Your AI Bill Is as Low as It Will Ever Be

Goldman Sachs Research projects token consumption rising by 2400% by 2030. That number is easy to read as a story about AI adoption, but adoption alone does not explain the scale of what is coming. The bigger shift is architectural: we are moving from a world where AI is mostly used through chat interfaces to one where agents run work on their own.

That distinction matters because chatbots and agents consume tokens very differently.

A typical chatbot interaction is relatively small. A user opens a tool, asks a question, gets a response, and moves on. In many cases, that means a prompt in the range of a few hundred to a few thousand tokens.

An agentic workflow is different. It can consume tens of thousands - or even hundreds of thousands - of tokens to complete a single task because it is not just answering once. It is reasoning in steps, deciding what to do next, prompting itself, checking outputs, and continuing until the task is done.

This is why token growth is likely to accelerate much faster than many teams expect.

Chatbots depend on behavior change

The first wave of enterprise AI was shaped by chat. That created a fundamental limitation: people had to opt in.

A chatbot only creates value if someone remembers it exists, decides it is the right tool for the job, opens it, and writes a useful prompt. Every one of those steps introduces friction. Even when the technology is powerful, adoption gets capped by human behavior. If employees do not change how they work, usage remains limited.

That is one reason enterprise AI adoption has often looked underwhelming in broad market studies. If AI is presented as another destination, another app, or another workflow users have to consciously initiate, many organizations never get beyond sporadic experimentation. People may be impressed by the capability, but that does not mean it becomes embedded in day-to-day operations.

The result is a mismatch: high excitement, modest utilization.

Agents reverse the adoption model

Agents change that equation because they do not wait for a user to begin the interaction.

Instead of relying on someone to remember to use AI, an agent can be triggered by an external event: a new email, an incoming invoice, a support request, a document upload, or some other operational signal. Once triggered, it starts working immediately. It processes information, makes intermediate decisions, calls other systems if needed, and continues looping until the task is complete.

That flips adoption from opt-in to effectively automatic.

A chatbot asks a human to go out of their way. An agent meets the work where it already happens.

This is the real reason token consumption is poised to explode. The shift is not just from fewer users to more users. It is from occasional one-turn interactions to multi-step autonomous processes. When AI becomes part of the workflow itself rather than a separate tool beside the workflow, token usage stops being episodic and starts becoming structural.

In other words, consumption grows because AI is no longer something people “try.” It becomes something the business runs.

Why enterprises are a natural fit

This dynamic is especially strong in enterprise environments.

Consumer use cases can be unpredictable. People behave differently, inputs vary widely, and the value of automation is often tied to personal habits.

Enterprises are different. Most organizations already run on documented, standardized workflows. They know what data comes in, what steps usually happen next, and what outputs are expected.

That makes enterprise operations a strong match for agentic systems.

When workflows are repeatable, agents become easier to deploy with confidence:

• A finance-related process can be triggered by an invoice.
• A support-related process can be triggered by a ticket.
• A research-related process can be triggered by a new request.

In each case, the pattern is familiar, the inputs are known, and the result has a clear business purpose.

That is why many companies are not adopting agents one by one as isolated experiments. They are starting with a single high-value use case, seeing immediate ROI, and then expanding quickly into many more. Once the first workflow proves itself, the next dozen become easier to justify.

The adoption curve is not linear. It compounds.

The value is real, but so is the cost

The upside is obvious. Agents can unlock efficiency, responsiveness, and scale in ways chatbots often cannot. They reduce the burden on employees to manually initiate every AI interaction. They fit more naturally into real operational systems. And because they can work through entire tasks rather than isolated prompts, they can create more measurable business value.

But there is a second side to this story: cost.

If agents become the default way work gets done, token consumption does not just rise with user count. It rises with every triggered workflow, every loop, every reasoning step, and every automated handoff. That means many organizations are underestimating the infrastructure bill that will come with successful AI deployment.

For teams experimenting today, this is the cheapest their AI bill is likely to be.

That is not a warning against adoption. It is a reminder to plan for success correctly. If the future of enterprise AI is agentic, then token usage should not be treated as a temporary line item attached to pilots. It should be treated as a growing operational cost tied to real business throughput.

The more useful your agents become, the more they will run. The more they run, the more tokens they will consume.

That is not a failure of the model. It is evidence that the model is working.

Conclusion

The projected growth in token consumption is not just about more people using AI. It is about a deeper transformation in how AI gets deployed inside organizations.

Chatbots require humans to opt in. Agents do not. They are triggered by the flow of work itself, and they keep going until the work is done.

That shift has major implications for enterprise adoption, ROI, and cost planning.

The value of agents is becoming clear very quickly. So is the reality that autonomous workflows consume far more tokens than simple chat interactions ever did.

For organizations building in this direction, the opportunity is enormous.

So is the bill.