Skip to content
PRAXIS

R.01Research · Verification ledger

AI savings statistics: the verification ledger

Nine figures that circulate in decks and articles about what AI saves a business, each traced to the document it is credited to and searched for the number. Seven failed. This page records every entry, the verdict, the nearest true figure where one exists, the sources that could not be read, and the correction made to one of its own verdicts on 2026-10-06. It is maintained in place, so the URL is stable and the date of the last revision is stated.

Retrieval date
2026-09-01
First published
2026-09-09
Last revised
2026-10-06
Entries
9 checked, 4 found while checking

Correction

What this ledger got wrong, and when it was fixed

A ledger that checks other publishers’ figures is held to its own rule. Where a verdict on this page was wrong, the correction is stated here, dated, with what the entry said, what the check found and what the entry says now. Nothing is silently edited and nothing is removed: the struck sentence stays on the page so it cannot be quoted from it as a finding.

  1. 2026-10-06entry A1

    A figure absent from its document, labelled as verified

    The entry said
    Struck: Companies that scale AI report 20% to 30% cost savings in automated functions. Credited to Accenture's front-runners' guide to scaling AI, named and linked. Verdict: Verified with correction. Disposition: publish with the correction.
    The check found
    The cited document does not contain the figure. 20% appears nowhere in the 39 page PDF; 30% appears only as chart axis labels and unrelated data-source percentages, never as a cost saving. The positive control held: the same extraction returned 68,858 characters and 322 quantified tokens, 80 of them distinct, including 30%, 11%, 13% and 16%, so the probe would have found the figure had it been there. This ledger's own finding said as much on 2026-09-01 and then labelled the entry as verified, because a smaller, real figure from the same document was carried as a correction. That label was wrong. A figure that is not in its document is Not found, which is the first of the four failure shapes this page names.
    The entry now says
    A1 reads Not found, do not publish, and its claim is struck in Section A. The document's real figure, an expected 11% decrease in costs within 18 months, remains entry B4, verified in its own right. The tally is seven of nine, not six.
    Checked by
    Q4 2026 claim audit, row Q4-01
A fountain pen resting on an open, unwritten notebook beside a cup of coffee: an illustrative image for a ledger kept by hand.

The tally

What survived checking

Of the nine figures in the original set, two survive: A8, New York City Local Law 144's bias-audit rules, and A9, the EU AI Act's high-risk hiring obligations, only with its date rewritten. Seven cannot be published as claimed: A1, A2, A3, A4, A5, A6 and A7. One of those seven, A1, names a real document whose own figure is entry B4, roughly half the size and an expectation rather than a measurement. Until 2026-10-06 this ledger counted A1 as a corrected survivor; it was not, and the correction notice above says so. The four entries in Section B were not in the original set; they were found while checking the others and passed the same test.

Nine is a small sample and no claim is made about the population of AI statistics as a whole. These were the figures a competent team assembled in good faith, from the sources such a team reaches first, for its own use. That is the sample a buyer assembles too.

Count of ledger entries by verdict
VerdictCountEntries
Verified5A8, B1, B2, B3, B4
Verified with correction1A9
Unverifiable2A4, A6
Not found5A1, A2, A3, A5, A7

Counts are per entry, by each entry’s own verdict line. Entry B4 is the figure the A1 document actually states, so the Accenture document appears twice: once as A1’s absence and once as B4’s figure. A gate resolving a figure reads the entry’s verdict, never this rollup.

Method

The same five steps for every figure

  1. 01

    Find the document the claim names.

    Not an article citing it. The document itself: title, publisher, date.

  2. 02

    Read it and search it for the number.

    Where the source is a PDF, the full text was extracted and searched. Where it is a site, a site-restricted search of the publisher's own domain was run.

  3. 03

    Prove the search could have found it.

    Before accepting any absence, the probe was shown to be live: the extraction returned the document's real text, the site search returned that publisher's pages, and other percentage figures from the same file surfaced. An absence check that cannot demonstrate it would have found the thing is worthless.

  4. 04

    Check that the sentence does not overstate the source's sentence.

    A forecast is not a result. An expectation is not a measurement. One industry is not all industries. A prediction about 2029 is not a present-tense capability.

  5. 05

    Stop at a block.

    Several publishers answer automated requests with HTTP 403. Each was recorded as unread and abandoned on first refusal, with no retries and no attempt to work around the block. Those sources are listed in the appendix below. No paid tool, trial, subscription or account was used.

Section A

The 9 figures as claimed, checked

Each entry states the claim as it circulates, the verdict, whether the figure may be rendered anywhere on this site, the document actually read or the place the claim traces to, the finding, and the caveat that must travel with any surviving figure. Entries are cited by id; the ids are stable across revisions.

  • A1Not foundrevised 2026-10-06

    Claim as it circulates, struck

    Struck, not supported by its source: Companies that scale AI report 20% to 30% cost savings in automated functions.

    Disposition
    do not publish
    Source read
    Accenture (Accenture Research), The front-runners' guide to scaling AI: Lessons from industry leaders. Published 2025-05-06. Open the documenthttps://www.accenture.com/content/dam/accenture/final/accenture-com/document-3/Accenture-Front-Runners-Guide-Scaling-AI-2025-POV.pdf
    Quoted
    “an 11% improvement in customer experience and an 11% decrease in costs” (the source’s own words; not a Praxis result)

    The document exists; the number does not. The 39 page PDF was downloaded and all 68,858 characters of its text extracted and searched. The probe is live: it returns 75 distinct percentage tokens from the file. Against that working probe the string 20% occurs zero times, 20 to 30 occurs zero times, and the phrase "automated functions" never appears.

    What the source states is an 11% decrease in costs within 18 months, alongside a 13% productivity increase, a 12% revenue increase and an 11% customer experience improvement. Those are what the surveyed organisations expect, not what anyone measured; the source's own sentence is "these organizations expect". The corrected, publishable form is entry B4.

    The nearest larger figure in the document is confined to insurance underwriting, where companies scaling AI forecast a 16% decline in organisational costs. That is one industry, one function, and a forecast rather than a realised outcome, and it must be labelled as such if it is ever used.

    Carries with it. Do not publish the 20% to 30% figure. The 11% figure may be used only with the word expect.

    Revised 2026-10-06. Verdict relabelled from Verified with correction to Not found, and the disposition from publish with the correction to do not publish. The figure as claimed is absent from the document, which is this ledger's own first failure shape, and the earlier label let a corrected figure stand in for a verified claim. Nothing about the evidence changed. The document's real figure is still entry B4, verified in its own right. The relabel follows the Q4 2026 claim audit, row Q4-01, which re-derived the absence with its own probe on 2026-10-06.

  • A2Not found

    Claim as it circulates, struck

    Struck, not supported by its source: Back-office functions are 30% to 40% of mid-size operating costs (McKinsey).

    Disposition
    do not publish
    Traces to
    A single vendor page, stealthagents.com, "AI Back-Office Automation Statistics 2026" (2026-05-23), a virtual-assistant outsourcing company, which asserts the figure with the bare attribution "per McKinsey estimates" and links to no document, report, title or date. https://stealthagents.com/research/ai-back-office-automation-statistics-2026

    No McKinsey document states this. McKinsey's own reachable material says support functions run 15 to 20 percent of a global company's personnel expenses, and that fixed and support costs typically range from 5 to 15 percent of operating costs. Both are materially lower than the claim and are measured against different denominators.

    A vendor's uncited invocation of a consultancy's name is exactly the pattern a source rule exists to stop. The claim is not merely uncited; against McKinsey's reachable numbers it appears inflated by a factor of two to three.

  • A3Not foundrevised 2026-10-06

    Claim as it circulates, struck

    Struck, not supported by its source: Top-quartile AI customer-support deployments cut costs 53% (McKinsey).

    Disposition
    do not publish
    Traces to
    AI-generated statistics listicles (lorikeetcx.ai, thestacc.com, aibusinessweekly.net). The Lorikeet article, the most cited of these, was fetched and the 53% figure is not in it at all, despite the search engine attributing it there.

    A site-restricted search of mckinsey.com for 53 percent against cost reduction and top quartile returns McKinsey pages, so the probe reaches McKinsey's index and is live. The only 53% it surfaces on mckinsey.com is unrelated: the share of executives who thought cost-cutting helped their companies weather the financial crisis. No McKinsey page states a 53% support cost reduction.

    The attribution chain is not merely weak. It is partly fabricated by the retrieval layer itself: the intermediary named as the source does not contain the number either. A reader following the citation one hop, and finding a plausible-looking article, would come away satisfied and would be wrong.

    Revised 2026-10-06. The Q4 2026 claim audit, row Q4-03, records this same claim as Unverifiable rather than Not found. That edition did not re-run the mckinsey.com site-restricted search this verdict rests on, and it asserts no absence on a control it did not reproduce. The two verdicts rest on different evidence taken five weeks apart and agree on the disposition: do not publish. This entry keeps Not found because its control is recorded above and was live on 2026-09-01.

  • A4Unverifiable

    Claim as it circulates, struck

    Struck, not supported by its source: IBM says AI customer service reduces cost by roughly 30%.

    Disposition
    do not publish
    Traces to
    Gartner press release, "Gartner Predicts Agentic AI Will Autonomously Resolve 80% of Common Customer Service Issues Without Human Intervention by 2029", 2025-03-05 (title verified on gartner.com; body unreadable, HTTP 403). https://www.gartner.com/en/newsroom/press-releases/2025-03-05-gartner-predicts-agentic-ai-will-autonomously-resolve-80-percent-of-common-customer-service-issues-without-human-intervention-by-20290

    ibm.com returned HTTP 403 on both the IBM Institute for Business Value report and the IBM Think article; gartner.com returned HTTP 403 on the underlying press release. Each was abandoned on first refusal with zero retries.

    The 30% is near-certainly misattributed. It is not an IBM measurement of realised savings; it is Gartner's forecast that by 2029 agentic AI will autonomously resolve 80% of common customer service issues, "leading to a 30% reduction in operational costs". IBM's page cites Gartner for it. Gartner's own release title states the 80%-by-2029 prediction, which is first-party confirmation that the prediction is Gartner's. Presenting a 2029 forecast as a current, IBM-measured result would be a material misrepresentation.

    Carries with it. If a forecast is genuinely wanted, it can be published only as "Gartner predicts, by 2029", and only once someone can read the release.

  • A5Not found

    Claim as it circulates, struck

    Struck, not supported by its source: Finance reports 50% to 70% workload reduction.

    Disposition
    do not publish
    Traces to
    Exclusively software vendors and outsourcing firms selling finance automation (softco.com, ramp.com, mindstudio.ai, pinnasys.com, everworker.ai, wgcpas.com).

    No study, survey or primary research produces this range. Each vendor states a different range against a different denominator: 70% to 80% less processing time per invoice; up to 50% reductions in invoice processing and payment time; 30% to 50% close-time reduction; 50% to 90% of manual work; 60% to 75% efficiency gains. These are not the same measurement and cannot be combined into a single workload-reduction figure.

    The range appears to be a synthesis of incompatible vendor claims rather than a finding.

  • A6Unverifiable

    Claim as it circulates, struck

    Struck, not supported by its source: 50% to 80% of routine questions are answered automatically.

    Disposition
    do not publish
    Traces to
    The 80% upper bound is the Gartner 2029 forecast in entry A4. The 50% lower bound has no traceable source at all.

    The claim silently converts a forward-looking prediction into a present-tense capability. The Gartner release is behind a 403 bot block; stopped, zero retries.

  • A7Not foundrevised 2026-10-06

    Claim as it circulates, struck

    Struck, not supported by its source: 68% of mid-market vs 31% of enterprise AI reaches production; median time 4.2 vs 13.6 months.

    Disposition
    do not publish
    Traces to
    A single marketing blog, axistudio.io, "Why Mid-Market Companies Are Beating Enterprise on AI in 2026" (2026-05-04). https://www.axistudio.io/blog/mid-market-vs-enterprise-ai-adoption-2026

    The blog attributes the numbers to "a 2026 BCG analysis of 1,200 companies across North America" and "a parallel 2026 McKinsey survey of 580 operators" and provides no URL, no report title and no methodology for either.

    Liveness check on the attribution: a site-restricted search of bcg.com surfaces BCG's 2026 AI research, including AI Radar 2026 (roughly 2,400 executives including 640 CEOs across 16 markets), so BCG's index is reachable. No BCG study with a 1,200-company North American mid-market sample was locatable. The named studies could not be shown to exist.

    These four numbers are the most quotable in the set and the least supported. The honest framing for the mid-market argument is entry B1.

    Revised 2026-10-06. The Q4 2026 claim audit, row Q4-05, records this same claim as Unverifiable rather than Not found, for the reason it states on its own page: the bcg.com liveness search this verdict rests on was not re-run in that edition. Different evidence, same disposition: do not publish. This entry keeps Not found because its control is recorded above and was live on 2026-09-01.

  • A8Verified

    Claim as it circulates

    NYC Local Law 144: annual independent bias audit, published summary, 10 business days' candidate notice, liability on the employer rather than the vendor.

    Disposition
    publish with the correction
    Source read
    New York City Department of Consumer and Worker Protection (DCWP), Notice of Adoption of Final Rule, adding Subchapter T to Title 6 of the Rules of the City of New York, implementing Local Law 144 of 2021. Published Final rule adopted April 2023; proposed 2022-09-23, second version 2022-12-23; enforcement began 2023-07-05. Open the documenthttps://rules.cityofnewyork.us/wp-content/uploads/2023/04/DCWP-NOA-for-Use-of-Automated-Employment-Decisionmaking-Tools-2.pdf

    The 10 page official PDF was parsed. Liveness confirmed: "bias audit" occurs 55 times in the extracted text, so the probe reads the operative text rather than a wrapper.

    Verified provisions, from the rule: an employer may not use or continue to use an automated employment decision tool if more than one year has passed since the most recent bias audit (Section 5-301(a)); the date of the most recent audit and a summary of its results must be published on the employment section of the employer's website in a clear and conspicuous manner (Section 5-302); notice must be given at least 10 business days before use, satisfiable by website notice, job posting, or mail or email; and the duty runs to employers and employment agencies, not to the tool vendor, confirmed structurally by the rule barring an auditor tied to the vendor from counting as independent.

    One wording correction: the law does not command an annual audit in the abstract. It conditions use of the tool on an audit conducted within the preceding year. An employer that stops using the tool owes no audit. Write it as "a bias audit no more than one year old", not "an annual audit".

  • A9Verified with correction

    Claim as it circulates

    EU AI Act high-risk obligations for hiring take effect 2 August 2026, including human oversight, six-month logging and candidate notice.

    Disposition
    publish with the correction
    Source read
    European Commission, AI Act, Shaping Europe's digital future (page last updated 2026-08-03). Published 2026-08-03. Open the documenthttps://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai
    Quoted
    “including biometrics, critical infrastructure, education, employment ... will apply from 2 December 2027” (the source’s own words; not a Praxis result)

    Correct when the claim was written, now wrong. High-risk obligations for standalone Annex III systems, the category covering employment and recruitment, were postponed from 2 August 2026 to 2 December 2027. High-risk AI embedded in regulated products under Annex I moves to 2 August 2028. The instrument is Regulation (EU) 2026/1744 of 8 July 2026, amending Regulations (EU) 2024/1689, (EU) 2018/1139 and (EU) 2023/1230 (the Digital Omnibus on AI), published in the Official Journal 2026-07-24 and in force 2026-07-27.

    The substantive obligations are unchanged in content, only in date: deployer duties under Article 26 still include assigning human oversight to natural persons with the necessary competence, training and authority; retaining automatically generated logs for at least six months; and, where the deployer is an employer, informing workers' representatives and affected workers before putting a high-risk system into use at the workplace.

    Verification gap, stated plainly: the consolidated legal text at EUR-Lex could not be read. Both the ELI page and the Official Journal HTML endpoint returned an empty body (HTTP 202 interstitial), treated as a bot check and not retried. The date correction rests on the European Commission's own official page, not on the OJ text. Anyone relying on it should open the EUR-Lex text once in a browser and confirm the amended Article 113.

    Carries with it. Publish only the 2 December 2027 date. The 2 August 2026 date is now false.

Pattern

Four failure shapes, and they repeat

Every failed entry belongs to one of four types. The types are cheap to test for, which is why the method above catches most of what fails in under a minute per figure.

  1. 01

    A real document, and the number is not in it.

    Entry A1. The report is real and readable; the range attached to it appears nowhere in the file, and the true figure is roughly half the size and describes an expectation rather than a measurement.

  2. 02

    A consultancy's name used as a citation, with no document behind it.

    Entries A2 and A5. A figure, a firm's name, and nothing else: no title, no year, no link. A name is not a citation. If a figure arrives with a firm attached but no document, the firm has not said it.

  3. 03

    A forecast about 2029, wearing the tense of a present-day result.

    Entries A4 and A6. Underneath, one Gartner prediction about the end of the decade, re-attributed to IBM and re-tensed as a realised saving. The tense change is the worst of the three transformations, because a prediction is being used to underwrite a budget decision this quarter.

  4. 04

    An attribution manufactured in the retrieval layer.

    Entries A3 and A7. Neither the named publisher nor the named intermediary ever printed the number. A citation checked at one hop looks fine and is fabricated at two. This is one observation, not a study of retrieval systems, and no rate is claimed.

Section B

Figures that passed the same test

Found while checking the others. Each is from a government statistical agency, a central bank, a peer-reviewed journal, or the corrected reading of a document in Section A, each is free to reach, and none is published by a firm selling the thing being measured. The caveat on each entry is part of the figure and should never be detached from it.

  • B1Verified

    Figure as the source states it

    US AI adoption rises steeply with firm size: 37% of firms with at least 250 employees, 32% of firms with 100 to 249 employees, and less than 20% of firms with four or fewer. Use increased among firms with at least 20 employees and did not change significantly among firms with fewer than 20.

    Disposition
    publish
    Source read
    US Census Bureau, Large Firms With at Least 20 Employees Biggest AI Users, by Adam Grundy, Cory Breaux and Dhanapati Khatiwoda; Business Trends and Outlook Survey (BTOS). Published 2026-05-26. Open the documenthttps://www.census.gov/library/stories/2026/05/ai-use-businesses.html

    Federal statistical agency, named authors, named survey, reference period 2025-12-14 to 2026-05-03. This is the honest framing for the mid-market argument the A7 numbers were meant to carry, and unlike A7 it is real.

    Carries with it. Measures who is using AI, not what it did for them.

  • B2Verified

    Figure as the source states it

    About 18 percent of firms had adopted AI as of year-end 2025 (firm-weighted, BTOS); about 41 percent of workers used generative AI at work as of November 2025; 78 percent of the labour force works at firms that have adopted AI (employment-weighted). Professional services roughly 33% and financial roughly 30% adoption.

    Disposition
    publish
    Source read
    Board of Governors of the Federal Reserve System, FEDS Notes, "Monitoring AI Adoption in the US Economy", by Jeffrey S. Allen. Published 2026-04-03. Open the documenthttps://www.federalreserve.gov/econres/notes/feds-notes/monitoring-ai-adoption-in-the-u-s-economy-20260403.html

    Firm-level adoption is far lower than worker-level use, and adoption concentrates in large firms.

    Carries with it. This note measures adoption only. It contains no measured productivity or cost effect. Do not let it imply savings.

  • B3Verified

    Figure as the source states it

    Access to a generative AI assistant raised customer support agent productivity, measured as issues resolved per hour, by 15% on average across 5,172 agents, with the largest gains among the least experienced and lower-skilled workers; the most experienced and highest-skilled saw small gains in speed and small declines in quality.

    Disposition
    publish
    Source read
    Oxford University Press, The Quarterly Journal of Economics, Erik Brynjolfsson, Danielle Li and Lindsey R. Raymond, "Generative AI at Work", QJE vol. 140, issue 2, p. 889, DOI 10.1093/qje/qjae044 (author-hosted published version: http://danielle.li/assets/docs/GenerativeAIatWork.pdf; NBER working paper w31161: https://www.nber.org/papers/w31161). Published 2025-02-04 (NBER working paper issued April 2023, revised November 2023). Open the documenthttps://doi.org/10.1093/qje/qjae044

    Read in two independent versions. The strongest number in the set, and it is not close: peer-reviewed in a top-five economics journal rather than self-published by a firm selling the thing being measured; a staggered rollout across thousands of named-sample workers rather than an executive opinion survey; a realised, measured outcome rather than an expectation or a forecast; free to read in full.

    Version discrepancy, disclosed: the NBER working paper states 14% average and 34% for novice and low-skilled workers across 5,179 agents. The published QJE version states 15% across 5,172 agents. Cite the QJE figures as the version of record. Do not mix the 34% from the working paper with the 15% from the journal.

    Carries with it. A productivity result from a single large firm's support organisation, not a cost-savings result. It must be presented as another company's published finding, in its own terms, with the sample and the journal named. It must not be converted into a percentage of cost, annualised, or extrapolated to a reader's business.

  • B4Verified

    Figure as the source states it

    Organisations scaling generative AI expect an 11% decrease in costs and a 13% productivity increase within 18 months.

    Disposition
    publish
    Source read
    Accenture (Accenture Research), The front-runners' guide to scaling AI: Lessons from industry leaders. Published 2025-05-06. Open the documenthttps://www.accenture.com/content/dam/accenture/final/accenture-com/document-3/Accenture-Front-Runners-Guide-Scaling-AI-2025-POV.pdf

    The corrected, publishable form of entry A1. Same document, same URL.

    Carries with it. These are what surveyed executives expect, not what was measured. Any copy rendering this figure must say "expect".

Appendix

Sources that could not be read

Recorded rather than worked around. Each was abandoned on first refusal, with zero retries and no attempt to bypass a block. A source that could not be read is a different state from a source that disagrees, and it is kept distinct here so that a gap cannot quietly become a fact later.

Sources refused or unreadable during verification
SourceURLResult
NYC DCWP AEDT landing pagehttps://www.nyc.gov/site/dca/about/automated-employment-decision-tools.pageHTTP 403
American Legal Publishing, NYC Admin Code Section 20-871https://codelibrary.amlegal.com/codes/newyorkcity/latest/NYCadmin/0-0-0-135843HTTP 403 bot block
EUR-Lex, Regulation (EU) 2026/1744https://eur-lex.europa.eu/eli/reg/2026/1744/oj/engHTTP 202, empty body
EUR-Lex OJ HTML endpointhttps://eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=OJ:L_202601744HTTP 202, empty body
Gartner press release, agentic AI 2029https://www.gartner.com/en/newsroom/press-releases/2025-03-05-gartner-predicts-agentic-ai-will-autonomously-resolve-80-percent-of-common-customer-service-issues-without-human-intervention-by-20290HTTP 403 bot block
IBM Institute for Business Value, AI customer servicehttps://www.ibm.com/thought-leadership/institute-business-value/en-us/report/ai-customer-serviceHTTP 403
IBM Think, future of customer servicehttps://www.ibm.com/think/insights/customer-service-futureHTTP 403
McKinsey (several article URLs)https://www.mckinsey.com/Repeated 60 second timeouts
Oxford Academic QJE article pagehttps://academic.oup.com/qje/article/140/2/889/7990658Returned a navigation shell only; resolved via the author-hosted published PDF instead

Revisions

What changed, and when

  1. 2026-10-06Entry A1 relabelled from Verified with correction to Not found and its disposition set to do not publish; the claim is now struck in Section A and a dated correction notice sits above the tally. Prompted by the Q4 2026 claim audit, row Q4-01. Entries A3 and A7 carry a note reconciling their verdicts with that audit's rows Q4-03 and Q4-05. The tally and the lede restated accordingly: seven of nine, not six. No evidence changed.
  2. 2026-09-12Relocated from the Journal to this permanent research page. Entries, verdicts and sources unchanged from the 2026-09-01 retrieval. The Journal URL now redirects here.
  3. 2026-09-09First published, as a Journal post, with the tally and the four failure modes.

Cite this ledger

Praxis Consulting Company, "AI savings statistics: the verification ledger", retrieved 2026-09-01, revised 2026-10-06, https://praxisconsultingco.com/research/ai-savings-statistics

Cite an individual figure by its entry id and the source named beside it, not by this page alone: the page is the record of the check, the document is the evidence. A reader who cited an earlier revision can compare it against the changelog on the left.

Have a business case checked this way

If you are evaluating an AI business case and want the numbers in it traced to their documents, tell us what you are looking at. The first conversation is free and ends with a plain list of what held and what did not.

Request a free quoteEmail usCall (415) 634-5434

No obligation · a scoping conversation first