Corporate artificial intelligence bills rise despite falling token prices

Companies that can control the rapidly rising cost of artificial intelligence could gain a competitive advantage as AI models become increasingly commoditised, according to PwC.

The professional services firm said cost-control tools for AI had become widespread and increasingly standardised, meaning they were necessary but no longer sufficient to give companies an edge.

A new operating model focused on controlling AI spending could instead allow businesses to generate savings and reinvest them in further AI initiatives, creating what PwC described as a compounding advantage.

PwC said one global technology company using this approach cut the cost of each run by 65 to 80 per cent, allowing it to run three to five times as much AI for the same budget.

The firm said AI spending was rising despite falling token prices because lower costs were encouraging companies to use AI across more processes, potentially increasing complexity and expenditure.

“Everyone tries to use AI everywhere, even if it just makes workflows more complex and expensive,” PwC said, arguing that widespread use of the same AI models meant spending alone did not necessarily provide a competitive advantage.

PwC said companies were often consuming more tokens than necessary without having systems capable of identifying where waste was occurring.

AI costs could also be difficult to track because tokens accumulated across planning, tool use, retrieval, reasoning, orchestration, safeguards, logging and review, while indirect infrastructure costs were often excluded from initial budgets.

The firm also warned that AI agent costs could compound rapidly, as agents created plans, delegated tasks to other agents, retrieved information and repeated processes when results were unsatisfactory.

At the same time, the price of one million tokens could range from pennies to as much as $50 depending on the model and tier, making model selection an important cost decision.

PwC stressed that choosing the cheapest model was not always the most economical approach because a weaker system could generate additional work, poor decisions or compliance failures.

Instead, companies should match the model tier to the task, balancing performance against cost.

PwC said traditional financial analysis could then be applied to AI spending by examining three elements of cost overruns: rates, such as supplier price increases; volume, including excessive calls and retries; and mix, meaning the use of inappropriate model tiers.

The proposed operating model centres on four disciplines: assessing the cost and value of AI projects before development, redesigning systems to reduce waste, linking spending to business outcomes, and reinvesting savings into further AI projects.

This could involve measures such as reducing unnecessary context, combining multiple tasks into fewer calls, imposing spending limits and automatically routing work to the least expensive model capable of completing it.

PwC said these controls should be built directly into AI systems, with mandatory budget limits, routing rules, workflow thresholds and audit trails.

Human oversight would remain important, with technology automatically flagging decisions for people and providing the information needed to act according to business priorities.

In the technology company case study, the approach also cut average runtime from 12 hours to four hours, while an end-to-end audit found that output quality was maintained.

PwC said businesses should establish how much each AI workflow costs per business outcome, put AI spending on the chief financial officer’s agenda and prepare for suppliers increasingly moving towards outcome-based pricing.

It also recommended giving designated finance owners both responsibility and technical support to enforce spending discipline, rather than simply providing them with dashboards.

The firm said companies should begin with their most valuable AI applications, where cost assessment, implementation and governance could have the greatest impact.

“The next round of AI advantage won’t go to whoever runs the most powerful models,” PwC said, arguing that companies would increasingly use the same underlying systems.

“Advantage will likely go to whoever runs them with more discipline,” the company concluded.