For the last two years, many companies have treated AI progress like a rollout metric. More users. More prompts. More internal dashboards. More teams proudly reporting that people are finally "using AI".
In the first phase, that was understandable. New technology needs experimentation. Employees need confidence. Leaders need visible signals that the organisation is moving.
But that logic will not carry companies through the next phase of AI.
Because AI is moving from answering questions to executing work. And once AI starts to execute work autonomously, usage is no longer just a behaviour metric. It becomes a cost driver. That is the shift many leadership teams are still underestimating.
We have seen this pattern before in digital markets
Many digital platforms started with low-friction promises: unlimited access, cheap subscriptions, generous usage, simple pricing and almost no visible marginal cost. Consumers and companies built habits around that convenience. Then the market matured. Prices increased. Premium tiers appeared. Advertising was added. Usage rules became more specific.
Streaming followed that path. Marketplaces followed that path. Cloud software followed that path. Media platforms followed that path.
The early promise was simplicity. The mature reality became segmentation, pricing power and monetisation discipline.
AI may follow a similar pattern, but faster. The difference is that AI usage is not only tied to access. It is increasingly tied to work execution.
Agentic AI turns usage into variable operating capacity
A chatbot is one cost logic. An AI agent that researches, browses, writes, checks, retries, calls tools, triggers workflows and repeats tasks at machine speed is something very different.
Once agents become part of the operating model, companies are no longer only buying software. They are buying variable operating capacity. And that capacity will not stay cheap forever.
Most companies still look at AI through a SaaS lens. They are used to predictable software logic: buy a seat, roll it out, drive adoption, track usage. The financial model feels familiar because enterprise software has trained us to think in licenses, users and monthly subscription fees.
Agentic AI breaks that mental model. A human user has a natural usage limit. People take breaks. They get tired. They stop prompting. An AI agent does not have the same boundary. It can run multiple steps, use external tools, generate intermediate outputs, retry failed attempts and consume tokens continuously if the workflow allows it.
"All-you-can-eat AI" is unlikely to survive serious enterprise use
The broader signal is already visible. Certain AI subscriptions are becoming more restrictive when they are used for external agent-tool workflows. More usage is moving behind credit meters, workflow packages or token-based pricing.
The license becomes the entry ticket. The real bill starts moving with usage.
This is not a criticism of any one provider. It is the natural direction of the market. AI infrastructure, inference, model access, context windows, workflow automation and agentic execution all have underlying costs. Vendors will not subsidise enterprise-scale usage forever.
Token consumption is not productivity
High token usage is already becoming a strange status signal in parts of the AI world. "Tokenmaxxing" is treated almost like proof that someone is advanced in AI. Some teams celebrate heavy usage. Some dashboards highlight prompt volume. Some organisations are starting to confuse AI consumption with AI value.
But token consumption is not productivity. More prompts do not automatically mean better decisions. More generated content does not automatically mean stronger marketing. More automated workflows do not automatically mean better customer experience. More AI activity does not automatically mean more business impact.
If companies celebrate usage, they will get usage. If they celebrate token burn, they will get token burn. If leadership teams build AI dashboards around activity instead of outcomes, people will optimise for the dashboard.
AI adoption is not the same as AI impact. Token consumption is not the same as productivity.
The next AI productivity logic needs to measure value, not activity
The next phase of enterprise AI will require a more mature productivity logic. Usage is only an input. The real questions are different: Does AI improve decision quality? Does it reduce cycle time? Does it increase customer value? Does it strengthen brand relevance? Does it improve commercial outcomes? Does it free people to do higher-value work? Does the value created justify the compute consumed?
These are not technical questions. They are leadership questions.
AI governance needs a cost dimension
Most companies are currently building AI governance around data protection, compliance, legal risk, security and responsible AI. All of that is necessary. But it is not enough.
AI governance also needs a cost dimension. Which workflows are worth automating? Which tasks justify frontier model usage? Where is a smaller or cheaper model good enough? Which agents need budget caps? Which workflows should require human approval? Who owns the cost of AI usage across departments? And how do we measure whether AI creates business value instead of just activity?
Without this discipline, the problems are predictable. Some teams will overuse expensive models. Others will automate low-value work. Dashboards will show adoption, but not impact. Budgets will be consumed before anyone understands which use cases actually create value. That is not transformation. That is unmanaged experimentation at scale.
Commercial teams need AI judgment, not another activity machine
In marketing and commercial teams, this distinction matters especially. AI can help create, analyse, test and personalise at scale. But more content is not automatically better brand building. More automation is not automatically better customer experience. More experiments are not automatically better growth.
AI should not become another activity machine. It should become a value machine.
That requires companies to prepare for the cost reality of agentic AI now. Not by using less AI, but by using AI with more judgment.
The next AI surprise may be financial, not technical
Many companies are prepared for AI pilots. Fewer are prepared for AI cost governance. Many have usage dashboards. Fewer have unit economics. Many are rolling out tools. Fewer are deciding which workflows deserve compute, which models should be used for which tasks and which AI agents need clear budget boundaries.
That gap will become visible as AI moves deeper into operating models.
The early phase of AI was about adoption. The next phase will be about monetisation discipline.
The AI race will not be won by the companies burning the most tokens. It will be won by the companies that know which work is worth the compute.