Senior leaders pushing treasury and other finance teams to leverage agentic AI models must manage the risk that elevated computing costs could undermine the benefits of agents. Keeping a tight lid on expenses is hard, though, because the extreme speed of AI advances is outpacing cost controls, forcing some companies to scramble.
- At a virtual session this month, the head of technology at one NeuGroup member company said a leading AI provider had recently released a new, highly capable model that “uses 10 times-plus the token load as any other model in the system” into its enterprise platform without any advance notice, forcing the tech team to act fast to prevent costs from spiraling.
- “It just kind of appeared and we had to call them and tell them, ‘We need this turned off,’” she said. “The capability to turn it off is not even in the administrator panel. So the reactivity and the speed at which you need to move to keep pace with all the innovation is really, really important.”
- The member works for a company owned by a private equity firm participating in NeuGroup for Value Creation, which facilitates peer collaboration, structured dialogue and shared learning among portfolio company finance leaders.
The need for access control and orchestration. A NeuGroup member from another company said his FinOps team was now doing AI optimization and usage analysis, but that earlier on, managing budget limits by individual users was problematic. "It felt like [the provider] was behind on some of the ability to influence what people had access to," he said.
- The technology leader said this very real problem is becoming easier to address, noting that a leading AI provider recently released role-based access control, enabling administrators to designate usage limits and controls by role or persona. Another AI leader offers something similar, though less fully featured, she added.
- In the same vein, determining which models are used for specific tasks and how models are chained together in a workflow is part of what’s known as AI orchestration, which helps control computing expenses. It’s necessary in part because a frontier model capable of deep reasoning costs significantly more per interaction than a smaller, efficient model designed for classification or data extraction.
- People who have the choice “will typically pick the most powerful model to do the most basic of tasks," the technology leader said. "And that is a very costly way to go about things." The remedy is two-fold: A default model allows the company to direct all users to a less costly version while still allowing them to switch; for more advanced control, enterprises can build or implement an orchestration layer that constrains model selection by use case.
- “When you get into that fourth and fifth category of automation in agentic, it's really important to understand the unit cost against what it costs to deliver,” she said. “If you didn't make a thoughtful selection about the model and the orchestration, you run a very high risk of delivering something that costs more than the human who was doing the work.”
Numbers tell the story. According to slides shared at the session, AI API pricing (the per-token rate charged by the provider for model access via API) spans a roughly 600-times range across model tiers in 2026. Routing simple, high-volume tasks such as invoice classification, GL coding and data extraction to a lighter model, while reserving more powerful models for complex analysis and agentic workflows, can reduce AI spend by 80% or more. Stacking that with batch processing (a flat 50% discount for non-real-time work submitted in bulk) and so-called prompt caching (which lets you pay once to store a prompt, then a fraction of that cost on every reuse or “re-read”) can bring total cost reduction close to 95%. - Most finance-relevant workflows—month-end close document processing, bulk invoice coding, audit prep assembly—don't require real-time response, which means they're prime candidates for batch submission. And because finance runs many tasks on recurring prompts (policy rules, chart of accounts, standard templates), prompt caching pays off almost immediately: the breakeven point is just two reads.
The value equation. Amid the focus on cost, the session addressed how finance leaders can measure ROI when so much about the benefits of agentic AI is still unknown. The technology leader said not to start with cost-benefit analysis but with the outcome you are trying to achieve, whether that's growth, productivity or cost reduction. Then track time. Her team knows, at a "time and motion" level, how long specific tasks take before AI assistance. That baseline makes the productivity gain visible, even when it doesn't yet translate to a hard-dollar number.
- "Right now, we're in the religion category," she said. "You've got to have faith that the learning, the enablement, and all these things will yield value later. You're really building muscle."
That muscle is being built across meaningful territory. According to a slide presented, areas where finance AI is already delivering real results include:
- FP&A and forecasting. AI-assisted variance narratives and rolling forecast updates cut cycle time dramatically. Teams that previously spent days on commentary are generating first drafts in hours, freeing analyst time for the judgment calls that matter.
- Month-end close. Automated reconciliation exception flagging, GL coding and journal entry review are compressing close cycles. Early adopters report 20% to 30% reductions in manual close hours.
- AP and procurement. Invoice matching, PO validation and duplicate detection running without human review, with freed capacity redeploying to vendor negotiation and contract management.
- Audit and compliance prep. AI-assisted document assembly and audit trail summarization compress preparation time; finance teams arrive at fieldwork with organized, complete packages.
- M&A due diligence. Contract review, revenue recognition analysis and normalization adjustments benefit from AI-assisted document processing. For companies running frequent acquisitions, the compounding returns are significant.
- Board and LP reporting. AI generates first-draft commentary on actuals vs. budget, cash position and KPI narratives from structured data. Review becomes editing, not writing.
