How to measure AI ROI at a manufacturer or distributor
Measure AI ROI one job at a time, in dollars, against numbers you write down before the work starts.
Where AI returns stand in 2026
Few companies report a significant financial return from AI. Six surveys published in 2026 asked different groups of people, and the table gives each result in the survey’s own terms.
| Survey | Who was asked | What they found |
|---|---|---|
| McKinsey, The state of AI in 2026 | 1,719 participants in 97 nations, May 4 to June 8, 2026 | 37% of respondents attribute at least some EBIT impact to AI use. EBIT is earnings before interest and taxes. |
| PwC, 29th Global CEO Survey | 4,454 chief executives in 95 countries and territories | 56% say their company has seen neither higher revenues nor lower costs from AI, and 12% report both. |
| Gartner, 2027 CIO and Technology Executive Agenda survey | 2,002 CIOs and technology executives, May to July 2026 | 13% of CIOs say their enterprise has realized significant or transformational value from AI. |
| Deloitte Private survey | 100 leaders of US private companies with $100 million to more than $1 billion in revenue, March 18 to 24, 2026 | 64% of companies with $500 million or more in revenue report moderate or significant ROI from AI, against 11% of companies under $500 million. |
| Grant Thornton, 2026 AI Impact Survey, manufacturing | 100 manufacturing respondents within a survey of 950 business leaders, February 23 to March 18, 2026 | None reported significant revenue uplift or significant cost savings from AI, against 12% of executives across all industries. Nearly half, 48%, are piloting AI. |
| MNP Digital Report, The business of AI 2026 | 250 business leaders across Canada | 59% characterize their returns as modest, and 16% report significant gains. |
Smaller companies trail larger ones. In the Deloitte survey, the share reporting moderate or significant ROI falls from 64% above $500 million in revenue to 11% below that line. That sample starts at $100 million in revenue. Below that size, the most useful evidence is the company’s own before-and-after numbers on one job.
Why most AI spending shows no return
The research points at how companies choose and measure the work. BCG wrote in August 2026 that companies commonly count activity, such as hours saved and tasks automated, instead of money in the income statement. The same article puts the share of companies seeing meaningful value from AI, measured in lower costs and higher revenue, at 6%.
Saved time is also lost when nobody plans where the hours go. BCG’s example: if AI makes a task 20% faster, the time saved has to go to other tasks, or the benefit disappears.
A satisfied team is not proof of a return either. In Bain’s survey of 197 executives, respondents said 80% of their generative AI use cases met or exceeded expectations. Only 23% could tie the initiatives to new revenue or lower costs.
Many companies also add AI to the old way of working. In the McKinsey survey, one-quarter of respondents outside the high performers report fundamentally redesigning workflows because of AI use.
What the companies with returns do differently
- They redesign the workflow. Nearly three-quarters of McKinsey’s AI high performers report fundamentally redesigning workflows because of AI use. McKinsey counts as high performers the respondents who attribute 5% or more of EBIT to AI and report significant value.
- They name one owner. In Bain’s survey of 1,125 sales and marketing leaders, 15% of laggards report shared or undefined ownership of AI initiatives, against 1% of winners.
- They measure the impact. McKinsey’s high performers are twice as likely as others to report defined processes to measure the impact of AI initiatives.
- They write rules for stopping weak pilots. Gartner found that organizations with guidelines for when an AI pilot should not move to production are 2.3 times more likely to be vanguards. Gartner’s vanguards are CIOs who report better business outcomes and high CEO-perceived AI value.
BCG puts 70% of the financial impact from AI on changes to people, processes and workflows. Data and the technology stack account for 20%, and algorithms for about 10%.
Measure one job before you buy more AI
Tell Derik the job you would give AI first. He will tell you what a four-week first stage would cover and how the result gets measured.
Start a conversationHow to measure AI ROI for one job
Start with a single job, such as quoting or order entry. A single job has a volume, an owner and records, so the numbers before and after come from data the company already keeps.
- Pick one job and name the owner. Choose work that happens every day and that one person answers for. AI strategy and roadmap shows how to choose the job.
- Check that the records exist. The AI readiness assessment scores whether the records behind the job can be exported, agree across systems and have an owner.
- Write down the baseline. Measure the job as it runs today, over four normal weeks, before anyone builds anything.
- Count the full cost. Add the build, hosting, model usage or seats, staff time and upkeep.
- Write the stop rule. Agree before the build which result at week four keeps the work going, changes the scope or stops the work.
- Measure again on live work. Take the same measures at week four and at month three, and count only gains that show up in records.
- Work out the payback period and the first-year ROI. Use the formulas below.
The AI ROI formula
ROI compares what a job returns with what the job costs, over the same period. Count both in dollars, by the month and for the first year.
| Measure | Formula |
|---|---|
| Monthly gain | Counted hours × loaded hourly cost + errors avoided × cost per error + added gross margin |
| Monthly running cost | Hosting + model usage or seats + upkeep hours × loaded hourly cost |
| Payback period, in months | One-time cost ÷ (monthly gain − monthly running cost) |
| First-year cost | One-time cost + 12 × monthly running cost |
| First-year ROI | (12 × monthly gain − first-year cost) ÷ first-year cost |
A loaded hourly cost is a person’s wage plus benefits, payroll taxes and overhead, divided by the hours the person works. Counted hours are freed hours that go to other work, as the section on hard and soft returns explains.
Judge the payback period the way you judge a new machine or a new hire, against the months you are willing to wait. For a first AI pricing pilot, McKinsey advises picking the use case through a clear ROI assessment. The aim is measurable impact within three to six months.
First-year ROI understates a job that keeps running, because the whole one-time cost falls in year one. Report the payback period beside the first-year ROI.
Write down the baseline first
A baseline is the job’s numbers before any change. Record the baseline over four normal weeks before the build starts, because a baseline is hard to rebuild afterwards.
- Volume: orders, quotes, invoices or reports a month.
- Time: minutes per unit of work, and the people who do the work.
- Errors: the share of work corrected after the fact, and what fixing one error costs.
- Speed: quote turnaround or order cycle time, from request to reply.
- Money: margin on the work, and the win rate on quotes.
- Extra capacity: overtime and outside help on the desk that does the job.
The baseline also helps a proposal get funded. Nigel Wallis, a research vice-president at IDC, summarized the Canadian results of the 2026 IDC manufacturing survey. He writes that a baseline, a payback window and a named owner carry more weight in AI proposals than a broad transformation argument.
Count the full cost
Count every cost of the job for the same period as the gain.
- One-time cost: the build or setup, staff hours for the test set and training, and any change to the ERP or the mailbox.
- Running cost: hosting, model usage or seats, and the hours someone spends on upkeep each month.
- Review time: the minutes a person spends approving each draft. Count review inside the minutes measured after go-live, so the same time is not counted twice.
Running costs can exceed the build. McKinsey notes that traditional IT typically costs 10% to 20% of the initial build to run each year. Generative AI at scale can incur recurring costs that exceed the initial build investment.
The cost of AI for a small business gives the formula for each part of the cost, from seats to staff time.
Set a stop rule for week four
A stop rule is a written test that decides, on a set date, whether the work continues, changes or stops. Write the rule before the build starts.
- The test: the share of real past cases, with known answers, that the system must get right.
- The live measure: a number from live work at week four, such as minutes per order.
- The decision: what happens if either number misses, such as narrowing the job or stopping the work.
Most of the organizations Gartner surveyed have no such rules. Only 31% of the CIOs have guidelines for when an AI pilot should not move to production. The share with guidelines for abandoning low-value AI already in production is 17%.
Measures by job
The baseline depends on the job. The table lists measures for four common first jobs, and what to leave out of the dollar figure.
| Job | Baseline to record | How the gain becomes dollars | What not to count |
|---|---|---|---|
| Quoting | Days from request to quote sent, estimator hours per quote and win rate | Estimator hours that go to other work, and added orders × gross margin once the win rate holds for a quarter | A win rate that rose because prices fell or the market grew |
| Order entry | Orders a month, minutes per order, the share of orders with a price or item error, and overtime | Hours that go to other work, overtime not paid, and errors avoided × cost per error | Minutes saved while the team still rekeys every line to check the draft |
| Pricing | Margin by product and customer, and the share of prices below current cost | Margin points × revenue on the repriced items, after any lost volume | Price increases that customers refused, or volume that moved to a competitor |
| Reporting | Hours to build each recurring report, and days after month end until the numbers are out | Controller or analyst hours that go to other work, and outside fees no longer paid | Faster reports that nobody uses for a decision |
Price is a strong lever on profit. McKinsey’s benchmark is that a 1% price increase gives an 8.7% increase in operating profits on average, assuming no loss of volume.
Quoting moves on speed and win rate. In a Bain client case, a building materials manufacturer redesigned quoting around a better tool. Once the tool was scaled, quote turnaround days fell by 90% and win rates rose 10 points. Bain’s page gives no date or timeline for that result.
Hard and soft returns
A hard return shows up in a record someone can check, such as payroll, overtime, invoices from outside services, credit notes or gross margin. A soft return is real but leaves those records unchanged, such as faster answers to customers or fewer interruptions at the order desk.
Count hard returns in the ROI. Track soft returns beside the ROI, and leave them out of the sum.
Time saved is the easiest return to overstate. Count an hour only when the hour goes to other work, such as overtime that stops or a hire the company no longer needs.
The Statistics Canada survey for the second quarter of 2025 asked businesses that had used AI about jobs. Of those, 89.4% reported no change to their employment levels after implementation. Plan for the gain to come from the same team doing more work, and name that work before the build.
BCG makes the same point: productivity gains stay theoretical unless they are linked to specific actions, such as spending less on outside providers.
Revenue gains need more care. A higher win rate can come from a price cut or a busy market. Compare the same months of the previous year, or keep part of the work on the old process for a few weeks.
Measure again at month three. Results drift when the volume, the staff or the model behind the system changes.
How ThriveAI helps
ThriveAI builds private AI systems on a company’s own data for manufacturers and distributors in Canada and the United States. Working sessions run by video call or on site.
Each first stage covers one job over four weeks, at a fixed price agreed before work starts. In week one, the client and ThriveAI agree a test set of real past cases with known answers. The drafts are checked against that test set, then against live work, before the team works from them.
If the client stops after the first stage, everything built in that stage is the client’s to keep. A named person at the client approves every action before anything is sent or saved.
The platform ThriveAI builds on is designed to keep each client’s data on a server dedicated to that client, in Canada or the United States. The client chooses where the model runs. One option is an open-weight model, a model whose files are published, running on that same server. The other is a frontier model under a written zero-data-retention control, with the model tier named. A hosted model may process requests in another country.
ThriveAI’s offers describe the four weeks for each job, from quoting to the order desk.