There is a recognisable moment in many South African digital programmes. The new system goes live. The forms are online, the approvals route by email, the documents are stored centrally. And six months later the process takes about as long as it did on paper, the same people complain about the same delays, and the report to the steering committee quietly changes from "transformation" to "adoption challenges".
The system usually works as specified. The problem is what it was specified to do. It took a process designed around paper, signatures and physical movement between offices, and reproduced it on a screen. Every step survived, including the steps that only existed because paper had to travel.
This closing part of the series looks at why that happens, why AI can make it happen faster, and what a better sequence looks like.
Vaalrand Metro and Strelitzia Bank are fictional composites, and their processes and figures are illustrative.
A small example with a large lesson
Vaalrand Metro's leave process, before any digitalisation, had nine steps. An employee completed a form. Their supervisor signed it. The form went to the section head for a second signature. A clerk checked the leave balance in a register. Another clerk checked whether the section had cover. The form returned to the supervisor if cover was missing. Human resources captured it. The payroll office was notified. The original was filed.
The first digital version had nine steps too. The form became an online form, signatures became approval clicks, the register check became a lookup the clerk did in a different system, and the file became a document library. It was faster in the sense that nothing was lost in internal mail. It was not simpler, and the bottlenecks were the same people doing the same checks.
The redesigned version had four. The employee requests leave. Rules check the balance and the cover roster automatically, using data the systems already hold. A supervisor approves, seeing the rule results. The record updates human resources and payroll directly. People still handle the exceptions, such as insufficient cover over a holiday period, but they no longer perform checks a system can do reliably.
None of this required AI. It required someone to ask, for each step, why it existed. The second signature existed because a supervisor had once approved their own leave without oversight; a rule that flags self-approval and unusual patterns addresses that risk better than a signature nobody reads.
Why programmes keep digitising the old process
The pattern is not caused by poor technical work. It has organisational causes that are easier to address once named.
Requirements are gathered from the current process. Business analysts ask users what they do, and users describe the current steps faithfully. Unless someone asks why each step exists, the requirements become a specification for the old process.
Nobody owns the process end to end. In Vaalrand, the leave process crossed four units. Each owned its own step. No one had the authority to remove a step owned by someone else, so every step survived.
Change is scoped as technology. The project charter said "implement a leave management system". It did not say "redesign leave approval", so the people who could have changed roles and rules were never in the room.
Controls are copied, not reconsidered. Every step that looks like a control is kept, because removing a control feels risky. Many of those steps were weak controls even on paper.
AI makes the old mistake faster
The same pattern now repeats with AI, often with more momentum. An organisation takes an existing process and adds an AI assistant at each step: summarise the form, draft the approval note, check the attachment. Each addition is useful in isolation. Together they produce a process with the same number of steps, more technology, more cost and a new set of failure modes.
The better question is the one the leave redesign asked. For each step: why does it exist, what risk or decision does it carry, and is there a simpler way to carry that risk now? Sometimes the answer is an AI assistant. More often it is a rule, a data integration, a changed role or a step removed entirely.
We have made the same argument from the delivery side in Build AI into the business you have and from the discovery side in The process map is usually fiction.
The layers that have to move together
A digitalisation programme succeeds when it moves several layers at once. Changing only the technology layer is what produces nine digital steps.
Strategy and outcomes. What the change is for, expressed as an outcome someone can measure, such as approval time or error rate, not as a system going live.
Process. The work, redesigned before it is automated. This is where steps are removed, combined or replaced with rules.
People and roles. Who does what in the new process. When the clerk no longer checks balances, what do they do instead? Unanswered, this question turns into quiet resistance.
Data. Shared definitions and accountable owners for the data the process depends on. The leave rules only work if the cover roster is accurate and someone is accountable for keeping it so.
Technology and AI. The systems that carry the new process, including any AI that genuinely improves a step.
Governance. How decisions are made, evidenced and reviewed, and how benefits are tracked after go-live rather than assumed.
A practical sequence
For any process proposed for digitalisation or AI, a short sequence avoids most of the trap.
- Name an end-to-end owner with authority across every unit the process touches.
- Map the real process including workarounds, and record why each step exists.
- Classify each step: decision, control, information transfer, or legacy of paper.
- Redesign by removing paper legacies, replacing weak controls with stronger rules, and keeping human judgement for genuine decisions and exceptions.
- Fix the data the redesigned process depends on, with named owners.
- Then choose technology, including AI where it improves a decision or removes effort the rules cannot.
- Measure the outcome you defined at the start, and review it after go-live.
The sequence looks slower than buying a system. In practice it is faster, because it avoids the second programme that is usually needed to fix the first.
What the series adds up to
The four parts of this series have looked at enterprise AI from different directions: a bank whose AI programme rested on reliability work, a service assistant whose value depended on restraint, a procurement reform that treated the platform as the fourth decision, and a leave process that showed what digitalisation actually means. The common thread is that useful AI in South African organisations is downstream of ordinary, disciplined work on foundations, processes and accountability.
That is not a reason to delay AI. It is a reason to aim it at redesigned work rather than at the old work with more technology on top.
This is the final part. Start again at Part 1, Reliability comes before AI, or see all parts on the series page.
How CloudNala can help
We help organisations redesign processes before automating them: mapping the real process, classifying each step, designing the new roles, rules and data ownership, and only then selecting technology and AI. The output is a redesigned process with an owner, a measurable outcome and a delivery plan that addresses all six layers rather than only the system.
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