CFOs could cut agent costs

Fortune reported on May 19 that CFOs may be able to cut agentic-AI costs by as much as 60% by fixing a basic enterprise-data problem: systems that store information without enough business context. Gartner research presented at the company’s Data & Analytics Summit in London said companies that prioritize semantics in AI-ready data can improve agentic-AI accuracy by up to 80% and reduce costs by up to 60% by 2027. That framing shifts the cost debate away from model prices alone. Gartner analyst Rita Sallam said agent outcomes depend on “semantic representations of data,” and warned that without clear relationships and rules inside company data, agents cannot operate accurately. Fortune, citing Gartner’s findings, said the problem is not only hallucinations or bias from models, but unreliable outputs produced when enterprise data lacks context. (finance.yahoo.com) ### Why are finance chiefs being told to look at semantics instead of model pricing? Gartner’s argument is that traditional schema-based data models are no longer enough for agentic systems. Sallam said a semantic, or “context,” layer needs to sit at the core of enterprise data infrastructure so agents understand what fields, entities and rules actually mean inside a business. (finance.yahoo.com) For CFOs, that makes agent spending look less like a pure software-procurement issue and more like a capital-allocation problem. Fortune said Gartner views semantic coherence as a cost-control and trust strategy, with possible relevance for regulators and audit committees reviewing how AI-generated outputs feed into reporting and disclosures. (finance.yahoo.com) ### Where do semantic failures actually make agents more expensive? The immediate cost shows up when agents have to work across multiple enterprise systems. If customer, product, contract or ledger data is defined differently in different applications, agents either fail, require more human review, or trigger rework downstream. Gartner’s warning, as relayed by Fortune, is that companies can keep spending on tools that appear powerful but still produce unreliable results because the underlying data lacks context. (finance.yahoo.com) PwC and other consulting firms are describing a similar problem from the deployment side. Forbes reported on May 19 that PwC launched a tool called “agentic scaffolding” to help enterprises design, simulate, visualize, stress-test and stand up agent-driven processes, with the company positioning change management as part of the implementation challenge rather than something handled after rollout. (finance.yahoo.com) ### Why does change management keep coming up in agent deployments? Forbes reported that PwC partner Rima Safari said “every major AI transformation” PwC is taking on with clients is using scaffolding to drive operating-model change. The publication described the tool as an application layer that visually orchestrates AI workflows while generating code needed to execute and manage those agents. (forbes.com) That matters because agents do not stay fixed after launch. Forbes said PwC’s approach reflects a broader shift in which enterprises need agile change-management processes as pilots move into production. A 2025 MIT report cited by Forbes found that 95% of AI pilots failed to deliver notable return on investment, which the article linked in part to weak integration with enterprise workflows. (forbes.com) ### How are large firms trying to operationalize that view? PwC has been building finance-focused agent programs with multiple model providers. On May 14, PwC and Anthropic said they were expanding their alliance, establishing a joint Center of Excellence and training and certifying 30,000 PwC professionals on Claude, while PwC said its Office of the CFO would be the first at-scale expression of that work. (forbes.com) On May 6, CFO Dive reported that OpenAI and PwC were building agents for planning, forecasting, reporting, procurement, payments, treasury, tax and accounting, with PwC saying the work was aimed at “practical, high-value workflows” under human supervision. In April, Sage and PwC also announced an agentic-AI delivery model for Sage Intacct implementations designed to reduce manual effort in design, configuration and testing. (pwc.com) ### So where does the spending move if companies follow this advice? The likely budget shift, based on Gartner’s and PwC’s descriptions, is toward semantic layers, workflow orchestration, testing, controls and change-management tooling rather than only toward cheaper inference. That is an inference from the sources, not a direct quote, but it follows from Gartner’s focus on context and PwC’s emphasis on implementation infrastructure. (cfodive.com) McKinsey made a related point in 2025 when senior partner Jorge Amar said companies were deploying agents first in “very deterministic” environments with clear processes, including IT help desks, software development and customer service tickets. That suggests the companies most likely to contain costs are the ones that narrow scope, define processes clearly and keep humans in the loop as agent programs expand. (finance.yahoo.com) PwC’s next public milestones are already on the calendar. The firm said on May 14 that it would roll out Claude Code and Cowork starting with U.S. teams and expand training toward a global workforce of hundreds of thousands, while its finance-focused work with OpenAI and implementation work with Sage continue across named enterprise workflows. (pwc.com) (mckinsey.com)