The headline from MIT’s State of AI in Business 2025—that roughly 95% of organizations see no measurable return on their generative AI investments—has made some leaders consider tapping the brakes. That’s the wrong conclusion, especially for the finance function. MIT’s analysis shows the problem isn’t AI models; it’s the approach. Companies that reap the biggest benefits from AI embed it into real workflows, ensure systems learn from feedback, and measure business outcomes.

The earliest and clearest returns from AI are showing up in support functions such as finance, procurement, and operations through reduced external spending, faster cycles, and tighter controls, even without broad headcount cuts, according to the MIT report. Finance processes tend to be repeatable, data-rich, and policy-bound—the very conditions in which generative AI delivers substantial benefits.

What’s more, budgets remain skewed toward visible, top-line pilots, while back-office automation—where the payback is often faster—remains underfunded. For CFOs, the report’s message isn’t “spend less on gen AI.” It’s “spend differently.” Aim more of the next dollar at finance use cases that affect cash, cost, and risk. Don’t buy or build AI tools for what they do but for the business outcomes they deliver.

Embedding AI in workflows

Anthropic’s Economic Index offers a detailed look at how enterprises actually use large language models in production. When models are embedded via application programming interfaces (APIs) rather than chat interfaces, companies don’t “converse,” they delegate. Anthropic reports that 77% of its enterprise API usage is linked to automating tasks, and that share is rising over time as users shift from iterative prompting to clear, directive commands (see Figure 1).

Figure 1

That behavioral shift is key for finance. It suggests that the most effective use of enterprise AI is automation, not chat-related assistance. Automation is the best way to improve straight-through processing, working capital performance, and cycle times.

Anthropic’s second major finding reinforces this point: The main limitation isn’t price; it’s data that provides context. When companies give AI complex tasks, they tend to provide much longer inputs. But the payoff is limited: Every 1% increase in input length yields only about 0.38% more output. In other words, making inputs longer doesn’t boost results much. The study also found that demand doesn’t change significantly when token prices change, suggesting that cost isn’t what’s holding back adoption. Rather, it’s how well AI can handle complex or detailed information.

In finance, the main challenge is building a solid data foundation. That means creating well-managed data products for things like the chart of accounts, vendor and customer records, and policy documents. These need to connect smoothly with enterprise resource planning (ERP) and enterprise performance management (EPM) systems, while ensuring data is retrieved securely and within proper boundaries—so the systems can operate seamlessly and stay compliant with the Sarbanes-Oxley Act (SOX) and International Financial Reporting Standards (IFRS).

In our experience, leading companies are moving from application-centric systems to data-driven platforms, then to intelligent orchestration, and, ultimately, to AI agents that manage multistep workflows (see Figure 2). The destination is a finance function that operates continuously—where closing, forecasting, and review processes run in real time and human attention is needed only for exceptions. That goal becomes possible when AI is embedded directly in the flow of work and is powered by clean, trusted data.

ROI in finance

Finance AI use cases with the highest returns share three traits: They’re bounded, repeatable, and tied to financial decisions that create value.

Accounts payable and receivable are becoming AI-powered engines of efficiency. In payables, AI now captures and classifies invoices, matches them to purchase orders, applies tolerances, auto-approves entries, and flags only the exceptions for review. On the receivables side, AI parses remittances, matches payments to open invoices, posts them automatically, and segments collections by risk—freeing up teams to focus where it matters most. When paired with APIs or robotic process automation, these agentic designs shorten cycle times and scale throughput without sacrificing control.

Management reporting is another proven win for automation. It’s highly repeatable and template-driven, making it ideal for AI to generate tables, charts, footnotes, and narrative commentary. Even modest adoption delivers returns, given the time these tasks normally consume. By contrast, general “finance knowledge bots”—tools that index everything—require extremely high adoption to deliver returns. They are low volume, require broad seat licensing, and demand extensive human oversight. In most cases, the people and license costs far outweigh the technology expense.

The rule of thumb: Focus AI on critical business decisions that affect cash, margin, or risk (for example, a procure-to-pay leakage agent that flags off-contract rates or duplicate billing) rather than diffuse Q&A tools that answer everything.

Not all copilots deliver enterprise value out of the box. These assistant-style tools can boost analyst productivity—especially in spreadsheets—but without integration to systems of record, they rarely change cost structures or improve compliance. Similarly, processes that rely on scattered or tacit knowledge, such as bespoke revenue recognition across one-off contracts, tend to underperform until their data is cleaned and connected, MIT’s report suggests.

Achieving more value from AI

Design for automation, not just assistance. Anthropic’s enterprise data shows that once models are embedded, users shift from cocreating outputs to fully delegating tasks. In finance, that means building straight-through workflows with confidence thresholds, approval logic, and audit trails. The goal isn’t better spreadsheets; it’s measurable gains in touchless processing, faster cycle times, and more effective control.

Build the context. For finance AI, the real constraint isn’t token cost but whether the context is ready. That means aligning on shared definitions for the chart of accounts and master data, integrating ERP and EPM systems with document repositories under a policy-aware retrieval layer, and embedding telemetry across every AI-enabled process. Every action should be traceable, with automated logs that show which data was used, the rules or models applied, and who approved or overrode decisions. This not only ensures compliance under SOX and IFRS—it also builds the trust finance needs to scale AI with confidence.

Focus on outcomes and partner to scale. MIT’s report shows that AI solutions built with external partners are nearly twice as likely to reach deployment as those built in-house. Successful CFOs hold vendors accountable for results such as touchless processing rates, faster close cycles, fewer exceptions, and lower external spending. They also insist on systems that get smarter with use. The key: Measure impact based on business outcomes, not in prompts used or hours saved.

Tie AI to finance modernization. Gen AI is an accelerator, not a shortcut. To deliver optimal results, it needs to move in step with broader efforts like data cleanup, workflow design, and process modernization. That’s how you create a seamless flow from insight to decision to action. Anchor AI programs to a multiyear roadmap, including continuous close, rolling forecasts, and real-time working-capital control.

Leading finance executives already are wielding AI to shape how decisions are made, how work gets done, and how finance creates value. The opportunity isn’t in experimenting for its own sake—it’s in embedding AI into core finance processes. Early adopters who link gen AI to a modernization agenda will see benefits increase over time as their data, governance, and automation foundations strengthen. In a rapidly automating landscape, those strengths will become a key competitive edge.