How to benchmark AI budgets to optimize ROI

Despite talk of an AI bubble, AI spending is not slowing down. Indeed, companies plan to commit 1.7% of their annual revenue to AI initiatives this year, up from 0.8% last year, according to research from the Boston Consulting Group.

At the same time, however, PwC’s 2026 Global CEO Survey shows that a growing number of companies are not seeing a return on their investment. A survey of more than 4,400 CEOs found that 56% have yet to see a financial benefit from their AI initiatives, while just 12% have seen both an increase in revenue and a cost reduction. Only a third are confident of revenue growth this year amid struggles to optimize AI’s ROI.

The disconnect can be put down to the fact that too many companies are buying AI tools, rolling them out to employees or customers, and only wondering several months later why they haven’t seen ROI. The problem is less to do with companies not getting value out of AI and more to do with the fact that they haven’t defined what value actually means for their business.

“A common mistake is spending heavily on tools before really understanding the problem [that needs to be solved],” says Jack Rickhuss, managing director and co-founder of tech consultancy Journi.

“AI only delivers real value when it’s in the hands of people who know how to use it properly.”

Tie AI investments to clear outcomes

Benchmarking AI budgets is becoming a critical discipline that leaders need to master to ensure they’re extracting value from AI deployments. Without benchmarking an AI project’s spend, companies have no objective way of knowing if AI is performing well or poorly and delivering value for money.

Adam Hofmann, partner (AI and people transformation) at challenger consultancy Elixirr, says that the trap he sees leaders fall into is deploying AI tools and celebrating personal productivity gains while the profit and loss doesn’t move.

This can happen because companies are “benchmarking off peers that don’t know what they’re doing either, which leaves you behind the curve or overspending on someone else’s confusion.”

Before getting started on benchmarking AI budgets, leaders should build a clear picture of what they’re currently spending on model API costs, such as per token, infrastructure such as compute, hosting, and storage, human validation, maintenance, and training. Hofmann warns that “if you can't trace the line from spend to outcome, you're measuring activity, not AI."

Rickhuss agrees, adding that most companies “are still measuring the wrong things”. He advises tying AI investment to clear outcomes, such as improving decision-making, saving employees’ time, and speeding up delivery of projects. “If you can’t link [AI investment] to something tangible, it’s hard to measure success.”

Review what industry peers are doing and then pilot

To get started, leaders should compare their spend against the peers that do know what they’re doing. One way they can do this is look at other companies in their industry using their earnings calls and 10-K filings, as well as analyst reports. They could use a large language model to extract data on peers’ AI spend as a percentage of revenue.

The next step is to break down AI use cases into various categories. This would cover the infrastructure (e.g. cloud compute and GPUs), people (e.g. the salaries for AI talent), licensing (e.g. cost-per-token), and data (e.g. storage).

Once AI budgets have been broken down, Shiro Theuri, CTO of Spanish on-demand delivery company Glovo, recommends taking “a pilot-and-test approach”. This is because piloting tools in a controlled environment can help to prevent feature creep and shadow IT. Pilots can also surface hidden costs that might otherwise have been overlooked.

An IDC and DataRobot survey carried out last year showed that 92% of enterprises deploying agentic AI at scale admitted that the costs incurred were higher than they had projected. The survey of 318 senior decision-makers at companies with more than 1,000 employees found that token consumption and hallucination remediation were the top unexpected costs, while inference was another common issue.

Continue to benchmark throughout a project’s lifecycle

To prevent their AI projects from getting stuck in pilot mode, leaders should continuously monitor the cost of running the project and benchmark this against the company’s own expectations.

Monitoring token consumption can be an effective way to keep budgets under control. AI costs can quite easily skyrocket, especially if employees end up using more tokens than forecast, running up higher bills and leading to companies exceeding their AI budgets. Luke Budka, AI director at marketing and training firm Definition, adds that tracking token usage can “reveal super users and also help identify employees who need support”.

While token usage can be used as a useful metric for cost management, leaders should be careful of ‘tokenmaxxing’ – the trend of enterprises measuring individual employee token usage and using it to measure productivity. As Budka explains, high token consumption can occur when employees are purposely inflating their usage to look busy, but it can also be a sign that employees are struggling with AI tools and could benefit from more training.

Ultimately, rather than focusing on the cost-per-token, leaders should measure the cost-per-business outcome. Determining the number of tokens required and tracking the tokens consumed to complete each workflow will help to optimize AI’s ROI.