09Numbers, and whose they are
How much AI saves: the published figures and where each comes from
Search for what AI saves a business and you get percentages. Twenty to thirty percent off the cost of the work. Half the support queue gone. A back office that costs a third of what you pay for it. The figures usually arrive with a large consulting firm's name attached, which makes them feel settled.
01
The numbers everyone quotes, checked at the source
We sat down to source the nine figures this site was planned around, because we were not willing to print a number we had not read in its own document. Seven of the nine did not survive that reading: five are not in the document they are credited to, and two could not be read at the publisher, so we cannot show them to be true. Among them: one is not in the report it is credited to. Two trace to a single forecast about the year 2029, written in the present tense. One traces to a marketing blog citing two studies we could not show exist.
So this page does the thing that nobody selling AI automation seems to do. It lists the numbers we use, who published them, and when. Then it lists the ones we threw out and why. You can check every line of it in a browser, which is the point. The entry-by-entry record behind it, with every URL, verdict and the sources that refused to be read, is kept at the verification ledger, which is revised in place and dated.
02
What is actually measured, and by whom
Four sources carry every savings or adoption figure we use anywhere on this site. That is the honest size of the evidence base. Not one of them is a savings percentage you could apply to your own business, and we would rather show you that than fill the gap.
15%
more issues resolved per hour by 5,172 customer support agents, at one company, given access to a generative AI assistant
The Quarterly Journal of Economics, Generative AI at Work (Brynjolfsson, Li and Raymond), vol. 140, issue 2, p. 889, published . Read back .
The strongest number in the set, and it is not close: peer reviewed in a top economics journal rather than published by a company selling the thing being measured, a realised measurement across thousands of workers rather than an opinion survey, and about the exact kind of work we get asked to automate. The gains landed overwhelmingly on the least experienced workers; the most experienced ones got a little faster and a little worse. A version trap worth knowing about, because it is a small lesson in how numbers drift: an earlier working paper version of this study reports an average one point lower, across 5,179 agents. We cite the journal, because that is the version of record. If you see this study quoted anywhere, check which version the quoter read. People mix figures across drafts constantly and it is almost never flagged.
11%
lower costs, which the executives surveyed expect within 18 months of deploying and scaling generative AI
Accenture, The front-runners' guide to scaling AI: Lessons from industry leaders (Accenture Research), published . Read back .
The load-bearing word is expect. The same organisations expect a productivity gain of a similar size over the same 18 months, alongside a rise in revenue and in customer experience. All of it is what senior people think will happen, not what anyone measured. We searched the document itself, the full PDF, and that is the figure it carries.
37%
of firms with 250 or more employees report using AI
US Census Bureau, Large Firms With at Least 20 Employees Biggest AI Users (Business Trends and Outlook Survey), by Adam Grundy, Cory Breaux and Dhanapati Khatiwoda, published . Read back .
About a third of firms with 100 to 249 employees do, and under a fifth of firms with four or fewer. Use rose among firms with at least 20 employees and did not move significantly among firms below that. The survey covers December 2025 to May 2026. This measures who is using AI, not what it did for them. It is the right number for larger firms have already started. It is the wrong number for anything about savings.
18%
of US firms had adopted AI by the end of 2025
Board of Governors of the Federal Reserve System, FEDS Notes: Monitoring AI Adoption in the US Economy, by Jeffrey S. Allen, published . Read back .
More than three quarters of the labour force works at a firm that has, even though most firms have not, and about two in five workers used generative AI at work as of November 2025. Professional services and financial services each run at around a third. Same caveat, stated again because it is the one people skip: this is adoption. It contains no cost effect and no productivity effect at all.
03
Two conversions to refuse
Both of these are one sentence away from the figures above. Both are wrong, and both are made constantly.
The Accenture figure is what executives expect, not what anyone measured. It comes from asking senior people what they think will happen within 18 months. An expectation is a plan, and plans slip. If we wrote "AI cuts costs by that much" we would have silently turned a survey of opinions into a measured result, and you would have no way to tell from the sentence which one you were reading. So it stays written as an expectation, every time it appears.
The QJE figure is a productivity result, not a cost saving. It counts issues resolved per hour. It does not count money. Whether extra throughput becomes a saving depends entirely on what happens next. If the queue was your constraint and you now serve more customers, it shows up as revenue. If you hold headcount flat and volume grows into the slack, it shows up as cost per contact. If nothing downstream changes, it shows up nowhere at all. The study measured throughput. It did not measure your payroll, and converting one into the other would be our arithmetic dressed up as their finding.
Saying this out loud costs us the two most useful numbers we have. We would rather lose them than have you find out later that we bent them.
04
Six figures this site will not publish
Named plainly, with the real number where there is one. Each is printed here as a claim that failed checking, under the ledger's verdict, so that nobody reads it as a supported one.
Claimed, ledger entry A1, verdict not found
20% to 30% cost savings in automated functionscredited to Accenture, The front-runners' guide to scaling AI
Not found
The report does not say it. We pulled the full text of the 39 page PDF and searched it. The claimed range appears zero times in the document, and the phrase "automated functions" does not appear at all. The search itself works, it finds 75 other percentage figures in the same file, which is how we know the absence is real and not a broken probe. The genuine figure is the expectation of lower costs in the table above, and it is an expectation.ledger A1
Claimed, ledger entry A3, verdict not found
Top quartile AI support deployments cut costs 53%credited to McKinsey, as the figure circulates
Not found
We cannot find it. Not on mckinsey.com: the only figure of that size we can find there is about executives and the financial crisis, a different subject entirely. We fetched the blog post the search engine credited for the figure and the number is not in that post either, which means part of the attribution was manufactured somewhere in the retrieval layer rather than merely repeated carelessly.ledger A3
Claimed, ledger entry A4, verdict unverifiable
AI customer service cuts cost by about 30%credited to IBM, as the figure circulates
Not found
It is not IBM's, and it is not a measurement. It is a Gartner forecast: that by 2029 agentic AI will autonomously resolve four in five common customer service issues, leading to lower operational costs. A prediction about 2029 is not a result you can put in a 2026 business case. Both the Gartner release and the IBM pages that cite it refused our requests, so we stopped there rather than working around them, and the figure stays unverified as well as misattributed.ledger A4
Claimed, ledger entry A6, verdict unverifiable
AI answers 50% to 80% of routine questionscredited to no named source; it circulates unattributed
Not found
The same forecast, wearing different clothes. The upper bound is that identical Gartner prediction about 2029. The lower bound has no traceable source at all. The claim also quietly converts a forward-looking prediction into a present-tense capability, which is the exact move this page exists to catch.ledger A6
Claimed, ledger entry A2, verdict not found
The back office is 30% to 40% of operating costscredited to McKinsey estimates, per one outsourcing vendor's page
Not found
The real numbers are far smaller. This traces to a single outsourcing vendor's page asserting it "per McKinsey estimates", with no document, title or date behind the attribution. McKinsey's own reachable material puts support functions at a fifth of personnel expenses at most, and fixed and support costs at a much smaller share of operating costs. Those are different denominators and much smaller quantities. The quoted figure is inflated by roughly two to three times.ledger A2
Claimed, ledger entry A7, verdict not found
68% of mid-market AI projects reach production against 31% of enterprise, in 4.2 months against 13.6credited to a 2026 BCG analysis and a 2026 McKinsey survey, per one marketing blog
Not found
One blog, no studies. The source is a single marketing post. It credits a BCG analysis of 1,200 North American companies and a parallel McKinsey survey of 580 operators, and gives no title, no link and no method for either. BCG's own 2026 research is easy to reach, so the search worked; the named studies still could not be located. These were the four most quotable numbers in the plan and the four least supported, which is usually how that goes.ledger A7
05
A seventh figure failed differently and is worth naming here. A compliance date we had drafted onto another page had become false between the plan being written and the page being drafted, because the law changed underneath it. The date that replaces it, 2 December 2027, is attributed to the European Commission, which is where we read it. That correction is on the hiring page, where it belongs.
06
Why a range, and not a single number
"What does AI save?" is not a question with a number for an answer, in the same way that "what does a building cost?" is not. It is a family of very different questions wearing one label. Four things move the answer more than the published figures differ from each other:
- 01
Which task. Answering a routine password reset and closing a quarter are both "automation". They behave nothing alike.
- 02
What you are counting against. Savings against a manual process, against the vendor you used last year, against the budget you set in January, and against what a competitor spends are four different baselines and four different percentages for the identical piece of work.
- 03
Who did the counting. A software company measuring its own product is evidence, but it is the weakest kind. It is also the kind that dominates search results for this topic.
- 04
How long after go live. Month one includes the novelty and excludes the maintenance. Month twelve includes the maintenance and the model updates and the person who now owns the thing.
Change any one of those and the number moves further than the gap between most of the figures people quote at each other. That is why a point estimate is a false object here. There is a second reason a range is better, and it is the one that matters to a buyer. A range tells you the spread. If reported outcomes for a piece of work run from roughly nothing to roughly a third, then a third is possible and nothing is also possible, and you should plan for both. A single number hides the bottom of the range, which is precisely the end you need to survive.
07
What to measure in your own business instead
A benchmark is a number produced by a company that is not yours, doing work that is not yours, counted by someone who was not you, under conditions nobody wrote down. Even an excellent one is a weak guess about your building. You can do much better than a guess, cheaply, before anything is built, and you can estimate it on your own numbers first. Four numbers, collected on the task itself:
- 01
Volume. How many times a month does this actually happen? Take it from the system that already records it: the ticket queue, the ledger, the applicant tracker, the phone log. Not from memory, and not from the person who owns the process.
- 02
Time per unit. How long does one take from start to finish, including the waiting and the chasing, not just the typing. Time twenty of them rather than estimating one.
- 03
Fully loaded hourly cost of whoever does it. Salary is not the number.
- 04
What breaks. Error rate, rework rate, how often it is late. Automation that is faster and wrong is not a saving, it is a faster problem.
Multiply the first three and you have the current cost of that one task, in your business, in your currency, this year. That is a real quantity, and no published percentage can compete with it, because it is measured on the actual thing.
08
Then measure the same four numbers the same way, out of the same system, one month and three months after the automation goes live. The difference is the result. It is yours, it is checkable, and if it is disappointing you learn that from your own data instead of from us.
Two rules we hold to on this. The baseline is agreed in writing before the build starts, because a baseline negotiated afterwards is not a baseline. And it is read out of a system that was already running before we arrived, not out of a spreadsheet built to make the case. If a number only exists because someone wanted to prove something, it is not evidence.
That measurement is also what the commercial side of this rests on: the fee is a share of the measured gain, and if there is no measured gain, there is no share. Which is why we would rather the baseline were slightly unkind to Praxis than slightly generous.
09
How to check a savings statistic in ten minutes
This is the routine we used on the nine figures. Seven failed it.
- 01
Find the actual document. Not a blog quoting it, not a listicle. If nobody can produce a title, a publisher and a date, the number does not exist. This alone kills most of them.
- 02
Check the verb. "Expect", "forecast", "predicts", "by 2029" and "could" are not "measured".
- 03
Check what was measured. Productivity, revenue, cost, headcount and elapsed time are five different things and they get swapped for each other freely.
- 04
Check who paid for it. A vendor measuring its own product is not a neutral referee.
- 05
Check the denominator. A percentage of personnel expense, of total operating cost, and of one process's cost are three different numbers that look identical in a sentence.
- 06
Search the document for the figure. Not the summary of the document. If the search engine works on the file and the number is not in it, the number is not in it.
You can run that on any statistic anyone quotes at you, including ours.
Questions
What should you know before the first call?
- How much money can AI save a business?
- There is no honest single percentage, and anyone offering one is quoting something they have not read. What AI saves depends on which task, what you measure against, who counted it and how long after launch, and each of those moves the answer more than the published figures differ from each other. The best evidenced result available is a measured rise in issues resolved per hour among 5,172 customer support agents, published in The Quarterly Journal of Economics in February 2025, and that is a productivity finding rather than a cost saving. The only reliable number for your business is the one measured on your own task before and after.
- Is the "AI cuts costs twenty to thirty percent" figure real?
- No. That figure is usually attributed to Accenture's report on scaling AI, and the report does not contain it. Searching the full text of the 39 page document returns the claimed range zero times, against a search that finds 75 other percentage figures in the same file. What the report actually says is that the organisations surveyed expect a cost decrease of about a tenth, and a slightly larger productivity gain, within 18 months. That is an expectation from a survey, not a measured outcome.
- What is the best evidenced result for AI in customer support?
- Brynjolfsson, Li and Raymond, "Generative AI at Work", The Quarterly Journal of Economics volume 140 issue 2, published 4 February 2025: access to a generative AI assistant raised productivity, measured as issues resolved per hour, by a measured margin on average across 5,172 support agents, with the gains concentrated among the least experienced workers. It is peer reviewed, it is a realised measurement rather than a forecast, and the full text is free to read. Note that an earlier working paper version of the same study reports a slightly lower figure across 5,179 agents; the journal version is the one to cite.
- Why do so many AI savings statistics turn out to be wrong?
- Because most of them are never checked against their own source. A figure appears in a vendor blog, gets a consultancy's name attached to it for credibility, is repeated by aggregator articles, and by the fourth repetition the attribution looks solid while no link to a document has ever existed. Of nine widely repeated figures checked one by one against their primary sources, six could not be supported: one was absent from the report it was credited to, two were the same forecast about 2029 presented as current results, one was inflated two to three times over its real source, and two traced to blogs citing studies that could not be located.
- How should a business measure what AI actually saved it?
- Measure four things on the specific task before anything is built: how often it happens per month, taken from the system that already records it; how long one takes end to end including waiting; the fully loaded hourly cost of the person doing it; and how often it goes wrong or has to be redone. Multiplying the first three gives the current cost of that task in your own numbers. Measure the same four the same way one month and three months after launch, and the difference is the result. Agree the baseline in writing before the build starts, and read it from a system that was already running.
- Are any of these figures a claim about what Praxis will do for you?
- No. Every figure on this page is another organisation's published result, reported in that organisation's own terms with its publisher and date attached, and none of them is a forecast of what would happen in your business. They exist so you can see the state of the evidence, including how thin it is. What we would put in front of you instead is a measurement of your own work, taken before we touch it.
Where this leads
This page is one part of the whole offer: have it built, paid from profit.
Start with a free consultation
If you want a number for your own business rather than somebody else's, the honest route is a call and 2 weeks inside the work, not a percentage from a web page. The call costs nothing and ends with a plain list either way.
No obligation · a scoping conversation first