The AI scaling problem nobody talks about
- Rennie Naidoo
AI pilots often succeed. Scaling them is where technology collides with organisational reality.
Every week brings news of another successful AI pilot. A chatbot reduces call centre volumes. A predictive model identifies customers at risk of leaving. A generative AI assistant helps employees draft reports in a fraction of the usual time.
The demonstrations are impressive. The metrics look promising. Executives are reassured that the future is arriving on schedule.
Then, more often than many would like to admit, very little changes.
Months later, the pilot remains confined to a single department. The wider organisation continues to operate much as it did before. Productivity gains prove elusive. Employees return to familiar routines. Managers keep existing approval structures in place. The technology that seemed poised to transform the business becomes another item on a growing list of unrealised ambitions.
This raises a question that receives too little attention amid the excitement surrounding artificial intelligence.
If AI is as transformative as its advocates claim, why does scaling it prove so difficult?
When technology meets institutional reality
The standard explanation is technical. Organisations need better models, cleaner data, stronger governance, more computing power or deeper AI skills. These factors matter. A poorly designed system will not scale simply because leadership wants it to.
But the deeper challenge is that organisations are not machines. They are human institutions.
This distinction is easy to overlook. Modern management has long been attracted to the idea that organisations can be engineered like technical systems. Identify an inefficiency. Apply a tool. Measure the outcome. Repeat the process.
Software seemed to support this belief. Many digital tools could be deployed widely and quickly. Email, collaboration platforms, cloud applications and enterprise systems created the impression that technology could be rolled out across an organisation as if the organisation were a neutral surface.
AI appears, at first glance, to follow the same pattern.
But AI is different because it does not simply automate a task. It enters decision-making, judgement, knowledge work and coordination. It touches the places where organisations are most human.
That is where the scaling problem begins.
The protected world of the pilot
A pilot project exists in a protected environment.
Objectives are clearly defined. Participants are often carefully selected. Senior sponsors are attentive. Exceptions can be managed. Complexity is temporarily reduced so that the technology can prove itself.
That is why pilots so often work.
Scaling requires the return of complexity.
The model that performs well in one division encounters different workflows in another. Data that looked clean in a controlled setting reveals inconsistencies across the enterprise. Governance questions emerge. Legal, compliance and security teams ask reasonable questions. Managers defend established processes. Employees develop workarounds. Business units interpret the same recommendation differently.
The technology may remain largely unchanged. The surrounding organisation does not.
This is why many discussions about AI scaling feel incomplete. They focus on the intelligence of the system while paying less attention to the adaptability of the institution expected to use it.
A model can generate recommendations in seconds, but an organisation may still require weeks to approve them. An AI assistant can summarise information instantly, but employees must still decide whether to trust it. A predictive system can identify an opportunity with impressive accuracy, but business units must still coordinate their actions to realise value.
The bottleneck shifts.
It moves away from computation and toward coordination.
The human system beneath the technical system
This is not a new problem. It is an enduring human problem appearing in technical form.
Organisations are built from habits, incentives, loyalties, hierarchies, professional identities and informal rules. People do not simply adopt a tool because it is available. They interpret what the tool means. They ask, often silently, what it changes about their work, their judgement, their relationships and their place within the organisation.
Will this system help me do better work, or will it expose my mistakes?
Will it support my team, or weaken our influence?
Can I trust its output?
Who is accountable if it is wrong?
Does using it make my work easier, or does it create another layer of monitoring?
These are not irrational objections. They are reasonable human responses to uncertainty.
People are sensitive to trust, fairness, reputation and risk. They rely on familiar routines because constant reinvention is exhausting. They protect their professional standing because recognition and credibility matter. They prefer known processes not because they lack imagination, but because predictability provides a sense of safety.
AI does not remove these concerns. It often brings them to the surface.
This is what many AI strategies miss. Enterprise data is not just operational data. It is also a record of how people work, decide, trust, avoid risk and coordinate.
The system may appear technical. The patterns inside it are human.
Why technical success is not organisational success
Many organisations still measure AI progress through technical indicators. Model accuracy improves. Response quality increases. Processing costs decline. Adoption dashboards show usage.
These are meaningful signals, but they can obscure more important questions.
Has decision-making improved? Have workflows been redesigned? Have employees developed new capabilities? Have teams learnt to coordinate around the system? Has the institution become more effective?
Without affirmative answers to these questions, AI risks becoming another powerful technology that demonstrates remarkable potential while producing modest results in practice.
History offers a useful warning. Computers did not automatically revolutionise office work. The internet did not automatically flatten hierarchies. Big data did not automatically make decisions scientific. Digital transformation did not automatically make organisations adaptive.
Some promises were fulfilled. Many were diluted by the stubborn reality of institutions.
The reason was rarely that the technologies lacked capability. It was that organisations struggled to reorganise themselves around those capabilities.
Railways transformed economies not because locomotives existed, but because societies built supply chains, regulations, timetables and business models around them. Electrification reshaped industry not merely because electricity was available, but because factories eventually redesigned how work was organised.
The technology mattered. The surrounding systems mattered more.
AI may be following a similar path.
Technology changes faster than institutions
The prevailing narrative suggests that increasingly capable AI will inevitably drive organisational transformation. But technological capability and organisational change do not move at the same speed.
Progress in one does not guarantee progress in the other.
This matters for CIOs because scaling AI is not simply a matter of moving from prototype to production. It is a matter of moving from technical demonstration to human adoption. That requires new workflows, new accountabilities, new skills and new forms of trust.
It also requires honesty about what AI is being asked to do.
In many cases, AI is not being asked merely to improve efficiency. It is being asked to alter how people decide, collaborate, share knowledge and exercise judgement. These are not only software deployment problems. They are institutional change problems.
The future of AI will certainly depend on advances in algorithms, infrastructure and data. But it will also depend on something less glamorous and far more difficult: the capacity of institutions to change.
That capacity is not created by a pilot. It is built through trust, redesign, patience and a clearer understanding of the people inside the system.
AI will scale when organisations stop treating adoption as a technical rollout and start treating it as a human transition.
The technology may be new.
The human problem is an old and enduring one.
- Rennie Naidoo is an IS professor and research director at the Wits School of Business Sciences. An established NRF-rated researcher, his focus areas include data science, sustainable IT, artificial intelligence and cyber security. This article was first published on CIO South Africa.