How to Measure ROI on an AI or Automation Investment

Measuring the return on an AI or automation investment comes down to four steps:

count the full cost honestly, identify specific sources of value, measure performance before and after against a real baseline, and convert the difference into money over a sensible time period. The arithmetic is simple. What makes it difficult is that most of the costs are hidden, many of the benefits are indirect, and the most common way of measuring them, asking people whether the tool helps, turns out to be unreliable.

That difficulty matters. Businesses are spending heavily on AI tools, automation platforms and custom models, yet many struggle to show what they have gained. Some projects are cancelled because no one can demonstrate value, even when they are delivering it. Others carry on for years because no one checks whether they are delivering anything at all.

This guide explains how to build a credible ROI case for AI and automation, whether you are evaluating a proposal or reviewing something already in place. It covers the basic calculations, the costs most often missed, the main categories of value, how to measure properly, a worked example, the indicators to watch early on and the mistakes that most often distort the numbers.

Why AI and automation ROI is hard to measure

Traditional automation, such as a workflow that routes invoices or a script that reconciles two systems, is relatively easy to evaluate. It does the same thing every time, and you can count how many hours it replaces. AI is harder, for several reasons.

First, the effect varies widely between tasks and people. A well-known study by Erik Brynjolfsson, Danielle Li and Lindsey Raymond looked at customer support agents given an AI assistant. On average, they resolved around 14 percent more issues per hour, but the gains were concentrated among newer and less experienced agents, while the most experienced saw little improvement. A study of consultants at Boston Consulting Group, led by researchers including Fabrizio Dell'Acqua at Harvard Business School, found that AI substantially improved speed and quality on tasks within its capabilities, but made consultants less likely to reach the correct answer on a task that fell outside them. The researchers described this uneven pattern as a "jagged technological frontier". An average figure can hide large gains in one area and losses in another.

Second, people are not good at judging the effect themselves. In a 2025 randomised study by the research organisation METR, experienced open-source software developers took about 19 percent longer to complete tasks when using AI tools, even though they believed the tools had made them roughly 20 percent faster. The result applied to a specific group and setting, and other studies have found gains in different contexts, but it is a powerful reminder that perceived productivity and measured productivity can diverge sharply.

Third, many benefits are indirect. Faster responses may improve customer retention months later. Fewer errors may reduce compliance risk that never shows up as a line in the accounts. Time saved by staff only becomes financial value if it is used for something worthwhile.

Finally, costs are easy to underestimate. Licence fees are visible. Data preparation, integration, prompt and workflow design, monitoring, human review and change management often are not.

Industry surveys suggest the consequences are widespread. Gartner predicted in 2024 that at least 30 percent of generative AI projects would be abandoned after proof of concept by the end of 2025, citing unclear business value among the main reasons. Measuring ROI properly is how you avoid both abandoning valuable projects and continuing worthless ones.

The ROI basics: formula, payback and time horizon

Before getting into the detail, it helps to agree on the measures you will use. Most businesses need only three.

Return on investment (ROI) compares the net benefit with the cost. The standard formula is the total benefit minus the total cost, divided by the total cost, expressed as a percentage. If an automation project costs $50,000 over three years and produces $120,000 of benefit over the same period, the ROI is ($120,000 minus $50,000) divided by $50,000, or 140 percent.

Payback period is how long it takes for cumulative benefits to equal cumulative costs. It is intuitive and especially useful for smaller businesses where cash flow matters. A project with a high ROI but a three-year payback may be less attractive than one with a lower ROI that pays back in six months.

Net present value (NPV) adjusts future benefits and costs for the time value of money, using a discount rate that reflects your cost of capital or required return. It matters most for larger investments with benefits spread over several years. Your finance team or accountant can advise on an appropriate rate.

Choose the right time horizon

The period you measure over has a large effect on the result. AI and automation projects usually involve upfront costs, a period of setup and learning when benefits are low, and then ongoing benefits alongside ongoing running costs. Measure over too short a period and almost any project looks poor. Measure over too long a period and you risk counting benefits that may never arrive, especially given how quickly AI tools and pricing change.

For most business AI and automation projects, a horizon of two to three years is a sensible default, with payback tracked month by month. Be cautious about assuming that any particular tool, model or price will remain unchanged for longer than that.

Measure at three points

It is useful to estimate ROI three times: before the investment, as a business case based on assumptions; after a pilot, using real data from a limited rollout; and after full deployment, to confirm whether the benefits were realised at scale. Each estimate should replace assumptions with measurements wherever possible.

Counting the full cost

Understated costs are the most common reason AI and automation business cases look better on paper than in reality. A credible ROI calculation includes all of the following, even where some have to be estimated.

Upfront costs

These include software licences or platform subscriptions for the initial period, development or configuration work by internal staff or an external partner, integration with existing systems such as CRM, ERP, email or document storage, data preparation, which is often the largest hidden cost and covers cleaning, structuring and connecting the data the system relies on, security, privacy and compliance work, such as data protection impact assessments and vendor due diligence, and initial training for staff.

Ongoing costs

Many AI services are priced by usage, for example per user, per document processed or per amount of text handled by a language model. Usage-based costs can grow quickly as adoption increases, so model them at expected volumes, not pilot volumes. Ongoing costs also include hosting and infrastructure, maintenance and updates when connected systems or the underlying AI models change, monitoring of accuracy and performance, and support.

The cost of human oversight

Most business uses of AI keep people in the loop. Someone reviews AI-drafted emails, checks extracted invoice data, approves recommendations or handles cases the system cannot. This review time is a real cost and is easy to overlook. If staff spend half the time they save checking the AI's output, the net saving is half what the headline suggests.

Change and transition costs

New tools require process changes, updated documentation, management time and a period where productivity may dip while people adjust. Running old and new processes in parallel during a pilot also costs time.

Risk costs

Finally, consider the cost of things going wrong: errors that reach customers, decisions that need reversing, or a data protection incident. These are hard to quantify precisely, but a simple estimate of likelihood and impact is better than ignoring them.

Identifying where the value comes from

AI and automation create value in four main ways. Most projects deliver a mix, and it helps to identify each source separately, because they are measured differently and carry different levels of certainty.

1.Cost reduction

This is the most direct form of value: less staff time spent on a task, lower spending on outsourced services, fewer errors that need correcting, or reduced software and infrastructure costs. It is also the easiest to measure, but it comes with an important caveat. Time saved is only a cost saving if it reduces spending, for example through lower overtime, fewer temporary staff, avoided hiring or reduced outsourcing. If people simply have more free time, the value depends on what they do with it.

2.Capacity and speed

Often the more realistic benefit is that the same team can handle more work, or handle it faster. A finance team that processes invoices in half the time can absorb growth without hiring. A sales team that spends less time on admin can make more calls. A support team that responds faster may handle peaks without backlogs. To turn this into financial value, you need to show what the extra capacity is used for, such as the cost of hires avoided as the business grows, or additional revenue from the time released.

3.Revenue growth

Some AI applications aim directly at revenue: better lead scoring, personalised recommendations, faster quoting, improved conversion on a website or better customer retention. These benefits can be significant but are harder to attribute, because many other factors affect revenue at the same time. Controlled tests, described in the next section, are the most reliable way to measure them.

4.Quality and risk reduction

Fewer errors, more consistent decisions, better compliance and earlier detection of problems all have value, even if they rarely appear as a single line in the accounts. Estimate them by putting a cost on the problems avoided: the cost of correcting an error, the cost of a typical customer complaint, or the expected cost of a compliance breach multiplied by a realistic reduction in its likelihood.

Be honest about certainty

It is good practice to separate "hard" benefits, which can be measured directly and translate into cash, from "soft" benefits, which are real but estimated. A business case that relies mainly on soft benefits is not necessarily wrong, but decision-makers should know that it is less certain. Some organisations calculate ROI using hard benefits only, then show soft benefits as additional upside.

Setting a baseline and measuring properly

You cannot measure improvement without knowing where you started. Yet many AI projects begin without any baseline at all, which leaves the business relying on impressions, and as the METR study showed, impressions can be badly wrong.

Measure the current process first

Before introducing a new tool, record how the process performs today. Useful measures include volume, such as the number of invoices, tickets or enquiries handled per week; time per task or per item, ideally from system timestamps rather than estimates; error or rework rates; turnaround time from start to finish; cost per transaction; and outcome measures such as conversion rate, customer satisfaction or first-contact resolution. Collect at least several weeks of data, and longer if the work is seasonal, so that you have a realistic picture rather than a single good or bad week.

Compare like with like

The simplest approach is a before-and-after comparison, but it has a weakness: other things change at the same time. Sales rise in a busy season, staff become more experienced, a new product launches. To isolate the effect of the AI or automation, compare against a group that does not use it where you can.

The strongest method is a controlled comparison, such as an A/B test, in which similar groups of users, customers or transactions are randomly assigned to use the new tool or continue as before. Where full randomisation is not practical, a phased rollout can work almost as well: introduce the tool to one team or region first and compare its results with those of similar teams over the same period.

Measure what matters, not just what is easy

Usage statistics, such as the number of prompts sent or documents processed, show adoption, not value. Time saved per task is useful, but only if you also check quality. If an AI tool helps staff draft replies twice as fast, but customer satisfaction falls or more replies need correction, the net value may be lower than it appears. Track outcome measures alongside efficiency measures.

Combine data with feedback

User feedback is still valuable, especially for understanding why something is or is not working. Use it to explain the numbers rather than replace them. Short surveys and interviews can reveal workarounds, frustrations and benefits that the data does not capture.

A worked example

To show how these pieces fit together, here is a simplified, hypothetical example. The figures are illustrative and chosen to make the calculation clear, not to represent typical results.

A growing business processes around 2,000 supplier invoices a month. It is considering an AI-powered tool that reads invoices, extracts the key details and enters them into its accounting system, with staff reviewing the results.

The baseline

At present, each invoice takes about six minutes of manual work, or roughly 200 staff hours a month. With a fully loaded staff cost of $35 an hour, that is $7,000 a month. Around 3 percent of invoices contain entry errors, and each takes about 30 minutes to find and correct, adding roughly 30 hours, or $1,050, a month. The baseline cost of the process is therefore about $8,050 a month.

After automation

In a pilot, staff take about 1.5 minutes to review each automatically processed invoice, around 50 hours a month. About 10 percent of invoices are exceptions that still need full manual handling, adding another 20 hours. The error rate falls to 1 percent, needing about 10 hours of correction. Total effort falls to around 80 hours, or $2,800 a month, a gross saving of about $5,250 a month.

The costs

Setup, including configuration, integration with the accounting system, data preparation, testing and training, costs $30,000. Ongoing software and usage fees, plus monitoring and maintenance, cost $1,500 a month.

The result

Assume benefits reach only half their full level in the first three months while the team adjusts. Over three years, total benefits are about $181,000, and total costs about $84,000. Net benefit is roughly $97,000, an ROI of about 116 percent, with payback in around the tenth or eleventh month.

The important caveat

This calculation assumes the 150 or so staff hours saved each month are genuinely valuable, for example because the business is growing and would otherwise need to hire another person, or because staff move to higher-value work such as supplier negotiation or credit control. If only half the time saved is put to valuable use, the three-year benefit falls to around $91,000 and the ROI drops to below 10 percent. The project still pays back, but only just.

This is why a sensitivity check is so useful. Showing how the ROI changes under more cautious assumptions, such as lower time savings, higher usage costs or slower adoption, gives decision-makers a realistic range rather than a single optimistic number.

Leading indicators and when to expect returns

Financial returns often take months to appear, and waiting until the end of a project to find out whether it worked is risky. Leading indicators help you see early whether an investment is on track.

Why returns are often delayed

Economists Erik Brynjolfsson, Daniel Rock and Chad Syverson have described what they call the productivity J-curve. When organisations adopt a general-purpose technology such as AI, they first have to invest in complementary changes, such as new processes, skills, data and ways of working, that do not show up immediately as output. Measured productivity can therefore dip or stay flat at first, before rising once those changes take hold. At the level of an individual business, this means early disappointment is not necessarily a sign of failure, but it does need to be distinguished from a project that will never pay off.

Leading indicators to watch

In the first weeks and months, focus on measures that predict future value. Adoption shows whether the intended users are actually using the tool for the intended tasks, rather than reverting to old methods. Accuracy and quality track how often the output is correct, how much needs editing and how many exceptions require manual handling. Time per task shows whether the measured time for the process is falling, not just whether people feel faster. Throughput shows whether the same team is handling more volume. And user feedback reveals whether people find the tool helpful and where they are struggling.

If adoption is low, accuracy is poor or time savings are not appearing after a reasonable learning period, it is better to investigate and adjust early, or stop, than to wait for disappointing annual figures.

Lagging indicators

Over the longer term, confirm the value with financial and outcome measures: cost per transaction, overtime and temporary staff spending, hires avoided, revenue or conversion changes, customer satisfaction and retention, and error-related costs. Compare these with the original business case and update your ROI estimate with real figures.

Set review points in advance

Agree at the outset when you will review results, typically after the pilot, at three to six months and at one year, and what level of performance would justify expanding, adjusting or stopping. Deciding these thresholds before seeing the data reduces the temptation to rationalise disappointing results.

Common mistakes that distort AI ROI

Counting time saved as money saved. This is the most common error. Hours freed up only become financial value if they reduce costs or are used for work that generates value. Be explicit about what will happen to the time.

Using vendor benchmarks instead of your own data. Vendors' case studies and headline percentages come from other businesses, processes and data. They can be a useful starting point for a business case, but they should be replaced with your own measurements as soon as possible.

Relying on self-reported productivity. As the METR study illustrated, people can feel faster while actually being slower. Use timestamps, system data and controlled comparisons wherever possible.

Ignoring the cost of checking AI output. Human review is often essential, especially for anything customer-facing or high-stakes. Include its cost, and track whether review time falls as the system and the team improve.

Measuring at pilot scale only. Pilots often run with enthusiastic users, clean data and close support. Results can change at scale, both positively, as fixed costs are spread, and negatively, as usage costs rise and edge cases multiply. Recalculate after full rollout.

Forgetting that AI products change. Pricing, features and underlying models change frequently. A model update can improve or degrade performance on your specific tasks. Monitor quality continuously and revisit the business case when significant changes occur.

Ignoring quality and risk effects. A tool that speeds up work while quietly increasing errors or compliance risks can have negative ROI even when the time savings look impressive. Always measure quality alongside speed.

Automating a poor process. Automating an inefficient process makes it faster, not better. Often the largest gains come from simplifying the process first, then automating what remains.

If you are evaluating an AI or automation investment, it can help to work with a partner who is willing to build the business case with you, define the baseline and measure results honestly, including when they fall short. The aim should be clear evidence of value, not just a working tool.