By Ayowole Delegan

Artificial intelligence has generated no shortage of predictions about the future of work.
Some believe it will dramatically improve productivity. Others expect it to eliminate large numbers of jobs. There are forecasts of autonomous companies, digital employees and intelligent systems capable of performing much of today’s knowledge work with limited human involvement.

These questions matter. But they may also be distracting businesses from a more immediate and commercially useful question:

How can organisations combine AI with the capabilities they already possess?
The most valuable applications of AI may not come from using a general-purpose model to perform isolated tasks. They may come from embedding those models within existing operations and surrounding them with proprietary data, institutional knowledge, established workflows and human judgment.

If that is correct, access to AI will not, by itself, become a lasting competitive advantage. Access will become increasingly widespread. The advantage will belong to organisations that can teach AI enough about their businesses to make it genuinely useful.

From General Intelligence to Organisational Context
Frontier AI models are already capable of writing, analysing documents, producing software code, interpreting images and supporting increasingly complex reasoning tasks. Those capabilities will continue to improve.

But the leading models are also becoming more closely matched. Stanford University’s 2026 AI Index observed that frontier model performance had become increasingly concentrated within a narrow band at the top of widely used rankings. This does not mean that progress in artificial intelligence has stopped. It means that access to powerful baseline intelligence is likely to become less differentiated over time.

When several companies can provide models with comparable capabilities, the strategic question shifts. The important issue is no longer simply which organisation has access to AI. It becomes what each organisation is able to build around it. A general-purpose model may understand how businesses operate in principle. It does not automatically understand a company’s customers, commercial definitions, reporting logic, risk appetite, operating history or decision-making culture.

That context must come from the organisation. This is why the next phase of AI adoption may be less about prompting models and more about connecting them to the systems through which companies already create value.

Building an AI Associate
I encountered this distinction while trying to improve the way I prepared a monthly business report. I wanted an AI agent that could support the analysis and produce the report. But simply asking a general-purpose model to write it was not enough. The model could generate fluent language, but fluency was not the real requirement.
The report needed to use the correct data. It needed to understand the commercial meaning behind the numbers. It also needed to reflect how I structured an argument and distinguished an important business movement from ordinary variation. There was an additional risk. If the model did not have reliable access to the appropriate information, it could confidently introduce inaccurate figures or unsupported explanations.

The solution was not a better prompt alone. I built a data pipeline that supplied the agent with the relevant business information. I used semantic questions to provide context about the business and teach it how I approached the analysis. I added samples of my previous reports so that it could follow my writing logic, structure and reasoning.

Then came the most important part: iteration. Every week and every month, I used the agent on actual analytical and writing assignments. I reviewed its output, corrected its interpretation and refined the context available to it. Over time, it became better at recognising what mattered and presenting it in the way the business required.
Today, it prepares much of the report, with relatively minor adjustments from me. The model did not become valuable simply because it could write. It became valuable because it was connected to accurate data, business context, accumulated examples and a continuing feedback process.

I did not replace myself with AI. I built something closer to an associate—one that could handle a repeatable part of my work while allowing me to concentrate more attention on interpretation, challenge and decision-making. That distinction is important. Many businesses are currently evaluating AI as though it were a finished employee that can be placed into an organisation and expected to understand how everything works.
A more productive approach may be to treat it as a capable new associate that still needs access, context, examples, supervision and feedback.

When Internal Knowledge Meets Frontier Models
I saw the same principle in a more specialised setting in another role.
Working with a team of analysts, I was involved in developing an AI-enabled clinical decision-support agent using medical data. The system received information about a patient’s symptoms and produced possible explanations, expressed probabilistically. It could also recommend corresponding tests and provide credible reference material explaining the basis of its assessment.
The team did not attempt to create a new frontier model from the beginning. We combined the capabilities of existing models with specialised medical data, analytical design and a structured validation process. In internal testing, the system achieved accuracy levels ranging from approximately 80 to 90 percent, depending on the test conditions and task evaluated.
That result did not make the system a replacement for a doctor. Medical decisions carry consequences that require clinical judgment, appropriate testing, governance and human accountability. A probabilistic output is not the same as a confirmed diagnosis.
But the exercise demonstrated something commercially significant. The underlying model supplied broad reasoning capability. The organisation supplied the specialised data, domain knowledge, validation logic and intended workflow. The practical value came from combining both.
The same pattern can apply in many industries. A bank can connect AI to its transaction history, risk controls and customer-service processes. A manufacturer can combine it with maintenance records, production data and the experience of plant engineers. A distributor can embed it within inventory planning, route economics and customer ordering patterns. A law firm can connect it to precedents, internal opinions and review standards. In each case, the general model may be available to competitors. The organisational knowledge surrounding it is not.

AI Will Reward Existing Capabilities
Research on business technology adoption already points in this direction.
An OECD study of firms across 11 countries found that companies using AI tended to be more productive, particularly among larger firms. But it also found that complementary assets—including digital skills, infrastructure and the use of other technologies—played a critical role in those productivity advantages.
This is consistent with the history of previous general-purpose technologies. Buying technology is rarely the same thing as capturing its full economic value. Businesses often need to redesign processes, develop skills, improve data and change how decisions are made before productivity gains become visible. Early evidence from generative AI presents a similar picture. In one field experiment involving 7,137 knowledge workers across 66 firms, employees given access to an AI tool integrated into applications they already used spent less time processing email. Yet the researchers did not find a broad change in the overall composition or quantity of tasks performed.
Access created some efficiencies. It did not automatically transform the organisation. This suggests that the AI productivity question cannot be reduced to how many employees have subscriptions to an AI platform. Nor is adoption adequately measured by how frequently workers use chatbots.
The more consequential questions are operational:
Is the AI connected to reliable internal data? Does it understand the organisation’s definitions and commercial context? Is it embedded within a real workflow? Can employees evaluate its output? Is there a feedback process through which the system improves? Are there safeguards for decisions where mistakes carry material consequences?
Without these complementary capabilities, organisations may produce more content without improving decisions. They may automate activities without addressing whether those activities create value. They may also generate answers faster while remaining uncertain about whether those answers are correct. That is not necessarily productivity. It may simply be faster activity.

The Internet Offers a Useful Parallel
The Internet eventually became available to almost every serious business. But equal access did not produce equal outcomes. Some companies used it mainly for communication and visibility. Others redesigned distribution, payments, customer acquisition, supply chains and entire business models around it.
The difference was not possession of the internet. It was what firms combined with it.
AI may follow a similar path. Frontier intelligence will increasingly be available through multiple providers, embedded into common software and offered at declining costs. Powerful models may eventually become infrastructure: essential, widely accessible and insufficient on their own to create differentiation. The firms that benefit most will be those that possess what the models do not—proprietary information, operating experience, customer relationships, specialised expertise and a clear understanding of how work actually gets done.
This creates a different kind of AI race. Companies may be tempted to focus on model selection: which platform is more intelligent, which one produces the best answers or which new release leads the latest benchmark. Those questions have value, but the answers may change quickly.
A more durable investment is to make the organisation ready for whichever models become strongest.

That means improving data quality, documenting institutional knowledge, creating secure connections between models and internal systems, redesigning important workflows and developing employees who can supervise and challenge AI-supported work.
It also means identifying areas where the organisation already has meaningful expertise. AI is most likely to strengthen an advantage when there is an advantage available to strengthen.

The Capability Around the Model
The debate about whether AI will replace people will continue. Some tasks will almost certainly be automated, some roles will change, and organisations will make different choices about how work is divided between humans and machines.
But replacement may be an incomplete way to understand the larger opportunity.
The more immediate value of AI lies in amplification: enabling an analyst to examine more information, helping a doctor assess possibilities more systematically, allowing a manager to prepare recurring reports faster, or helping a company apply its knowledge more consistently across the organisation.
That amplification will not occur simply because a model has been purchased. AI needs the right data. It needs business context. It needs to be placed within an operating process. It needs human expertise capable of assessing when its conclusions are useful, incomplete or wrong.
Over time, the leading frontier models may continue to converge in capability. But the organisations using them will not converge in their capacity to create value. Some will have better information. Some will have deeper institutional knowledge. Some will build stronger feedback systems. Some will understand precisely where AI should support judgment and where human accountability must remain decisive.
AI may become as widely available as the internet. The competitive advantage will not be access to intelligence alone. It will be the organisational capability built around it.