AI Automation ROI: How to Measure It and Prove It

AI budgets are growing and finance teams are starting to ask what they bought. That question will only get louder over the next year. This is the measurement approach I use when scoping automation projects, plus what the large surveys say about who is actually seeing returns.

Most companies cannot yet show bottom-line impact

McKinsey's State of AI 2025 survey of 1,993 respondents found that 88% of organisations use AI in at least one business function, but only about 39% report any impact on enterprise-level EBIT, and most of those put it below 5%. Only about 6% qualify as "AI high performers," attributing 5% or more of EBIT to AI.

IBM's 2025 CEO Study, which surveyed 2,000 CEOs between February and April 2025, found that only 25% of AI initiatives had delivered their expected ROI over the previous few years, and only 16% had scaled enterprise-wide. Yet 85% of those CEOs expected their efficiency-focused AI investments to show positive ROI by 2027.

The next year or two is when many of those expectations get tested. Teams that can show numbers will keep their budgets. Teams that cannot are the ones Gartner has in mind when it predicts over 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs and unclear business value. The same pressure is reshaping software budgets, which I cover in AI agents vs SaaS.

What the high performers do differently

McKinsey's high performers are nearly three times as likely as others to have fundamentally redesigned their workflows when deploying AI. That matches what I see. Bolting a model onto an existing process saves a few minutes per task. Redesigning the process around what the model does well, and putting people on the exceptions, is where the larger savings come from.

An example of the difference: automating invoice data entry so a person no longer types fields is a task automation. Redesigning invoice intake so documents are photographed, extracted, validated against a schema and only the failures reach a person, as in the platform I built at Avenuor, changes who does what and how many items need human attention at all.

A simple ROI formula

Keep it boring. Finance teams trust boring.

Annual ROI = (annual value - annual cost) / total investment

Where:

• Annual value = hours saved × loaded hourly cost, plus errors avoided × cost per error, plus any revenue directly attributable (for example faster quote turnaround converting more deals).
• Annual cost = model and API usage, hosting, monitoring, maintenance, and the human time spent reviewing AI output.
• Total investment = build cost, integration, data preparation, and the time your own team spends on the project.

The most common mistake I see is leaving out the review time. If an agent drafts 500 replies a week and a person spends a minute checking each one, that is over eight hours a week of cost that belongs in the formula.

The metrics to track

Pick one or two primary metrics per workflow. Tracking everything is a way of avoiding a decision.

Count the full cost of running AI

AI has running costs that traditional software does not, and they grow with usage.

• Model usage. Agents make many model calls per task. Gartner predicts inference costs per agentic workflow will increase more than fivefold through 2028. Cap steps and tokens per task, cache where you can, and send simple tasks to smaller models.
• Coding tools. If your ROI case includes developer productivity, include the tool bills. Gartner predicts AI coding costs will surpass the average developer's salary by 2028 as token consumption grows.
• Quality assurance. Evaluation suites, monitoring and periodic human audits. These are not optional; they are what lets you trust the error-rate number. I cover them in AI agent evaluation.
• Maintenance. Model versions change, vendor APIs change, your own process changes. Budget for ongoing engineering time.

How to run a measurable pilot

1. Choose a workflow with volume

ROI is volume times savings per item. A task done 20 times a month rarely justifies a build; one done 2,000 times often does.

2. Record the baseline for two to four weeks

Time, volume, errors and cost per item. Without this, every later conversation about ROI turns into opinion.

3. Define success before you build

Write it down: "Reduce average handling time per invoice from X to Y minutes with an error rate no worse than today's." If you cannot agree on this, the project is not ready.

4. Ship a narrow first version with human review

Run the AI on real items with a person approving output. Track how often the person edits or rejects it. That edit rate is your early warning signal.

5. Compare after 60 to 90 days

Same metrics, same method. Include all running costs. Then decide: scale, adjust, or stop. Stopping a project that does not pay is a good outcome if it happens early.

Soft benefits: count them, but separately

Faster responses, happier staff and better customer experience are real, but they are hard to price. List them alongside the ROI figure rather than inside it. A business case that needs soft benefits to break even is a weak business case.

Work with me

I scope AI automation projects around a measurable outcome and build them with the logging needed to prove it. If you want a business case your finance team will accept, see workflow automation and AI integration, or book a scoping call.