How do you control an AI budget when tokens get cheaper but the bill keeps growing? Measure the cost of completing a business process successfully, not just the cost of calling a model. That calculation must include integration, oversight, employee time and error correction. Otherwise, falling technology prices can conceal rising operating costs.

Research by the Capgemini Research Institute, conducted in September 2026 among 500 senior executives at large organizations, suggests that this gap is already showing up in budgets.

  • 59% reported higher spending than expected.
  • 20% exceeded their budgets by more than 30%.
  • 53% were concerned about AI costs.

These are executive reports, not independently audited financial results, and the sample does not necessarily represent small and midsize businesses. Even so, the findings highlight an important management problem: a budget built around licenses and tokens does not capture the cost of operating a business capability.

Tokens got cheaper. The process may not have.

When the price per unit falls, it is tempting to assume total spending will follow. But an organization may expand usage, feed more documents into a model or deploy agents that make a sequence of calls rather than answer a single question. System integrations, access controls, quality checks and exception handling add further costs.

Other costs barely register on an invoice: an employee correcting a summary, a manager verifying a recommendation or a team repeating work already classified as automated. In these cases, the system is not necessarily reducing work. Sometimes it simply moves that work somewhere less visible.

The right unit of measurement is therefore cost per successful outcome: a support case closed without being reopened, an invoice processed correctly or a sales proposal approved for use.

A cheap model that creates extensive rework can cost more than an expensive model that gets the job done correctly.

The calculation should include model usage, infrastructure, maintenance, oversight and rework. Spread implementation costs over a reasonable operating period and workload, and state the assumptions. Otherwise, even a successful process can look expensive at the start or misleadingly cheap later.

Human oversight is a management mechanism, not an approval queue

AI can handle tasks that require interpretation and judgment, rather than simply execute fixed rules. That makes human oversight necessary: the outcome is not guaranteed in advance. But if every action requires someone to read everything and approve it again, meaningful efficiency gains become difficult.

Good design distinguishes routine, reversible actions from decisions with financial, legal or customer-service consequences. Allow bounded automation for low-risk activities, route exceptions for review and require approval before sensitive actions. Even the automated path needs monitoring and spot checks.

The goal is not to remove people from the process. It is to let them oversee much more activity without losing control. Measure the volume of exceptions, the time spent resolving them and the errors discovered after approval, not just the automation rate.

Watch out: Unbudgeted oversight is a hidden cost. Savings estimates are inflated if they ignore the time employees spend reviewing and correcting outputs. Design oversight around risk and include its cost in process-level measurement.

This distinction matters especially for agents operating inside enterprise systems. An inaccurate chat response needs correction. An incorrect action in an operational system may also require a rollback, an investigation and follow-up with a customer.

Multiple models can save money, but they also add complexity

According to the survey, three in five executives expect to adopt a multi-model approach over the coming year. There is a sound rationale: extracting a field from a document is not the same task as analyzing a complex contract. There is no reason to assume the same model is right for both.

Still, adding providers brings its own costs: evaluation, contracts, security, monitoring and maintenance. Using multiple models is a way to improve process economics, not a technology goal in itself.

  • One model throughout: Operational simplicity, with the risk of overpaying for straightforward tasks.
  • Routing between models: Better task fit, with added routing evaluation and maintenance costs.
  • Escalation to a more capable model: Targeted use of expensive capabilities, with repeat calls and additional latency to account for.
  • Human exception handling: Professional judgment, with labor costs and exception workload to manage.

Before changing models, test the candidate against representative work from your organization. A lower price is not an improvement if it increases rework or slows completion. Even switching between highly capable models requires testing: general benchmarks alone cannot establish whether a model fits your process.

The most important gains may not reduce payroll spending

Capgemini proposes a metric called AI Yield to document the value generated by AI investment. The direction is useful, but the measure comes from a consulting firm with a commercial interest in the field. An organization should first define the value it wants to demonstrate and how it will avoid double counting.

In particular, distinguish three outcomes:

  • Realized financial savings: Spending actually eliminated or reduced, net of the solution's costs.
  • Freed-up capacity: Working time saved that can support additional tasks, provided the work is actually reorganized.
  • Business improvement: Faster service, fewer errors or additional revenue, where the connection to the process change can be substantiated.

Freed-up time is an operational asset, but it does not automatically translate into cash savings. Equally, a budget overrun is not necessarily a failure if activity has grown and cost per outcome has fallen. The CFO needs to see both total spending and the output that spending purchased.

Connect finance, operations and IT

Accountability cannot sit with the technology leader alone. The process owner defines success, finance validates the value, and IT owns infrastructure, permissions and monitoring. AI expertise connects these responsibilities, but it does not replace a working understanding of the business process.

  1. Define the business outcome: Choose a process and specify what successful completion means.
  2. Establish the baseline: Record costs, handling times, errors and rework before implementation.
  3. Budget for the full operation: Include implementation, usage, maintenance, security and human oversight.
  4. Test in a controlled environment: Compare performance on representative cases and define stopping and escalation conditions.
  5. Scale based on results: Expand only when quality and cost per outcome justify it.

This also requires internal capability: employees who know how to work with models and assess their outputs, and a team that can build and manage agents on shared infrastructure. A personal tool requires people to change their working habits. An agent embedded in an existing process may demand less change from employees, but more engineering and management discipline behind the scenes.

When choosing an implementation team, look for relevant education, deep AI expertise and experience with real business processes. An impressive demo is no substitute for being able to explain what happens when the system gets something wrong, who handles it and how much that response costs.

The leadership question is not simply how much we spent on AI, but how much it cost us to achieve a better outcome. Once that measure is defined, budget decisions become grounded in evidence: stop an inefficient deployment, improve a promising process or increase investment where value has been demonstrated.