McKinsey Regulates the Use of AI Tokens

The shift to consumption-based billing is putting pressure on AI budgets. McKinsey emphasizes transparency and education, with individual alerts, a prompt gateway, circuit breakers, and caching. EY claims a 60% reduction in tokens since April, while developers at Deloitte report hitting new quotas. Debasish Patnaik recommends measuring cost per outcome and further encouraging adoption.

Among peers, visible cuts and felt quotas

EY has established an AI Value Realization Office and an invisible router to direct usage towards the most suitable model. According to EY, these tools, along with other governance initiatives, have led to a 60% reduction in token consumption since April. In June, a senior software engineer at Deloitte US indicated that the change in GitHub's pricing model had altered work expectations, with developers quickly reaching the new monthly caps set. In September, OpenAI reported that its most intensive coding agent users were consuming over $7,000 worth of tokens per day.

Gateway, circuit breakers, and caching to smooth consumption

At McKinsey, an internal gateway optimizes queries before they reach model providers. Circuit breakers can temporarily suspend access in cases of particularly high token consumption, allowing time to assess whether the usage is productive. Caching enables the reuse of responses to repeated questions and pools costs at the enterprise level rather than locking them into individual licenses. Patnaik believes that usage could become more pronounced in six months but notes that additional controls are already in place.

Governance focused on autonomy and initial effects

Alerts sent to employees are not intended to limit AI usage but to make consumption visible and help accomplish the same tasks more efficiently. Patnaik asserts that consultant autonomy remains a principle and that training in tool usage has been essential for managing consumption. He states that the company has already observed changes in practices since employees gained more visibility. He also indicates that internal usage has not yet reached a problematic level and that, for productivity-related or deliverable acceleration uses, the cost-benefit ratio remains favorable.

Individual tracking, summer deployment, and high volume concentration

McKinsey tracks token consumption for each user and sends email notifications in case of overages, a system that was rolled out during the summer period. According to a company blog post, nearly 10% of users account for about 65% of the volumes, with consultants and engineers among the primary users, and the company expected to process around five trillion tokens per month by May 2026. Patnaik places this approach within a framework of transparency and awareness. More broadly, the company has adopted internal controls to manage usage, prioritizing information over restriction.

Consumption-based pricing and advice: aiming for cost per outcome

The end of flat-rate subscriptions in favor of consumption-based billing earlier this year forces companies to monitor the number of tokens used. McKinsey implemented a formal practice this spring to guide its clients towards profitable deployments. Patnaik recommends evaluating the cost of tokens per outcome rather than per employee and warns against optimization aimed solely at reducing tokens, which risks stifling valuable uses. He also highlights that a cost reduction in one area can generate costs elsewhere, and reports that, in the last quarter, clients became aware of hidden costs to weigh against expected competitive benefits. For now, he believes it remains relevant to encourage adoption. His role primarily involves providing AI capabilities to clients, while McKinsey is already advising large companies on budget management. Consulting firms have, for their part, widely introduced AI and are now facing the associated costs. Patnaik compares the approach to information on mobile data consumption on a work phone. He leads QuantumBlack in the UK.