How To Track Your ROI on AI?

The AI ROI paradox

AI is accelerating business investment: more bets, made faster, with more uncertainty. Yet most of those decisions still live in spreadsheets, email threads and disconnected workflows, with little learning loop once the money is spent.

The pressure is real. Around 87% of finance leaders say they need to tie AI spend to outcomes within a year, but only about 22% say they can do that today. AI clearly improves productivity. The hard question is how that productivity becomes ROI.

The measurement gap

Research published in 2026 shows the gap clearly. McKinsey's State of AI ^[1] finds that 80% of organizations say AI has improved individual productivity, and Deloitte reports that 66% see productivity or efficiency benefits ^[2].

But far fewer see it in the financials. Only 37% attribute any operating profit (EBIT) impact to AI, just 6% qualify as AI high performers with at least 5% EBIT impact, and only 20% report increased revenue ^[1][2]. PwC finds that 74% of AI-driven returns are captured by the top 20% of companies ^[3]. The difference between those groups is rarely the model. It is how the value is measured and owned.

Four questions every AI return must answer

What makes the return credible? A realized, concrete change in operating metrics, translated into ROI.

What did it really cost? Implementation, ongoing operation and human oversight, not just licenses.

Did AI cause the improvement? You need evidence that separates the effect of AI from other changes happening at the same time.

When does value arrive? Delivery timing and the adoption ramp decide when benefits actually show up. The rule of thumb: same baseline, same period, full costs.

Productivity is not ROI until someone owns it

AI creates capacity. Someone has to turn that capacity into money. Faster support only becomes ROI when it leads to avoided hiring or lower outsourcing costs. Faster development only matters when it produces earlier releases with measurable business impact. Better sales support counts when it shows up as higher conversion and contribution profit.

That is why ownership has to be decided at approval, not after launch. Name the owners of each benefit when the investment is approved, then compare their commitments with results once it is live.

What real-world cases teach us

Klarna reported roughly 700 agent-equivalents of capacity from its AI assistant and claimed about $40M in annual profit improvement for 2024 ^[4], then later put human service quality back at the center of its operating model. Company claims are useful signals, not controlled attribution: test quality, repeat contacts and retention alongside cost.

Controlled studies give the strongest causal evidence, built on four steps: a baseline on the same task, a control group of comparable people without AI, outcomes measured as time plus quality plus completion, and clear boundaries on where AI helps and fails. Even so, a credible lift is not automatically a business case. It still has to be translated through workflow volume, adoption, quality guardrails and full cost.

Morgan Stanley reported about 98% adoption across advisor teams by putting trusted internal research into the daily workflow, with citations, feedback and human review ^[5]. It proves distribution and governance drive adoption, but not a public dollar ROI. JPMorgan has publicly discussed more than $1.5B in annual business value across AI and ML initiatives ^[6], managed like a portfolio where each use case moves from identified to validated to realized.

Measure the right things

Start with a baseline of volume, time, cost and quality before anything is deployed, and agree on it up front. Ask the key question early: are we comparing AI with today's costs, or with what would have happened without AI?

Track adoption and operating outcomes, and always count full cost: build, run, change and risk. Above all, pair efficiency with quality every time. A cheaper process that drives repeat contacts or churn is not a return.

Never mix proven and speculative value

Realized value and projected value should never sit on the same line. Keep them in separate columns, with separate owners and separate confidence levels. Like JPMorgan's portfolio approach, give each use case a baseline, an owner, a benefit type, a realization status and a review date.

Seven disciplines that turn AI claims into ROI

1. Baseline first. 2. Use a control group. 3. Measure quality and efficiency together. 4. Count full costs. 5. Name owners. 6. Report realized and projected value separately. 7. Use conservative assumptions.

A worked example: $2M for AI customer support

Picture a proposal to invest $2.0M in an AI customer support transformation, promising a 10% reduction in support costs, with 80% adoption and a six-month implementation. Should Finance approve the rollout, approve a limited pilot with milestones, ask for more evidence, or decline?

The answer depends on the disciplines above: an agreed baseline for handle time and CSAT, a clear path from productivity to value (lower cost, avoided hires, serving more demand or better customer outcomes), named owners for each benefit, and realistic ROI and payback math. Strategic considerations matter too, but they should be stated, not hidden inside the numbers.

Who really owns AI ROI?

Finance owns the numbers, Product owns the value and Engineering owns execution. ROI sits at the intersection of all three, which means it is everyone's job, and no one's, until teams work together. Without alignment, great ideas die in spreadsheets.

Uprali is built for exactly this: a shared place where Finance, Product, Engineering and business teams build and evaluate AI investment cases together, prioritize and approve the right bets, and track realized results against the original assumptions. We are looking for design partners. Join the pilot to make ROI a team sport.