AI spending is rising. But are companies getting enough back?

Gartner expects more than 40% of agentic AI projects to be canceled by the end of 2027, undone by escalating costs, unclear business value, and weak risk controls. The technology isn’t the problem, three technology executives argue. The problem is that most companies still can’t say what their AI spending is actually buying them.

According to Andreas Hassellöf, CEO of Ombori, businesses should look at “the cost per business outcome” rather than the cost of individual tokens. Mohammed Aboul-Magd, General Manager – Cybersecurity at SandboxAQ, says companies need to connect AI costs to specific pieces of work, measuring “cost per agent and per task” against a baseline. Alexandre Depret-Bixio, SVP, International at Anomali, takes the argument back to the data itself, saying that “nine times out of ten, the biggest cost driver isn’t the AI model, it’s the messy data organizations feed it.” Different starting points, but the same problem: AI spending is easy to see on an invoice. Its business value is harder to pin down.

THE REAL COST IS THE WORK AROUND AI

For Hassellöf, the fixation on token prices overlooks the larger shift underway within companies. “The wrong question is whether the cost per token is going up,” he says. “The right question is whether the cost per business outcome is going down, and whether the outcome itself has become more valuable because the process was redesigned.”

He compares the transition to the early days of electrification, when factories replaced steam engines with electric motors but continued to use the same line-shaft systems. The technology changed; the factory did not. The productivity gains only arrived when manufacturers redesigned the way production worked around the new source of power.

AI presents a similar choice. Companies can spend their time squeezing the cost of individual interactions, or they can rethink the workflow those interactions sit inside.

“AI should be evaluated like any operational investment that changes the architecture of work,” Hassellöf says. That means looking at what it costs to complete a task today, what it costs after the process has been redesigned, and what new value becomes possible as a result.

Meanwhile, Aboul-Magd sees the same measurement problem from a different angle. Token spend often appears as a technology cost, separated from the business activity it supports.

“Most companies today can see the cost but not the return, because token spend arrives as one line on a technology invoice, disconnected from the work it paid for,” he says.

His answer is to bring AI measurement down to the level of the task: a proposal, a customer ticket, a sales close or another defined piece of work. If a process that previously took three weeks can now be completed in days, that change can be measured.

The distinction matters because AI does not necessarily deliver value in a straight line. As an agent learns within a workflow, its usefulness can increase while its usage and accuracy figures tell only part of the story.

“Measure execution and outcomes as a trend rather than a snapshot,” Aboul-Magd says.

“The objective is to deliver useful business outcomes repeatedly, economically and at the speed each use case requires,” says Sherif Tawfik, Chief Business Officer at Core42. “Organizations must measure AI against outcomes rather than activity.”

STOP MAKING THE MODEL DO EVERYTHING

One of the quickest ways to drive up AI costs is to assume that every problem needs the most capable model available.

“We are still designing many AI systems as if every problem requires a conversation with a large language model,” Hassellöf says. “That is an expensive architectural assumption.”

His recommendation is intelligent routing: deterministic logic for straightforward rules, smaller models for simpler reasoning, retrieval when the information already exists, and frontier models only when they genuinely add value.

Tawfik makes the same point from an infrastructure perspective. “Not every task requires the largest model, the most expensive accelerator or a real-time deployment path,” he says.

“Organizations can reduce costs by routing each workload to the most appropriate model, accelerator and deployment environment. This improves utilization, avoids paying for idle or over-specified capacity and reserves premium infrastructure for applications that genuinely require it.”

The bigger question, however, is whether some steps in a workflow need to exist at all.

Companies should ask which processes remain in place simply because human cognition was historically scarce and expensive, Hassellöf says. Some of those steps should disappear rather than be automated.

Aboul-Magd points to another source of waste–repetition.

“The biggest hidden cost is repetition,” he says. Many AI agents begin every task from scratch, repeatedly consuming the same documents, context, and decisions. Every handoff between tools can also force an agent to reconstruct information it already had.

Persistent, shared context can remove much of that waste. So can matching the model to the task.

“A frontier model drafting a status summary is a senior hire doing data entry,” Aboul-Magd says. Smaller models can often deliver the same quality at a fraction of the cost, but companies need to measure both quality and cost to identify such opportunities.

Depret-Bixio sees a similar problem before the AI even gets involved,

duplicated, disconnected, and poorly structured information forces AI systems to spend time checking and rechecking it before they can do anything useful. In cybersecurity, where teams can already be overwhelmed by huge volumes of signals, that problem becomes particularly visible.

The fix is less glamorous than buying another model: improve the data first.

“Remove duplicates, standardize the format, add context upfront,” Depret-Bixio says. The result is fewer tokens, faster answers and more accurate outputs.

ACCURACY HAS A PRICE. SO DOES GETTING IT WRONG

The push to reduce AI costs creates another question: how much validation and context can companies afford to remove without undermining trust?

Aboul-Magd argues that the answer depends on what the AI is being asked to do.

“I’d resist the framing slightly, because it’s a tiering decision more than a tradeoff,” he says. An agent summarizing meeting notes does not require the same level of assurance as one handling financial figures or client data.

That means companies need to match the cost of validation to the consequences of being wrong.

Continuous evaluation can help determine where additional context genuinely improves an outcome and where it simply adds expense. Without that evidence, companies risk going to either extreme: validating everything or validating almost nothing.

“The cost of validation is visible on the invoice, while the cost of a wrong answer, a bad number in a board pack, a claim that doesn’t hold, stays invisible until it’s very large,” Aboul-Magd says.

Hassellöf makes a similar argument from an architectural perspective.

“More context does not automatically mean more truth,” he says. “At some point, you are paying increasingly more to reduce increasingly smaller amounts of uncertainty.”

His answer is to make confidence part of the architecture. Companies should define the level of certainty a decision requires, then add data, validation or human intervention only when that threshold is not met.

“Trust, not fear, should determine how much intelligence you deploy,” he says.

That also changes how companies think about proprietary data. Domain-specific information, closed-loop feedback and internal context can sometimes produce more reliable systems without continually adding layers of external information.

AI SPENDING IS NO LONGER JUST AN IT PROBLEM

As AI agents move beyond answering questions and into executing work, the economics become harder for boards to ignore.

“Because agents are no longer tools so much as actors,” Aboul-Magd says. They write code, draft documents, open tickets and interact with real data. That gives an AI cost line a different significance from a conventional software bill.

A growing expense with operational authority and no clear measure of return is, in his view, a board-level issue.

He points to Gartner’s prediction that more than 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls as reasons.

“The programs that get cut won’t be the ones that spent the most,” Aboul-Magd says. “They’ll be the ones that couldn’t answer what the spending returned.”

Hassellöf similarly sees AI as moving into the operating model rather than remaining an IT function.

“You cannot outsource the economics of a capability that is changing how the company operates,” he says.

That distinction is becoming important as AI touches more customer interactions, employee workflows and automated decisions. At that point, token spend starts to resemble a variable cost of doing business.

But Hassellöf’s warning goes beyond controlling that cost. Companies could create a bigger problem by optimizing the technology without redesigning the work around it.

Boards that treat AI as a token bill to be reduced risk making the same mistake as factories that installed electric motors without rethinking their production systems.

MEASURE WHAT CHANGED

The metrics companies need are therefore less about the model itself and more about what happened to the work.

Hassellöf recommends measuring cost per completed outcome, latency, exception rates, human intervention, rework and the financial value created. He also points to exception-to-resolution time and decision quality over time.

Aboul-Magd narrows his list to five: cost per agent and task, quality or success rate, cycle time, human-hours redirected from coordination to judgment, and ROI per outcome.

Tawfik argues that CFOs and CIOs need a shared framework for evaluating AI. “The CIO will naturally focus on performance, security, architecture, interoperability and sovereignty, while the CFO will consider financial predictability, return on investment and cost accountability,” he says.

“On the CFO side, it means evaluating AI based on the cost, speed and quality of achieving a business outcome rather than the price of an individual model interaction, and whether expenditure can be attributed by department, application, project or model.”

The last point is particularly important. An agent that improves as it learns within a workflow may look mediocre in a single quarterly snapshot and far more valuable over time.

Depret-Bixio focuses on the operational consequences. Companies should look at how long it takes to move from incoming data to a decision a security team trusts, how much analyst time has been removed from manual work, and what it costs to resolve an incident fully.

The common thread is simple: the number of tokens consumed tells executives very little on its own.

“If those numbers are moving the right way, the AI is paying for itself,” Depret-Bixio says. “If they’re flat, companies are probably just paying more for the same result.”

That may ultimately be the more useful way to think about AI ROI. The objective is not to build the cheapest system possible. It is to understand what the technology has changed, what that change is worth, and whether the company has redesigned itself enough to capture the value of that change.

As Hassellöf concludes, “AI ROI is not about how little intelligence you consume. It is about how much of the old constraint you remove, and whether you use that removal to build something better.”