AI is on your P&L. Most companies are only reading half of it.

Why AI budgets break

The mistake is managing AI as a technology expense instead of an economic system. AI isn’t like software, which has a far more predictable cost structure: buy seats, pay license fees, know your bill, and costs rise only when headcount grows or contracts renew.

With AI, that predictability disappears. Cost rises not just with the number of people, but what they ask, how deeply the model reasons, and how much work routes through AI systems. Increasingly, users aren’t people but agents and automated workflows that consume inference around the clock, with no human seat to count or license to cap.

In nearly every deployment we have examined, including our own, a small share does most of the spending: fewer than 10% of users and workflows drive the vast majority of the bill. The pull is relentless; in some of our environments, requests to lift individual spending limits arrive by the hundreds each day.

As AI moves deeper into the enterprise, consumption rises with every new workflow, agent and modality. Chatbots become agents. Agents become teams of agents. Text expands into multimodal generation with images, videos, and audio. And at the physical edge, autonomous robots draw on the same inference infrastructure, where every decision and exception carries a token cost the business did not plan for. As use cases multiply, the consumption curve compounds exponentially. Goldman Sachs projects token consumption will reach roughly 120 quadrillion tokens a month by 2030, about 24 times today’s level. Companies that treat current usage as the baseline plan against a world that will not last.

Security adds another hidden cost, and it is not one any organization can afford to skip. The more models, agents, data and connected systems a company adds, the more it has to defend. That defense work—constant scanning, tighter access controls, threat monitoring and fast response—puts a shadow price on every deployment. It rarely shows up in the first cost plan but grows with every new agent and connection.

This is the Jevons paradox in AI form: just as more efficient steam engines expanded Britain’s coal use rather than cut it, falling token prices push companies to run more AI, not less, as work once too expensive to justify suddenly makes financial sense.

Our simulation^[3]  found that if token prices fell 25% today, only 15% of organizations would bank the savings. The rest would reinvest in new use cases, expanded workloads or higher-quality models. Telling employees to use less AI is not the answer; it caps the productivity gains AI is meant to unlock. The answer is to route that new demand to the cheapest model that does the job well, so the same budget buys more useful work—and more measurable return.

A second force matters just as much: the Pareto frontier, the best trade-off between cost and value. Most companies overspend where the work doesn’t justify it or misroute tasks and leave value on the table. The most sophisticated companies close that gap, matching every task to the right model at the right cost, then shift the curve itself through prompt compression, smarter caching and model distillation.

Here’s a concrete example: a deck that once took three consultants 15 hours now takes one person two hours, at a token cost of a couple hundred dollars against thousands of dollars of freed time. That is the trade the frontier measures. The return is not the tokens saved; it is the freed capacity redirected to higher-value work. Measured against an outcome the enterprise already tracks, that reinvestment is where token spend converts into return.

Together, these forces are why tokenomics now matters strategically. As companies optimize to achieve more for less, the frontier moves outward: quality improves, costs fall, new use cases emerge and consumption rises. The discipline is knowing where AI creates value, when to harvest savings and how to reinvest them.

The model rule

The first rule is to match the level of intelligence to the value and risk of the work. Here’s a starting point: across industries, we estimate only 10-20% of enterprise tasks are complex enough to justify frontier or near-frontier capability.

The useful test: if a task breaks into well-defined steps, a smaller model will likely do; reach for frontier only when the work is irreducibly complex, when decomposition destroys the insight, when errors compound across long agent chains or when the answer sits at the intersection of legal, financial and regulatory judgment.

A few patterns clear this bar. Complex reasoning is one; when a chief strategy officer weighs competitive position, macro trends and an acquisition target's financials to name the real strategic risk, the task holds contradictory evidence in tension to land on non-obvious judgment. Also, agentic workflows with high error cost; when one mid-chain failure corrupts the whole output. Another is long-context synthesis; a 200-page due-diligence package demands rigor and coherence across the whole document.

Routing work to the right intelligence is a strategic choice, not a technical preference. Leaders need an operating discipline: classify work by complexity and risk, route it to the right level of intelligence, then tune and monitor consumption over time. The hard part is knowing which work matters most.

Companies that cannot distinguish routine tasks from cognitively demanding, higher-value work will misallocate in both directions: using expensive models where they are not needed and weaker models where advantage is at stake. Accenture’s experience shows model capability alone will not determine who wins the AI race. Economic discipline will.

The same discipline helped a telecom operator reduce annual AI spending by 68%. In each case, the breakthrough came from seeing consumption at the workflow level, then routing, tuning or redesigning work before waste became normal operating cost.

The C-suite mandate

In most companies, AI costs land in the technology budget while value shows up elsewhere—in customer service, engineering, marketing or whichever business unit uses the tool. The CFO sees cost, the CIO sees usage and business leaders see outcomes, but those views rarely connect at the enterprise level.