- Published
- Updated
- 8 min read

TL;DR: AI savings came in below plan for many large companies. Bain's 2026 survey of 951 firms found the largest group of those measuring AI savings got 10% or less, after targeting 11% to 20%, and 44% are funding the next wave from prior savings. A small business should bank real AI savings from one measured workflow before spending them.
Updated October 2026: an earlier version quoted a 4% figure and other numbers we could not find in Bain's published work; they were removed.
A survey went around in June, and the number that stayed with me was not the disappointing one about AI savings. It was the one about what companies are doing next with AI savings they may not have.
Bain & Company asked large companies how their automation and AI spending had paid off. The answers say a lot about how budgets get built, and the lesson travels well from a billion-dollar company down to a ten-person office.
What did Bain's survey find about AI savings?
The study is the Bain Automation and AI Pathfinder Survey 2026, with 951 companies in the sample. Bain's summary: "Nearly 40% of companies that measured AI cost savings landed below 10%, despite targeting 11% to 20%." (Bain & Company (opens in a new tab))
Bloomberg's coverage adds who was asked and when. "The survey, completed in April, was based on responses from executives at 951 companies" with "more than $100 million in revenue," and "Among companies measuring their AI cost savings, the largest share (40%) realized reductions of 10% or less." (Bloomberg via Insurance Journal (opens in a new tab))
The technology worked, the value did not
Bain's own line is short: "The value didn't arrive." (Bain & Company (opens in a new tab)) These are companies with data teams and finance departments, and many still landed in single digits on AI savings.
The original version of this note said only 4% of companies passed 30% savings. That figure is not on Bain's published page or in Bloomberg's report, so it is gone. The 40% number is enough to make the point.
Budgets are rising anyway
Here is the part that should make any owner pause. Bain writes that "rather than pausing to understand why, 90% of those same companies are now increasing their budgets again." (Bain & Company (opens in a new tab)) Spending more is not the mistake. Spending more before you know why the first round fell short is.
Why is funding AI from AI savings a risky bet?
Bain asked how companies plan to pay for the next wave of generative and agentic AI. The largest group, 44% of companies, "cited savings from prior automation programs." Bain's verdict: "Self-funding the next wave from past returns sounds like discipline. In reality, it is a circular bet with a structural leak." (Bain & Company (opens in a new tab))
Reinvesting a forecast
Bain's reasoning is worth reading slowly. "The self-funding model that 44% of companies are relying on only works if the prior returns are real." (Bain & Company (opens in a new tab)) If the first wave underdelivered, the pool of AI savings is smaller than the plan assumed. If the business case for wave two was built on projected AI savings, not banked ones, the company is not reinvesting profit. It is reinvesting a forecast.
The small business version
A large company can absorb a missed forecast. It becomes an awkward slide and a revised plan. In a small business, a forecast you already spent is not a rounding error. It is payroll, or the quarter.
I see the same pattern at small scale. A subscription gets added because it will "save ten hours a week." Nobody counts the hours. Six months later a second tool gets added on the strength of the first one's AI savings, which were never measured. The habit is the same as the large company's, just with smaller numbers and less room to absorb it.
Why did the AI savings fall short?
It is tempting to blame the models. Bain's data points somewhere less exciting and more useful.
Data access came first
"Data access and integration is the single biggest barrier to AI progress, cited by 41% of respondents." Bain goes further: "the No. 1 reason AI programs underperform is that companies cannot reliably get access to their own data." (Bain & Company (opens in a new tab))
In a small business that sounds like this: customer history in the email inbox, prices in a spreadsheet, job notes in a phone, invoices in the accounting app, and none of them connected. An AI tool cannot save time on work whose inputs it cannot reach. I wrote about this on my own site in AI without the data (opens in a new tab), and it is the quiet reason many AI savings never show up. Good data engineering is unglamorous, and it is usually the first real step.
Budgets assumed autonomy that mostly isn't there
The second gap is about what the AI actually does. Bain found "only 7% of companies running fully autonomous agents in production today." (Bain & Company (opens in a new tab)) If a business case assumed the AI would run a whole job start to finish, and in practice a person still reviews, corrects and approves, the AI savings shrink to whatever that review time leaves over.
That is not a reason to give up on automation. A person in the loop is often the right design. It is a reason to budget for the loop. Our note on "botsitting" looks at where that review time goes.
Measurement came after the money
The third gap is timing. Many companies sized their AI savings before launch and checked them after the next budget was already approved. This is a familiar move in how people reason about money: a number written in a plan starts to feel like a number in the bank. Nobody decided to treat the forecast as income. It just happened when nobody went back to check.
How common is this at small businesses?
AI use is spreading quickly, so the question of AI savings reaches more owners every quarter. The Census Bureau's 2026 AI supplement found that "18% of firms used AI in a business function, rising to 32% on an employment-weighted basis," and expected adoption "to reach 22% within six months." It also found the scope is narrow: "57% of users integrate AI in three or fewer business functions." (U.S. Census Bureau (opens in a new tab))
Narrow use is good news. It means most businesses can still measure AI savings one function at a time, before they stack tools on top of each other. We looked at what Massachusetts owners report in 51% of Massachusetts small businesses use an AI platform.
Cuts that come back
There is one more way AI savings go negative: cutting people first and finding out later the work still needs doing. We covered that in 55% of companies regret replacing workers with AI. The fix is the same as everything above. Measure what the tool actually does before you change the team around it.
How do you measure AI savings before you spend them?
The companies in Bain's survey that came out ahead did something plain: they funded the next wave from returns that had actually arrived. A small business can do a simpler version of the same thing.
Pick one workflow
Not a department and not a program. One workflow, close to the money: answering common customer questions, turning a phone call into a quote, chasing unpaid invoices, or drafting the weekly update. If you are unsure where to start, my guide on how to figure out where AI fits in your business (opens in a new tab) walks through the choice.
Write down the baseline first
Before you switch anything on, time the workflow for two weeks the way you do it now. How many times a week, how long each time, who does it, and how often it goes wrong. The NIST AI Risk Management Framework makes the same point for larger teams: monitor and document how production metrics "differ from the same metrics collected during pre-deployment testing." (NIST AI RMF Playbook (opens in a new tab)) Without a baseline there are no AI savings to measure, only impressions.
Run it for 90 days and count three things
- Time saved per week, counted from the baseline, not guessed. Include the time someone spends checking the AI's work.
- Quality compared with how you did it before. Same or better, or the AI savings do not count.
- Total cost: the subscription, usage charges, setup, training, and the slower weeks while everyone adjusted.
Brian has a 30-day version of the time test in does AI save time at work (opens in a new tab) if you want a shorter first pass.
Then decide what the next dollar does
If after 90 days you have banked real AI savings, you have real money to decide what to automate next. If you have not, you saved yourself from funding round two on a number that did not happen. Either result is useful, because both rest on something you measured, not something you hoped.
The trap is not AI. The trap is letting a forecast act like income.
Frequently asked questions
How much did AI save large companies in 2026?
Less than planned for many. In Bain's survey of 951 companies with over $100 million in revenue, the largest share of those measuring AI cost savings, about 40%, got 10% or less after targeting 11% to 20%.
What does "a circular bet with a structural leak" mean?
Bain's phrase for funding the next AI wave from savings the last wave was supposed to produce. If those savings were smaller than planned, the new budget rests on money that never arrived.
Why do AI savings fall short?
Bain's top reason is data: 41% cited data access and integration as the biggest barrier. Business cases also assumed full automation, while only 7% of companies run fully autonomous agents in production.
How should a small business measure AI savings?
Pick one workflow, time it for two weeks before you start, then run the AI version for 90 days and count time saved, quality and total cost, including the time spent checking the AI's work.
Should I stop spending on AI?
No. Fund the next step from AI savings you have actually banked, not from projections. A measured small win is a better base than a large plan nobody checked.
Where we fit
Kief Studio is in Shrewsbury, and our AI automation work follows the order in this note. We pick the workflow, define what good output looks like, and measure how it runs today. Then we build the automation with retries, error handling and quality gates on every output, keep a human path for anything the model is unsure about, and run it for you on LTFI, reporting what it handled and what it sent to a person. When the real blocker is data spread across tools, our data engineering work comes first.
We do not take on projects meant to replace staff with AI. We build systems that remove bottlenecks, and your people drive the work.
First conversation is free. No commitment.




