AI vendors say companies are measuring ROI the wrong way. What will convince CFOs to keep paying?
Artificial intelligence companies want their customers to use more AI. Their customers want AI to help them spend less money. That creates a tricky dynamic for an industry whose revenue depends on selling more of a technology that’s supposed to make businesses more efficient. As CFOs prepare their 2027 budgets, the question is becoming harder to avoid: Is all this AI spending actually paying off?
The industry’s preferred measures of progress offer an incomplete answer. Salesforce reports that its technology has performed 7 billion agentic work units, which count activities such as invoking a tool or completing a reasoning chain. A single business outcome can require many such units. Meanwhile, ServiceNow told investors on October 28 that its internal AI use generates more than $1 billion in annualized value, double its estimate at the end of 2025.
Those figures help vendors make their case to investors but offer buyers less certainty about what they’re paying for. More automated activity can mean useful work—or repeated attempts to complete the same task.
At Infosys, that tension is already affecting revenue expectations. AI revenue rose 29% from the previous quarter and reached 10.5% of sales, yet the company cut its full-year growth outlook to 1%-2% amid pricing pressure and demands for greater productivity. When AI lets a client buy the same service for less, the savings strengthen the client’s case for investing in AI while reducing the supplier’s contract value. New AI business must make up for the revenue lost as existing work becomes cheaper.
In a survey of 260 finance executives, the cloud cost management company CloudZero found that two-thirds said their boards required proof of returns before approving further AI investment. Yet a separate survey from Protiviti, a management consultancy, found that only 35% of finance organizations considered themselves effective at measuring AI returns.
In conversations with Fast Company, enterprise AI leaders described how competing definitions of success are shaping pricing and contract renewals. Outcome-based pricing offers little protection when a vendor counts a problem as resolved but leaves the customer with more work. Higher consumption bills are difficult to justify when a cheaper model delivers the same result. To secure another year of spending, suppliers may need to show customers which expenses they can eliminate, even when that means reducing their own fees.
Snowflake tells some customers to spend less on AI
Snowflake’s CEO, Sridhar Ramaswamy, argues that rising AI consumption is worthwhile only when it leads to operational improvements and replaces existing costs. Snowflake reported $1.49 billion in product revenue for its second fiscal quarter of 2027, up 37% from a year earlier, and 126% net revenue retention.
“Rather than celebrating consumption, I look for shrinking backlogs, faster ship cycles, and teams deploying previously impossible capabilities,” Ramaswamy tells Fast Company. He says companies should be eliminating expenses such as contractor fees, agency retainers, and redundant software as their AI use grows.
That thinking proved important during a contract renewal involving an enterprise customer’s inventory operations. Snowflake’s automated agents had reduced a two-week reconciliation process involving six analysts to just two hours. Yet the customer’s finance team initially resisted renewing.
“Our team was celebrating query volume and developer throughput, but the CFO did not care about consumption metrics,” Ramaswamy says. The deal closed after Snowflake demonstrated that the technology could prevent millions in obsolete product markdowns, retire two legacy business intelligence tools, and eliminate escalations from store managers. “Technical velocity is an internal metric,” he says. “Finance buys hard cost displacement and operational certainty.”
Ramaswamy says the same principle has led him to advise customers to scale back popular AI workloads. At one financial services company, employees were using expensive frontier reasoning models to handle routine support tickets and extract information from documents.
Snowflake recommended routing those tasks to smaller open-weight models through its Cortex AI Gateway, reserving frontier models for complex reasoning. The change cut workload costs by more than half without sacrificing accuracy, according to Ramaswamy.
“If a cheaper model gets the exact same job done with zero loss in fidelity, running the most expensive model on the market is just bad business,” he says.
Who decides what counts as a paid outcome at Zendesk?
Tom Eggemeier, CEO of Zendesk, sees outcome-based pricing as a way to tie spending more closely to the work a service organization completes. The company reports nearly 20,000 AI customers, and its AI annual recurring revenue surpassed $200 million in 2025; expectations are that this will exceed $400 million in 2026.
But deciding what constitutes a successful outcome is not straightforward. Zendesk uses a separate AI model to evaluate automated resolutions, and Eggemeier says closing a ticket alone does not necessarily mean an interaction qualifies for payment.
“Service has outgrown the ticket. A customer’s issue can span channels and multiple conversations, so an outcome should be tied to eligible work on the issue, not counted once for every message or contact,” he says.
Customers can trace charges to individual interactions and dispute classifications after a ticket closes. Zendesk’s broader Outcome AI Platform Plans remain in early access, and Eggemeier says the company is still analyzing how they work in practice.
For finance teams, the important question is whether automation reduces the cost of serving customers. Eggemeier says that calculation needs to account for repeat contacts, escalations, service quality, and the human work required to resolve an issue. Zendesk’s dashboards can provide some of that information, but they do not automatically calculate total cost to serve.
Under Zendesk’s pricing structure, the contracted price for an eligible outcome remains the same whether AI resolves the issue independently or a person helps. An AI-assisted handoff does not trigger a second charge.
“Pricing is tied to eligible work and the customer’s plan, rather than directly to the cost of running the model,” Eggemeier says. “So lower AI costs don’t automatically change the contracted price.”
Atomicwork counts an AI fix only when work resumes
Vijay Rayapati, cofounder and CEO of Atomicwork, takes a narrower view of what qualifies as a completed AI resolution.
“A closed ticket is an administrative event. An employee getting back to work is the outcome. If AI answers in 10 seconds but leaves someone with 30 minutes of troubleshooting, we’ve moved the work onto the employee,” he says.
Atomicwork and its customers agree on what constitutes a completed task before measuring AI performance. Answering a policy question correctly may be sufficient, but an access request is not complete until the employee can actually use the system. Providing instructions alone does not qualify as an autonomous resolution.
Atomicwork charges $1 to $3 per completed outcome, with annual minimums, and also offers a $25,000 annual platform tier. Its definition of autonomous resolution differs from Zendesk’s broader plans, which can count eligible outcomes even when humans help complete them.
Rayapati argues that outcome-based pricing can correct a longstanding problem with enterprise software: Vendors get paid for seats and usage regardless of whether customers see much value.
The company also guarantees customers migrating from ServiceNow a 50% reduction in total cost of ownership, a calculation that includes operating and maintaining the platform rather than just purchasing it.
“If we’re going to guarantee savings, customers should be able to hold us to the numbers,” Rayapati says.
Rackspace measures AI returns as deployments mature
Gajen Kandiah, CEO of Rackspace, argues that measuring AI returns requires accounting for the human work needed to keep systems running after deployment. Rather than focusing solely on how many tasks agents perform, he wants to know how often engineers must intervene.
“An agent that looks brilliant while an engineer quietly fixes its output isn’t doing the work. The engineer is,” he says.
Rackspace divides engineering time into two categories: “build,” for lasting improvements to a system, and “babysit,” for one-time corrections. Its goal is to reduce the latter as deployments mature. Required human approvals in regulated settings are tracked separately, since they are part of the service rather than evidence that an agent has failed.
During an early-stage deployment for a large U.S. healthcare provider, Rackspace found that engineers spent most of their time improving source data rather than correcting the AI itself. As the data improved, each update cycle required less engineering support. Kandiah says incorporating user feedback into an agent’s instructions and test cases can help prevent recurring problems.
Rackspace also sometimes recommends less AI than customers initially request. One company wanted large language models to run an entire master data management system covering customer and product records. Rackspace instead recommended traditional software for routine processing, reserving AI for cases where structured rules needed additional help.
The approach made operating costs more predictable but also reduced the amount of AI work Rackspace would perform.
“We have to be willing to help a customer spend less with us,” Kandiah says. “It affects our revenue, and I do not assume every dollar comes back through a new AI engagement.”
Rackspace is making similar trade-offs in its own business. In July, it lowered its annual revenue outlook by $150 million as it shifted away from lower-margin work and invested in AI. Its third-quarter results subsequently prompted a $35 million increase in the midpoint of its full-year revenue outlook.
The company also announced AI infrastructure agreements with Together AI and Blacksmith worth approximately $610 million in anticipated total contract value, subject to the agreements’ terms. Quarterly revenue totaled $609 million, down 9% from a year earlier, while adjusted operating profit rose 12%. Rackspace reported a $54 million net loss. Capital spending increased to $104 million from $34 million a year earlier, alongside $36 million in quarterly operating cash flow.
Kandiah says the company’s challenge now is turning its investments into sustained revenue, not simply announcing more AI capacity. Customers face a similar test: A lower infrastructure bill may free up money for additional investment, but implementation and support must still factor into the calculation.
“As delivery gets more efficient, our contracts and our own cost base have to follow,” he says. “Keeping charges in place for work a customer no longer needs is hard to defend.”
Reltio and Anthropic want AI measured against a baseline
Ansh Kanwar, chief product officer at Reltio, an SAP company, says AI costs can reflect the quality of the underlying data. When enterprise information is fragmented or poorly organized, agents may need more expensive models, additional tokens, and more time to answer the same questions.
Kanwar recommends comparing the cost, speed, and accuracy of identical AI tasks using fragmented versus unified data. That allows companies to distinguish improvements attributable to better data from those attributable to the AI itself.
At Anthropic, Kate Jensen, head of Americas, says companies are increasingly choosing AI projects based on specific business problems rather than the technology’s capabilities. The AI giant’s S-1 registration statement for its IPO reports $17 billion in third-quarter revenue and about 900% net revenue retention, excluding consumer plans.
Jensen recommends agreeing on what success looks like before deploying AI, then comparing the results with what a team could have accomplished without it. Hours saved are worth tracking, she says, but business outcomes and newly achievable work offer a more meaningful measure.
“As Claude takes on bigger projects, every leader should be able to say what got done, how well it was done, and where they would have been without the help of Claude,” she says.
McKinsey sees a gap between AI productivity and financial returns
Michael Chui, a senior fellow at QuantumBlack, McKinsey’s AI business, says improvements in individual productivity do not necessarily translate into financial gains for an entire organization.
McKinsey’s 2026 State of AI survey, covering 1,719 respondents across 97 countries, found that 80% said AI improved their personal productivity. Yet only 37% reported that AI contributed to their organization’s earnings before interest and taxes (EBIT), essentially unchanged from 2025. About 6% attributed at least 5% of EBIT to AI. Those figures measure different levels of impact and do not necessarily mean individual companies’ returns have stopped improving.
Chui argues that meaningful financial gains often require changing entire workflows or business functions rather than making isolated tasks faster. And even when those changes work, the results may take months to appear.
A sales team, for example, might use AI to prepare proposals faster, but deals could still take six months to close. In the meantime, finance teams can track whether more proposals reach short lists or advance through procurement, rather than judging the investment solely by the volume of proposals produced.
Those interim measures can help CFOs decide whether an investment is on track before its ultimate financial results become clear.
At the next renewal, CFOs can spell out what another six months of spending must deliver. But the contract terms also determine who gets to define success. If an unmet business target is replaced by a more flattering usage metric, each renewal risks becoming another paid experiment. Vendors asking customers for patience need to show how progress will translate into financial returns—and when buyers can expect to see them. Eventually, that evidence must inform what a customer is willing to pay, or whether it renews at all.
内容来自 Fast Company,版权归原作者所有。
阅读原文 ↗