Consulting firms quickly moved to make AI part of employees’ working lives. Now they are faced with the costs of using them on a large scale.
One of McKinsey’s strategies to manage rising costs is not to limit consultants’ use of AI, but to let them know when they are using too much.
The company tracks usage per user and alerts its employees via email when their AI usage becomes too high, says Debasish Patnaik, who heads QuantumBlack – McKinsey’s AI, data and analytics group – in the UK. The warning system was introduced company-wide in the summer.
“Just like using mobile data on a work phone, we tell them this is the way you can do things to make it more cost-effective for the business,” Patnaik said.
Companies that initially focused on encouraging AI experimentation are now faced with the costs of adopting the technology at scale.
The age of freewheeling AI token usage came to an end earlier this year when LLM providers switched from subscriptions to consumption-based pricing models. Consulting firms and other business users are now charged based on the number of tokens (the small pieces of text that an AI model reads and produces) they use, rather than a flat access fee, making efficiency more important as usage grows.
In September, OpenAI announced that they are now the most prolific users of AI coding tools consume more than $7,000 tokens worth per day.
McKinsey’s approach to managing those spending pressures relies on transparency and education, Patnaik said.
According to a company blog post, the company was processing approximately five trillion AI tokens per month by May 2026. Consumption was highly concentrated, with about 10% of users accounting for about 65% of the total, with consultants and software engineers among the heaviest users.
Debasish Patnaik is the leader of McKinsey’s QuantumBlack business in the UK. McKinsey
The AI usage alerts are not intended to discourage employees from using AI, but to show them what they are consuming and help them perform the same work more efficiently.
“We really believe in giving autonomy to the consultants,” said Patnaik.
Ensuring they receive proper training on how to use tools is critical to managing usage, he added. He noted that McKinsey has already seen shifts in the way employees use its tools as users gain more insight into their AI habits.
In addition to token warnings
McKinsey’s internal AI spending has not yet reached a problem stage, Patnaik said.
Most of the company’s AI use is for personal productivity or to drive faster engagement, in both cases “where the cost-benefit ratio still leans toward the benefit side,” he said.
Usage levels could be “egregious” in another six months, but Patnaik said the company had already implemented additional controls in addition to warnings to manage usage.
An internal AI gateway optimizes requests before they reach model providers, while circuit breakers temporarily pause access around particularly high token usage while the company assesses whether usage is productive.
Caching allows answers to repeated queries to be reused and aggregates consumption costs across the enterprise rather than locking up unused capacity in individual licenses.
Other consulting firms are implementing similar controls to manage AI spend.
The Big Four firm EY has setting up an ‘AI Value Realization Office’” to manage AI spend and has a “invisible” router behind some specialized AI tools that guide workers to the model best suited to a task. EY told Business Insider that the router, among other control measures, had reduced token consumption by 60% since April.
In June, a senior software engineer at Deloitte US told Business Insider that changes to GitHub’s pricing model were “already wreaking havoc” on expectations for work, with developers quickly burning through their new monthly quotas, which took effect that month.
Bringing the strategies to customers
Patnaik’s role leading QuantumBlack in the UK is primarily focused on delivering AI capabilities to McKinsey clientsrather than the company’s internal use of AI. McKinsey set up a formal practice this spring to advise clients on how to use AI cost-effectively.
Patnaik’s broad advice for clients reflects the firm’s internal approach.
Over the past quarter, customers have increasingly realized that there are “hidden costs that we haven’t fully thought about,” and that managers must weigh the potential competitive advantages of AI against the return on investment, Patnaik says.
Patnaik said companies should assess token costs per outcome rather than per employee. Simply optimizing for fewer tokens could discourage valuable use of AI.
“What you don’t want is to incur costs here, but on the other side you incur costs without knowing it,” he said. For now, he added, there is still value in encouraging adoption, rather than restricting it.
