As enterprise AI budgets continue to rise, so will the positive results–as long as companies pursue endeavors that are narrow, measurable, and auditable - and not moonshots.
This past summer's research revealed the gap between proof of concept AI endeavors that never touch real systems and those that are actually using GenAI in production, which means they're wired into enterprise workflows.
In the first instance, MIT research found that 95% of enterprise GenAI efforts are not producing measurable returns on investment for companies. In the second instance, Google's global ROI of AI survey reports that 74% of entities achieve ROI on at least one use case and, among agentic early adopters, that jumps to 88%.
So, rather than draw conclusions of AI bust in one case and AI boom in another, it pays to look deeper at what is being measured. In both cases, it is clear that pilots and proof of concept ideas often stall–which is no surprise–while endeavors that wire AI into workflows will often succeed. What makes the difference is systems engineering plus governance.
Even the MIT results, which rattled investors and executives alike, found that GenAI failures often involved a lack of integration into workflows and poor strategic alignment. Meanwhile, successful efforts included robust data integration, alignment with business needs, and clear processes for AI deployment.
AI in the Finance Function
For CFOs, this is all good news. The finance function lends itself to automation and to agentic AI that can decide and do things, like orchestrate steps, call enterprise APIs, and hand off or escalate via human guardrails.
That's because finance operations are often structured, policy-heavy, and instrumented for controls. This is a great proving ground for agentic AI without existential risk. I see agentic AI making early and successful inroads in four common finance workflows:
-
Closing the books. This has long been a finance job that's taken too much time and resources. AI agents can speed this up by collecting missing receipts and documents, coding expenses, and flagging anomalies. They can do this continuously, escalating alarms as closing time approaches, all while writing back to the system of record. No one in your team will miss this work.
-
Purchasing. Employees can now make plain-language requests in Slack or Microsoft Teams to purchase whatever is needed and unleash an AI agent to do the legwork. To ensure the right legwork, make sure the request is compliant with company policy and budgets and first routed to the right approvers for a gut check.
-
Travel. Companies with clear travel policies–married with traveler preferences–are ripe to use AI agents to purchase, then package a trip report and group expenses for one-click approval before exporting to company databases, where things can later be audited for compliance and other checks.
-
Payments support. Finance teams aren't making the most of their brainpower if they're stuck answering employee questions on such things as card blockages and expense reimbursement statuses. A trained AI agent can pick up that workload, leaving your highly trained finance staff for higher level work.
Staying on Track
Still, the risk is real - as MIT's research shows - that enterprises can get off the ROI beam with AI. And that can be real money lost, especially given predictions from IDC that agentic AI will take 26% of IT budgets within five years.
To ensure that doesn't happen:
Plug agents into source-of-truth systems. This means things like ERP, ticketing, CRM, and policy engines so the agents can deliver CFO-grade proof results. The right metric to look for is "rate of closed-loop automations per quarter that meet policy and pass audit," plus the dollars attached.
Give agents tools. To deliver ROI, agents need access to tools and data so they can open cases, update records, create tickets, post journal entries, and follow policy. Google emphasizes secure access to internal systems and its guidance is blunt: give agents governed access to enterprise systems and write the rulebook early. With good governance first, good performance will more likely follow.
Secure executive support. The companies reporting AI success will no doubt have strong executive support that demands a clear definition of value upfront around such things as speed, accuracy, cost, and or revenue.
Agentic AI as Participant
In and outside the finance function, the potential for ROI with AI is centering around the arenas of productivity, marketing, security, and customer experience. Most likely, it will take less time to achieve ROI when use cases are repeatable and the data informing the use case is clean and easily reachable. Security is emerging as a good domain for AI agents, in part, because the work is event driven and tool heavy.
All in all, enterprises that treat AI agents as transaction participants–and not as toys– will achieve the operational leverage they need to further increase investment. With more wins and institutional knowledge on how to do AI right, finance and other teams will successfully achieve ROI with more AI in more and more ways.
