The institution is an operating system
Every institution already runs a programme. Most of it is written in habit, not code. Redesign that first, or the technology you buy will simply run the old programme faster.
Walk into any bank, insurer, municipality or ministry and ask to see its operating system. You will be shown a technology estate: core platforms, a data warehouse, a cloud migration plan, perhaps an AI pilot or two. That is the hardware. The operating system is something else. It is the set of rules, mostly unwritten, that decides who may decide what, which information travels to whom, what gets escalated and what gets quietly absorbed, and what a person is rewarded for doing on an ordinary Tuesday.
That operating system is the real institution. The strategy document describes what it intends. The operating system determines what it does.
The programme runs whether you see it or not
Institutions accumulate their operating systems the way old cities accumulate streets. A process was designed for a regulation that has since changed. An approval step was added after a single bad loss and never removed. A spreadsheet became the system of record because the system of record could not be changed in time. Each decision was reasonable when it was made. Together they form a programme that nobody wrote and nobody owns, yet everybody executes.
This matters now because of what we are asking technology to do. For decades, institutions bought systems to record and to process. The new generation of tools, AI above all, is bought to recommend, to decide and to act. A tool that acts inside an institution inherits that institution's operating system. If decision rights are muddled, the tool will be muddled. If information is hoarded, the model will learn from fragments. If nobody owns the outcome, nobody will own the machine.
Why the value does not arrive
The evidence is no longer anecdotal. BCG's 2025 research on AI value found that 60% of companies report little or no value from AI, and only 5% achieve value at scale. It is tempting to read those numbers as a verdict on the technology. Read them instead as a verdict on operating systems. The same models are available to the 5% and to the 60%. What differs is the institution the models are dropped into.
Gartner expects that over 40% of agentic AI projects will be cancelled by the end of 2027. Again, the instinct is to blame the agents. But an agent is a new kind of worker. Hire a brilliant worker into an organisation where nobody can say who approves what, where the data is scattered across a dozen inboxes, and where success has never been defined, and that worker will fail too. We would not call it a talent problem. We should not call this a technology problem.
The same models are available to everyone. What differs is the institution they are dropped into.
Design the operating system on purpose
If the institution is an operating system, then transformation is not a procurement exercise. It is a design exercise, and it begins with questions that sound almost too simple to ask:
- What decisions does this institution exist to make well?
- Who holds each of those decisions today, and who should?
- What information does each decision need, and where does it actually come from?
- How will we know a decision was good, and who is told when it was not?
These questions are not technical. They are architectural. Answer them, and the technology choices become clearer and more modest. Skip them, and no platform, however capable, will rescue the result.
Three disciplines
The first discipline is to make the invisible visible. Map decisions, not departments. An organisation chart shows reporting lines; it says almost nothing about how judgement flows. A decision map shows where value is created or lost, and it exposes the approvals that exist only because someone, once, was burned.
The second discipline is to separate what must be human from what may be automated. Not everything a person does today deserves a person tomorrow. But some things do, permanently: judgements that carry moral weight, decisions that change a life, moments when the institution's legitimacy is on the line. Designing those in, deliberately, is what distinguishes an institution that uses machines from one that is used by them.
The third discipline is to version the operating system. No engineer would ship code without a change log. Institutions change their decision rules constantly with no record at all. An institution that can say what changed, when and why, is an institution that can learn. It is also one that a regulator, a board or a citizen can hold to account.
The Southern African opening
Southern African institutions are often described as behind. In one sense the opposite is true. Many have not yet hard-coded their operating systems into decades of legacy automation. They have the chance to design the operating system first and choose the technology second, rather than spending the next decade untangling the reverse.
That is an advantage only if it is used with intent. Operating models imported wholesale from very different markets carry assumptions about labour, regulation, infrastructure and trust that do not hold here. The better course is to design for the institution as it actually is, serving the people it actually serves, under the law that actually applies.
What we imagine
Imagine an institution whose decision rights are as legible as its balance sheet. Where every automated step has a named human owner. Where a new tool is judged first by one question: which decision does this improve, and how will we know? Where change is versioned, explained and reversible.
None of this needs a technology that does not yet exist. It needs the will to treat the institution itself as the thing being designed.
That is the art of the possible as we understand it. Not imagining machines that will save us, but imagining institutions worthy of the machines we already have.