How to Measure AI ROI: A Practical Framework for Growing Businesses

Measure AI return workflow by workflow, not tool by tool. For each job AI does, record a baseline on a number the business already tracks, count the full cost, then compare after a fixed period. Hours saved count only when they are spent on something measured. Most businesses fail to prove a return because they never set the baseline.

What is AI ROI?

AI ROI is the return a business gets from AI compared with what it spends to get it. The usual formula is net benefit divided by total cost, shown as a percentage. The hard part is not the arithmetic. It is deciding what counts as benefit, what counts as cost, and what would have happened without AI.

The formula looks simple:

AI ROI = (value gained minus total cost) ÷ total cost

In practice, both halves are slippery. Value might be revenue won, costs avoided, a hire not made or work done faster. Cost includes far more than the subscription. And without a record of how the work performed before, any number you produce is a guess.

That is why so many leaders struggle with it. According to Microsoft and LinkedIn's 2024 Work Trend Index, 59% of leaders say they are concerned about quantifying the productivity gains from AI. The same survey found 60% worry their company lacks a vision and plan for AI. The two problems are linked: it is hard to measure a return on something with no defined purpose.

Why is AI ROI so hard to measure?

AI ROI is hard to measure because AI usually arrives as tools rather than projects. People use it across dozens of small tasks, nobody records a baseline, and the benefit is spread thinly. Add self-reported time savings that may be wrong, and the result is spend that feels useful but cannot be proved.

Three problems recur.

No baseline. Few businesses record how long a proposal took, or how many leads a campaign produced, before AI was involved. Without that, there is nothing to compare against.

Diffuse use. When everyone uses AI a little, for everything, the benefit is real but scattered. It never shows up as a line in the accounts.

Unreliable estimates. People are poor judges of their own time savings. In a randomised trial published by METR in 2025, 16 experienced software developers took 19% longer on real tasks when allowed AI tools. Yet they had forecast a 24% speed-up, and afterwards still believed AI had made them 20% faster. That is one study, in one setting, with early-2025 tools. However, it shows why "how much time did AI save you?" is a weak survey question.

The scale of the problem shows in the wider data. According to McKinsey's 2025 State of AI survey, 88% of organisations use AI in at least one function, but only 39% report any earnings impact at the enterprise level.

What should you count as the cost of AI?

Count everything it takes to get the result, not just licences. That includes subscriptions and usage fees, setup and integration, the time people spend learning and checking AI output, outside help, and the ongoing effort to keep context and workflows current. Leaving any of these out flatters the return.

A practical cost list:

• Licences and usage. Seats, API or usage charges, and any add-on tools.
• Setup. Connecting systems, writing down methods, building the shared context AI works from.
• People's time. Training, and the time spent checking and correcting AI output.
• Outside help. Consultants, agencies or contractors.
• Upkeep. Keeping the context current and fixing workflows when tools change.

The checking time matters most and is most often ignored. If an AI draft takes ten minutes to produce and forty to fix, the saving is smaller than it looks. The guide to the hidden costs of AI goes further.

What should you count as the return?

Count changes in numbers the business already tracks: revenue or pipeline, cost per lead, time from request to finished work, error or rework rates, and hires avoided. Hours saved belong on the list only when the freed time goes into something measured. Otherwise they are capacity, not return.

Return usually falls into four groups:

1. Revenue. More pipeline, higher conversion, faster deals.
2. Cost. Lower cost per lead, less outside spend, a hire not made.
3. Speed. Shorter time from request to approved work, such as time to first draft.
4. Quality. Fewer errors, less rework, more consistent answers to customers.

Be strict about hours saved. If AI frees two hours a week for a salesperson, the return is whatever those two hours produce: more calls, more meetings, more pipeline. If the hours disappear into email, the return is zero. The article on why time saved is not ROI explains how to tell the difference.

How do you set a baseline before using AI?

Pick one workflow, choose a measure the business already records, and capture it for a few weeks before AI is involved. If the work is already underway, use historical records, such as past proposals or last quarter's campaigns. A rough baseline set honestly beats a precise one invented afterwards.

Good baselines share three traits. First, they measure the outcome rather than activity, such as meetings booked rather than emails written. Second, they come from a system, such as the CRM or a project tool, rather than from memory. Third, they cover enough time to smooth out a lucky or unlucky week.

For example, a team that drafts weekly client reports might record the time from data being available to the report being approved. Four weeks of that number is a baseline. After AI takes on the drafting, the same measure shows whether anything changed.

What does a simple AI ROI framework look like?

A simple framework has five steps: choose one workflow, set a baseline on a business number, count the full cost, run AI on it for a fixed period, then compare and decide. Repeat for the next workflow. Over time the business builds a portfolio of measured workflows rather than one big unprovable bet.

This shape has a useful side effect. Because each workflow is measured on its own, leaders can stop what does not work without abandoning AI altogether.

When should you expect a return from AI?

A single well-chosen workflow can show a measurable change within one to two months, because frequent jobs produce data quickly. Broader returns, such as revenue or headcount effects, take longer. Expect the first quarter to look flat while setup costs land and people learn the new way of working.

Be careful with headline studies here. MIT NANDA's 2025 GenAI Divide report found that 95% of organisations it studied saw no measurable return. It defined success as deployment beyond pilot with measurable results six months later. Critics, including Paul Roetzer of the Marketing AI Institute, called the method weak. The useful lesson is narrower: returns appear where AI is built into a workflow that is measured, and not where it is merely available.

To make that measurable from the start, Skein builds workflows that report on their own results, such as pipeline, cost per lead and time to first draft. The outcomes page shows what is tracked and how often. That suits a business that wants AI held to the same numbers as everything else. A team that only wants occasional help with writing may not need formal measurement at all.

Frequently asked questions

How do you calculate AI ROI?

Divide the net benefit by the total cost and show it as a percentage. Net benefit is the value gained, such as revenue, cost avoided or a hire not made, minus everything AI cost. Measure it per workflow against a baseline recorded before AI was used, or the result is a guess.

What is a good ROI for AI?

There is no standard figure, because returns depend on the workflow and the business. A useful test is whether a workflow beats its own baseline on a business measure after a fixed period, net of all costs. Compare each workflow with the alternative, such as hiring or outsourcing, rather than with an industry average.

Why is it so hard to prove AI ROI?

Most businesses adopt AI as tools spread across many small tasks, without recording a baseline first. The benefit is real but scattered, and self-reported time savings are unreliable. Measuring specific workflows against numbers the business already tracks makes the return visible.

Should time saved count as AI ROI?

Only when the freed time is spent on something measured. Hours saved are capacity, not value. If a saved hour becomes more sales calls or faster client delivery, count the result of that work. If it disappears into other tasks, the return is zero.

How long does it take to see ROI from AI?

A frequent workflow can show a measurable change within one to two months. Wider effects on revenue or headcount take longer. Expect the first quarter to look flat while setup costs land and people learn the new process, then judge each workflow on its own numbers.

Related reading