AI Bubble Narratives Misses the Token Waste Problem in Enterprise AI

Darragh Grove-White — Fri, 29 May 2026 19:20:05 GMT

"what did all that AI usage actually produce?"

AI has entered its “show me the value” phase.

For the last two years, many companies treated AI adoption as the goal. Get employees using it. Buy the licences. Run pilots. Encourage experimentation. Prove the organization is moving.

Now the bills are arriving.

Executives are asking a harder question: what did all that AI usage actually produce?

A postmortem is necessary. Some AI spending is poorly justified. Some valuations may be overheated. Some companies bought access before they had a clear plan for value. But the current cost panic is being interpreted too simplistically.

The contrarian take here is that high token consumption does not automatically prove that AI demand is fake. It often proves something more practical: companies scaled AI usage before they built AI operating discipline.

What is token waste in enterprise AI?

Token waste happens when companies spend heavily on AI usage without knowing whether that usage improved real work.

It shows up when teams burn tokens on long chats, oversized prompts, unnecessary agent runs, poorly scoped workflows, repeated rework, and vague requests that force AI systems to wander through too much context.

The problem is not token usage by itself. The problem is unmanaged token usage without context discipline, workflow boundaries, human review, and clear measures of accepted output.

Why high token usage does not prove AI demand is fake

For those unfamiliar, a token is the basic unit of language an AI system reads or writes. Prompts use tokens. Uploaded files use tokens. Long chat histories use tokens. Agentic workflows use a lot of tokens because agents search, plan, call tools, evaluate results, retry steps, and create intermediate work before producing the final answer.

That means a large AI bill can mean several things.

It can mean useful work is being done. It can mean employees are experimenting. It can mean the wrong model is being used for low-value tasks. It can mean agents are wandering through poorly defined work. It can mean teams are treating usage as productivity.

The problem is that most companies cannot yet tell the difference.

Said differently, many companies have not established context discipline with their teams. Without governance, controls, and clear workflow boundaries, enterprise AI token purchasing can become a blank cheque with limited return on investment attached to it.

That is the real story.

The first wave of enterprise AI adoption asked, “How do we get everyone using AI?”

The second wave will ask, “Which AI usage is creating durable work, and which usage is just burning budget?”

Where enterprise AI token waste begins

I see this pattern with leaders and professional-service firms trying to move beyond casual prompting.

They know AI can help. They also know their current use is messy.

People keep long chats open for months. Teams paste entire documents into prompts because they do not know which sections matter. Staff ask AI to “think deeply” about tasks that need a checklist. Managers see adoption dashboards but cannot say which workflows improved, much less why.

That is where token waste begins.

The issue is rarely that people are lazy or foolish. The issue is that most organizations have not been taught how to manage AI as a system.

They need to know:

Which work belongs in a simple workflow?
Which work requires an agent?
Which context is relevant?
Which context should stay out?
Which model is appropriate?
Where does human review belong?
How will we measure accepted output rather than activity?

This is especially important in regulated and high-trust work.

Insurance advisors, wealth firms, legal teams, accounting firms, and other advisory businesses cannot judge AI value only by speed. They also have to protect judgement, client trust, sensitive information, and case momentum.

A fast output that creates rework is expensive. A polished answer based on the wrong context is dangerous. An agent that searches, rewrites, and retries without boundaries may look productive while quietly increasing cost and risk.

Why token usage is a crude productivity metric

This is why Context-First AI matters.

The practical move is to make AI use more disciplined. Context needs to become infrastructure. Workflows need clear steps. Agents need harnesses. Outputs need review. Leaders need better metrics than “how much AI did we use?”

You would be far from alone if you thought the question was whether the company used more tokens this month.

But that is where many leaders miss the point. Token usage is a crude lead indicator disguised as a productivity metric. Really, it is just a raw input.

A raw input can tell you something is happening. It cannot tell you whether the work improved, whether rework dropped, whether judgement was preserved, or whether the company created a reusable asset.

Instead, look at the lag indicators: what happened downstream that you can measure, audit, and improve?

What leaders should measure instead

Did the work improve? If so, how?

Did cycle time drop? If so, when and where?

Did rework decline? If so, for whom, in which workflow, and by what evidence?

Did the output get accepted? By what standards and controls?

Did the team preserve judgement? If so, was it built into the process through human-gated reviews and embedded controls, or was it glazed on at the end?

Did the workflow become repeatable? If so, is it ready to become a structured AI workflow, or does it need more human feedback and evaluation cycles first?

Did the AI spend create a reusable asset? If so, who will use it, how will they use it, and when?

That is the beginning of token return on investment.

AI cost concern is healthy. It will force better questions. The danger is that leaders overcorrect and conclude that rising AI bills prove AI itself is failing.

A better conclusion is more precise: poorly managed AI use is expensive.

Three checks before your next AI budget review

Before your next AI budget review, ask:

1. Are we measuring useful work completed?

How are we measuring useful work completed? Are we effectively quantifying and qualifying its value, or are we only looking at AI activity, token usage, and the enterprise AI bill for the month?

If the only evidence is usage, you do not yet know whether AI improved the work.

2. Do our teams know what context belongs in each workflow?

Do our teams know what context belongs in each workflow?

Do they also know what context does not belong, so they avoid a context tax that makes model usage needlessly expensive?

Too little context creates shallow output. Too much irrelevant context creates cost, confusion, and avoidable review work.

3. Are we using agents only where autonomy is actually needed?

Are we using agents only where autonomy is actually needed, or are we using agents because we think we need more of them?

Do we have criteria for deciding whether an AI agent is valuable, whether an agentic workflow is the right fit, or whether a standard workflow would do the job better?

Some work needs deterministic process logic. Some work needs probabilistic reasoning. Some work needs both. But many tasks do not need an agent at all.

Frequently asked questions

Is high AI token usage always bad?

No. High token usage can mean useful work is being done, especially in complex workflows where AI needs to search, reason, evaluate, and revise.

The problem is high token usage without a clear connection to accepted output, reduced rework, faster cycle time, preserved judgement, or reusable assets.

What causes token waste in enterprise AI?

Token waste usually comes from weak context discipline, poorly scoped prompts, unnecessary agentic workflows, long-running chats, excessive document uploads, unclear review gates, and measuring usage instead of useful work completed.

How should leaders measure AI return on investment?

Leaders should look at downstream evidence: whether the work improved, whether cycle time dropped, whether rework declined, whether outputs were accepted, whether human judgement was preserved, and whether the workflow created a reusable asset.

When should a company use an AI agent instead of a standard workflow?

Use an AI agent when the task requires autonomy, tool use, branching decisions, retrieval, iteration, or multi-step work that cannot be handled by a simpler workflow.

If the task is predictable and repeatable, a standard workflow may be cheaper, safer, and easier to govern.

Why does context discipline matter for AI cost?

Poor context discipline increases cost because teams feed AI too much irrelevant information, keep long chats running, ask models to reason through poorly defined work, and use expensive models or agents where simpler systems would work.

The better first question

I speak to leadership groups and advisory firms about practical AI adoption in high-trust environments. I also work with firms to diagnose where AI is creating value, where it is creating waste, and where better context architecture would improve speed, quality, and governance.

If your leadership team is trying to separate useful AI work from expensive AI activity, I can help you diagnose where the value is real, where token waste is hiding, and where better context architecture would make AI safer, cheaper, and more useful.

For more practical notes on Context-First AI systems, subscribe to Context-First AI notes for leaders who need practical systems, not hype.

Grove-White Co Blog