One of the questions I was asked recently was this:
What happens when subsidised token prices go away and organisations start feeling the true cost of AI?
It is a very good question.
Not because I think the AI market suddenly collapses when token prices change. I do not think that happens.
The more interesting issue is that cheap or subsidised pricing has made a lot of AI usage look simpler, and more economically viable, than it really is.
It has encouraged experimentation. That has been useful.
It has accelerated adoption. That has also been useful.
But it has also delayed a harder conversation. And that harder conversation is not just about token price. It is about the full economics of using AI properly inside real enterprises.
That includes the cost of the tokens, yes.
But it also includes utilisation, orchestration overhead, repeated context packaging, safety layers, routing, specialised agents, approvals, reviews, observability, auditability, and the general cost of making AI trustworthy enough to sit inside real business operations.
That is where the real economics start to bite.
Cheap tokens created adoption. They also created some bad habits.
Low pricing has had an obvious effect on the market.
It reduced friction. It encouraged teams to experiment. It made it easier for vendors to include AI in products. It allowed organisations to get people used to AI-assisted work without having to scrutinise every interaction too closely.
That was always likely to happen.
Whether one sees that as strategic subsidy, aggressive market-making, or simply part of a familiar platform-growth pattern, the effect is broadly the same. Expectations shift. Behaviour changes. What was once optional starts to feel normal. Then, before long, it starts to feel necessary.
People begin to expect faster turnaround.
Executives begin to expect summaries, drafts, analysis, and acceleration.
Developers begin to expect AI in the IDE (where they develop).
Support teams begin to expect AI assistance in handling requests.
Users become less tolerant of repetitive manual effort.
Vendors begin designing their products on the assumption that AI is simply part of the baseline.
That is important, because once the norm changes, the economics become much more consequential.
Token price is only half the story
A lot of the current debate still focuses on price per token.
That matters. Of course it does.
But I think many organisations are asking the wrong cost question.
They are asking:
How much does the model cost?
The better question is:
How much does a governed AI outcome cost?
That is a very different thing.
A single assistant-style interaction can look cheap enough:
one user
one question
one model call
one answer
That is the AI most people have in mind when they think about cost.
But that is not the AI architecture serious enterprises will rely on for more consequential execution.
The moment AI moves beyond casual assistance and starts participating in governed business workflows, the cost picture changes materially.
Enterprise AI gets expensive in the places people do not first notice
What looks like one simple AI task is often not one task at all.
It is a chain.
A request has to be classified.
Context has to be assembled.
Policy may need to be loaded.
The right role boundary must be applied.
The task may need routing to a specialist agent.
A safety check may need to happen.
A tool may need to be called.
The output may need review, scoring, or validation.
The next step may need another context package.
The entire interaction may need to be logged, versioned, and made observable.
And if the task crosses a threshold, a human may need to step in.
So what looked like “just ask the AI to handle it” becomes a multi-step governed workflow.
And every one of those steps has a cost.
This is one of the reasons I keep saying that the real economics of enterprise AI are not just about token price. They are about the total cost of making intelligence safe, specialised, governable, and operationally useful.
The email example makes the problem obvious
Take something as ordinary as email.
At first glance, it sounds easy.
You might say: why not just let an AI read each email and reply?
Simple enough.
Until you think about what enterprise email actually is.
Email is not one use case. It is a front door to many use cases.
Some emails are support queries.Some are complaints.Some are procurement requests.Some are legal issues.Some contain sensitive information.Some are phishing attempts.Some require escalation.Some involve policy interpretation.Some need a draft only.Some need action.Some need to be ignored.Some need to be routed to a particular team or role.Some create compliance or regulatory implications.Some require CISO sensitivity.Some require legal caution.Some require financial control.
So very quickly, the idea of one generic email agent starts to break down.
If you want this to work properly, you usually need:
a classification step
a security or phishing screen
routing logic
specialist handling by query type
role-based context
policy-aware prompts
output constraints
legal or compliance boundaries
and in many cases a review path before action is taken
And once you start doing that properly, you may not be dealing with one agent anymore. You may be dealing with dozens of narrower capabilities, each with its own prompt discipline, context structure, controls, and evaluation logic.
That is good design.
It is also more expensive.
Not necessarily because each individual step is expensive in isolation, but because the total workflow now consumes more AI calls, more context handling, more orchestration, and more runtime governance.
That is the real cost picture.
The better we design AI, the more cost discipline matters
This is where the conversation gets interesting.
Good enterprise AI design usually pushes you toward:
more specialisation
more bounded roles
more routing
more policy context
more controls
more observability
more evaluation
more review points where needed
All of that improves safety and quality.
But it also increases complexity.
And complexity increases cost unless it is managed deliberately.
That is why mature organisations are going to need a much more disciplined approach to AI consumption than many currently have.
Not because AI is going away.
Because it is becoming operationally real.
Even if prices fall, utilisation can still explode
This is another point I think many people underestimate.
Yes, token prices may well come down over time. That is entirely plausible.
Model efficiency will improve. Competition will increase. Smaller models will get better. Routing will improve. Caching will get smarter. Infrastructure will mature.
But falling unit cost does not automatically solve the enterprise economics.
Because utilisation can grow far faster than price falls.
Once AI becomes embedded across more workflows, teams, systems, and operating expectations, usage can explode.
So the organisation may pay less per token, while spending much more overall.
That is the familiar cloud story all over again.
Compute got cheaper. That did not stop cloud bills from growing.
The same thing can happen here.
Which is why usage governance matters just as much as price discipline.
Assistant AI and agentic AI are not economically the same thing
I think this distinction matters a lot.
Assistant AI is often relatively light:
one request
one answer
limited workflow implications
limited orchestration overhead
Agentic or governed execution AI is very different.
It is usually:
multi-step
context-rich
policy-bound
role-sensitive
integrated into other systems
observable
reviewable
and sometimes split across multiple specialist capabilities
That architecture is more powerful.
It is also more demanding economically.
So when organisations say they want “more agentic AI,” they need to understand that they are not just asking for more autonomy. They are often asking for more orchestration, more controls, and more consumption overhead as well.
That is not a reason to avoid it. It is a reason to price and govern it properly.
The next maturity step is cost-governed AI architecture
I think mature organisations will respond to this in a few predictable ways.
First, model routing will become mandatory.Not every task should hit the biggest model. Some tasks should use smaller models. Some should use cached outputs. Some should use retrieval. Some should use deterministic code or rules engines. Frontier models should be reserved for frontier-value tasks.
Second, token efficiency will become an engineering discipline.Bloated prompts, unnecessary context windows, repeated retrieval, verbose chains, and poor memory strategies will stop being harmless design flaws and start being real cost problems.
Third, cost-per-outcome will matter more than cost-per-token.The question will become: what did this workflow cost to run safely, and what value did it produce?
Fourth, step-level cost tracking will become important.Not just monthly vendor bills. Real visibility into which workflows, teams, agents, and process steps are driving cost.
Fifth, specialisation will need discipline.Yes, specialist agents can improve quality and safety. But organisations will need to know when the improvement justifies the added orchestration and runtime cost.
Sixth, governance overhead itself will need to be measured.How much cost is being introduced by review loops, repeated context packaging, validation steps, and safety layers? Where is it worth it? Where is it over-designed?
Those are serious questions. But they are the right questions.
The market will become more discerning
I do not think true-cost AI causes the market to retreat.
I think it causes the market to mature.
A lot of low-value, casual, speculative usage will get cut. That is normal.
The workflows that survive will be the ones with clearer economics:
replacing expensive labour
reducing risk
improving service outcomes
accelerating development
improving decision speed
shortening turnaround
reducing operational friction
or enabling revenue
In other words, the use cases that continue will be the ones that can justify themselves.
The rest will start to look like theatre.
Final reflection
Cheap tokens helped the market adopt AI.
That was probably always going to happen.
But the next phase is not really about cheap AI versus expensive AI.
It is about whether organisations are ready to treat AI as an enterprise resource with real economics attached to it.
That means measuring usage properly.
It means designing with cost in mind.
It means understanding that governed AI is more expensive than casual assistant AI.
It means realising that what appears simple at the surface often becomes much more complex once you add the architecture, controls, routing, specialisation, and oversight required to trust it.
And it means moving from “How much does the model cost?” to a much better question:
What does it cost to produce a governed AI outcome that the business can actually rely on?
That is the question mature organisations will answer well.
And I suspect that is where the next real separation in the market will come.
Cheap tokens created adoption.
True economics will create discipline.