77% of Finance Orgs Deployed AI. Only 35% Can Tell You What It Returned.

A Protiviti survey of 902 finance leaders found that 77% of finance organizations had deployed AI, but only 35% said they could confidently gauge its return. That gap isn't a measurement problem. It's a measurement design problem. And it's sitting in your next budget review.

If you're defending AI spend to your CFO, the question isn't whether AI "works." It's whether you can show what it produced in terms Finance will recognize, at a cost that includes everything you actually spent. Most teams can't. A DoiT survey of 500 finance leaders reported that only 15% could calculate AI ROI without significant bottlenecks, citing differing definitions of success between Finance and Engineering (37%) and unclear financial attribution (36%) as leading causes. Meanwhile, 79% reported cost overruns.

The Three-Bucket Problem

Most AI ROI narratives collapse because they mix unlike things. A defensible framework separates outcomes into three buckets: incremental revenue (or gross profit), realized cost savings (spend that actually decreased), and productivity or capacity gains (hours freed and redeployed). The critical word is "redeployed." If headcount stays flat and vendor invoices don't shrink, you haven't saved money. You've created capacity. That's real, but it's different, and Finance knows the difference.

An editor who spends 10 fewer hours per week on first drafts didn't save the company $500. She has 10 hours available. The ROI story starts when those hours produce something measurable: more published content, faster campaign turnaround, or a new workflow that wasn't possible before. Without a redeployment plan and evidence it worked, "time saved" is a capacity claim dressed up as cash impact.

Your Cost Sheet Is Probably Incomplete

Token usage, third-party integrations, implementation labor, training, governance overhead, and the human time spent reviewing AI outputs: these all belong on the cost side. Trintech CFO Omar Choucair flagged the risk that a small number of heavy users can drive disproportionate token consumption, a variable cost that's easy to miss until the invoice arrives. He also stressed building verification into AI workflows and ensuring employees can explain outputs to auditors. If your team can't explain how an AI-generated recommendation was produced, auditors and Finance will discount the result.

A PPC Land summary of TransUnion findings showed that 65% of surveyed marketers primarily measured AI impact through estimated time or cost savings, while only 35% used formal lift tests or experiments. Estimated savings are a hypothesis. Lift tests are evidence.

Before-and-After Isn't Proof

A before-and-after comparison shows correlation, not causation. Your pipeline improved after deploying an AI lead-scoring model, but you also changed your targeting, swapped the landing page, and ran a new offer during the same period. Which lever moved the number?

Stronger approaches use randomized holdouts or matched comparisons and document other changes that could explain results. The hypothesis format works here: "If we deploy AI-assisted lead scoring on Segment A while holding Segment B as a control, then SQL conversion rate will increase by X% because the model surfaces higher-intent signals earlier in the funnel." If you can't run a holdout, at least document every concurrent change so Finance can assess what else might have caused the shift.

Run It This Week: A Workflow-Level ROI Pilot

Setup: Pick one workflow with a single owner (an editorial queue, a paid campaign, a lead-scoring model). Define what counts as accepted output before you start: an editor-approved article, a sales-accepted lead, an opportunity that progressed past discovery.

Baseline: Save samples of completed work with review history and elapsed time from request to acceptance. Use a period representative of normal workload.

Cost sheet: Subscriptions plus usage charges, plus implementation time, plus human review and rework hours. Add a "Rework Reason" field. Repeated corrections will tell you where the workflow needs repair, something a clean subscription total won't show.

Measurement: Separate capacity (hours freed) from financial benefit (revenue earned or spend reduced). If hours are freed, record what task received them. Link to the output. Until that output produces a measured result, report the hours as capacity recovered, not savings.

The hypothesis: If we run AI-assisted [workflow] for [time period], then [primary metric] will improve by [X%] because [mechanism]. Success = [primary metric]; guardrails = [quality metric doesn't degrade]; stop-loss = [if cost per accepted output exceeds baseline by 20%, pause and diagnose].

Where the Real Returns Hide

KPMG surveyed 1,013 senior finance leaders across 20 countries and found that among those dissatisfied with AI ROI, the leading barrier was slow organizational adoption and change management, not the technology itself. The binding constraint on AI returns is often the operating model around the tool, not the tool. A Deloitte survey of 1,326 global finance leaders found that among respondents actively using AI, 21% believed their investments had already delivered clear, measurable value. That's a signal that most organizations are still in the measurement-design phase, and the ones who get measurement right first will scale spending with confidence.