AI is making intelligence cheap. What comes next?

For much of the past two years, the investment debate around artificial intelligence has centred on intelligence itself. Which company has the smartest model? How quickly are those models improving? And what happens to software companies when artificial intelligence can perform much of the work their products were designed to facilitate? These are reasonable questions. But they may increasingly be the wrong ones.

The cost of intelligence is falling rapidly. Capable models are proliferating, open-weight alternatives are improving, and businesses can increasingly choose between models or use several at once. Intelligence is becoming cheaper, more accessible and more interchangeable. This does not diminish the importance of AI. It changes where value is likely to accrue. As intelligence becomes abundant, the bottleneck shifts from producing work to determining what work should be done, whether it is correct and how it fits within a broader organisation. The scarce resource is no longer generation. It is judgment.

As investor Clifford Sosin recently observed, “Coming up with ideas was never the hard part. The hard part is how fast reality answers them.”

AI can generate more ideas, designs, code and analysis than any organisation could possibly pursue. But it cannot eliminate the need to choose which ideas deserve to be tested, expose them to reality and learn from the result. In fact, by making generation almost free, AI makes judgment, experimentation and feedback more important, not less.

From processing information to exercising judgment

I have experienced this shift personally.

As an investor, I once faced a practical limit on how much information I could process. I could read perhaps 25 annual reports in a week. AI now allows me to interrogate hundreds of reports, earnings calls and industry documents over the same period. However, the bottleneck is no longer access to information or even the ability to summarise it. The bottleneck is deciding what matters. Which management teams should be trusted? Which competitive advantages are real? Which apparent risks are temporary, and which ones threaten the underlying economics of a business? What information contradicts the prevailing investment thesis?

AI can assist with each question, but it cannot remove the need for judgment. In many respects, it makes judgment more important. Producing 20 plausible answers is easy. Choosing the right one and accepting responsibility for that decision however is not. The same shift is occurring across the economy.

A lawyer can generate a contract in seconds, but someone must determine whether it protects the client. A software developer can generate thousands of lines of code, but someone must decide whether the product solves the right problem. A designer can produce dozens of concepts, but an organisation still has to agree on which design serves its customers and brand. AI dramatically increases the supply of potential output. It does not automatically increase an organisation’s ability to coordinate, evaluate and act upon that output.

Figma and the changing bottleneck in design

Figma provides a useful example.

Historically, producing designs was time-consuming. Designers had to create each concept, component and variation manually. The cost of generating alternatives placed a natural limit on how many ideas an organisation could explore. AI is removing that constraint. A team can now produce multiple interfaces, prototypes and websites in minutes. But this does not mean design becomes unimportant. It means the bottleneck moves.

Once an organisation can generate 20 designs almost instantly, someone must still decide which one is right. The work must conform to the company’s design system, reflect the product strategy, incorporate customer feedback and satisfy multiple stakeholders with competing priorities. The difficult part of design was never simply drawing rectangles on a screen. It was reaching agreement about what should be built.

We believe this strengthens Figma’s role. Its value is not limited to the act of producing a design. Figma is becoming the shared environment in which product intent is expressed, debated and ultimately approved. Developers may generate code in Cursor or Claude. Designers may create an interface using Figma Make, Google Stitch or another AI tool. Product managers may begin with a written prompt. But these outputs still need somewhere to be reviewed, refined and reconciled.

That collaborative canvas can become the system of record for product intent: the place where the organisation determines not only what a product looks like, but what it is meant to do. The code may change constantly. The underlying intent must remain coherent.

Why software does not disappear

The assumption that AI will “eat software” often rests on a narrow view of what enterprise software does. Jensen Huang takes the opposite view, arguing that “the software industry in the future will be much larger than the software industry of today.” His reasoning is that software companies will no longer provide only tools for people. They will also provide specialised agents capable of using those tools. As the number of digital workers expands, so too could the demand for the software through which they operate.

Software is not merely a collection of menus and forms that helps employees produce an output. The most valuable platforms coordinate workflows, enforce permissions, preserve institutional memory and create accountability. They determine who is authorised to do what. They maintain the official version of a customer record, transaction, design or contract. They document how a decision was made and provide an audit trail when something goes wrong. AI can generate a plausible answer. It does not, by itself, make that answer authoritative.

These questions become more important as AI performs more consequential work. This is why systems of record may become more valuable rather than less valuable. Intelligence can be supplied by many different models, but the application controlling the workflow retains the customer relationship, proprietary context and responsibility for the outcome. The model generates. The software governs.

Models will be components, not necessarily destinations

Today, users often visit a particular model directly. They open ChatGPT, Claude or Gemini and begin a conversation. Over time, this intelligence is likely to appear inside existing workflows. An accountant may use AI through Intuit without selecting the underlying model. A designer may use several models through Figma. A large company might route different tasks between proprietary, open-weight and specialised models depending on cost, accuracy, privacy and latency. The application becomes the orchestrator.

This limits the power of any single model provider. If intelligence is available from multiple sources, software companies can switch models, use them selectively or run smaller models within their own environments. The model remains essential, but it risks becoming one component within a broader system.

The software platform possesses something the model provider often lacks: workflow context, proprietary data and an established position of trust. The closer a platform sits to the customer’s actual work and the more responsibility it assumes for the result, the greater its opportunity to capture value from AI.

The opportunity for Accenture

The same shift from production to judgment may benefit companies outside traditional software. AI can already generate code, analysis and presentations. But producing these outputs is not the same as transforming a large organisation. The harder task is deciding where AI should be deployed, redesigning the surrounding workflows and integrating it with legacy systems, proprietary data and existing controls.

This is where Accenture may become more valuable. Large enterprises are unlikely to rely on a single model. They will use a combination of proprietary, open-weight and specialised models, selected according to cost, capability, security and privacy. Someone must connect those models to existing systems, determine what data they can access, establish where human approval remains necessary and monitor whether they continue to perform as intended.

The more abundant and fragmented intelligence becomes, the greater this coordination challenge becomes. Accenture therefore does not need to own the leading model. Its opportunity is to become the integration and governance layer between increasingly capable AI and the complicated reality of the enterprise. It can help clients redesign workflows, deploy agents, establish guardrails and turn isolated demonstrations into dependable production systems.

This also changes the nature of consulting. As AI automates the production of code, analysis and presentations, consultants can spend more time understanding how organisations work and determining how people, software and AI should fit together. The opportunity for Accenture is not to charge clients for producing the same work more efficiently. It is to help clients reorganise themselves around a new form of intelligence. AI commoditises some of Accenture’s traditional output, but increases demand for the higher-value work of implementation, integration and organisational change.

Where value accrues

The central investment question is therefore not simply: who owns the smartest model? It is: who controls the workflow when intelligence becomes abundant? The strongest businesses are likely to combine several advantages:

• Distribution that places AI directly inside an existing customer relationship.
• Proprietary context that improves the relevance of its output.
• A system of record that remains authoritative.
• Governance that makes AI safe and accountable.
• Ownership of the workflow and responsibility for the result.

Intelligence will continue to improve, and its economic impact will be enormous. But intelligence alone may not be the scarce asset investors once assumed. As the cost of generating work approaches zero, the value of choosing, coordinating and taking responsibility for that work rises. The winners may not be the companies that produce the most intelligence. They may be the companies that help us decide what to do with it.