.


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