Skip to content
PRAXIS

R.02Research · Claim audit

The Mid-Market AI Claim Audit, Q4 2026

Of 43 quantified AI claims checked in Q4 2026, 34 could not be traced to a primary document containing the figure. A quarterly, instrumented check of the quantified claims that mid-market AI consulting pages publish about what AI does to cost, time, headcount and output. Each claim is read word for word, traced to the document it is credited to, and searched for the figure, and every absence carries the positive control proving the search would have found it. The sample is small and the method is published, which is the shape that makes it citable. Praxis Consulting Company publishes this page, is inside the population it audits, and carries this edition's only proven miss.

Disclosure

The publisher of this page is inside the population it audits

  • Praxis Consulting Company publishes this page and is one of the firms whose claims it checks. Praxis is owned by Evan Dang, who also owns EliteSEO Consulting, a separate practice that does not sell AI implementation and is not in this frame.
  • Praxis sells AI implementation consulting, so it sits inside the population this audit examines. Its own figures went through the same pass, in the same order, with the same probe.
  • The result is that Praxis carries this edition's only proven NOT FOUND, on its own page, against a document it named and linked itself. Three of its seven rows name no document a reader could open; a fourth names one that does not contain the figure.
  • No firm paid to appear here, no firm was given sight of its rows before publication, and no firm was left out because of what its rows would say.
  • The frame was derived from a page published before this one, by a script, so it could not be shaped to produce a better headline. The script and the raw observations are in the site's repository.
  • Quotations are reproduced word for word with one exception: an em dash or en dash inside a quotation is rendered as a comma, because this site's copy gate permits neither character. No word is added, removed or reordered.

Finding, Q4 2026

One number, and the sample it came from

Of 43 quantified AI claims checked in Q4 2026, 34 could not be traced to a primary document containing the figure.

The Mid-Market AI Claim Audit, Q4 2026. Retrieval window 2026-10-06, one day. https://praxisconsultingco.com/research/mid-market-ai-claim-audit

43

Claims checked

11 of 23 in the frame

Publishers with a claim in the table

89 / 81 / 8

Pages requested, read, unread

34 of 43

Could not be traced to a document holding the figure

Sheets of white paper laid out on a wooden table: an illustrative image for a register of claims set out one per row.

43 claims from one day of reading is a small sample. It has a published method, a frame derived by a script from a page published before this one, and a positive control on every absence, and those are the properties that make a small sample worth citing. The numbers on this page describe these 43 claims and nothing else. No rate, trend or industry generalisation is offered or supported.

Method

How each row was produced

Stated before the findings, and reusable word for word next quarter. The scripts are in the site’s repository, the raw observations are committed next to them, and re-running them on the same day reproduces the same rows.

  1. 01

    Fix the frame before looking at anything.

    The frame is derived from an already-published document, not chosen. Every firm compared on /research/ai-implementation-firms-compared is in it, and that page's own cells name the sampled pages: the URLs its criteria 4, 5, 6 and 7 record, plus each firm's homepage. A script reads them out of the source file, so re-running it reproduces the frame exactly and no page can be added because it looked promising.

  2. 02

    Count one quantified claim at a time.

    The unit is a single public claim, not a company and not an article, so one page contributes several rows. A claim counts when it states a number about what AI does to cost, time, headcount or output, on a public page selling or advising on AI to mid-market buyers, presented as fact rather than as the publisher's own labelled forecast. Everything dropped is listed below with the condition it failed.

  3. 03

    Read the text around the claim before recording its source.

    The extractor works sentence by sentence, so a source printed in the next sentence, under a chart or in a caption is invisible to it. Reporting those as unattributed would understate every publisher's sourcing and would be a false finding about a named firm, so the attribution recorded is the one a human read near the claim.

  4. 04

    Find the document, then search it for the number.

    Not an article citing it. The document itself. Where it is a PDF the full text is extracted and searched; where it is a page, the page text is. A figure is matched in every written form the audit treats as the same number, which errs towards verifying rather than towards accusing.

  5. 05

    Prove the search could have found it.

    Every row reporting a figure as absent carries its own control: the extracted character count and the number of other quantified tokens the same extraction returned. A probe that returns nothing is a failed probe, and a failed probe produces UNVERIFIABLE, never NOT FOUND. An absence check that cannot demonstrate it would have found the thing is worthless.

  6. 06

    Read every match in its sentence.

    A present token is not a verdict. 30% appears in one of the documents below only as chart axis labels, so the claim credited to it still fails. Presence has to be read; only absence is settled by the presence test and its control.

  7. 07

    Mutation-test the checker, both directions, before publishing.

    A figure known to be in a fixture must come back present, and a figure known to be absent must come back absent, along with the variant forms and the liveness rule. A checker that could not find a figure genuinely present would report every claim as unsourced, and this page would be both worthless and unfair to every firm in it. The script refuses to run at all if that test fails.

  8. 08

    Stop at a block.

    A 403, a 202 challenge, a timeout or a page that serves only a script shell is recorded with its status and abandoned on the first refusal. No retries, no header change, no address change, nothing solved. Those pages and documents are listed below. No paid tool, trial, account or subscription was used.

Sampling frame

Who is in the sample, and why they are

The frame is derived from a page this site published first, not chosen for this one. Every firm compared on the eight-criteria comparison is in it, 23 firms in all, and that page’s own cells name the pages sampled here: the URLs its criteria four, five, six and seven record, plus each firm’s homepage. A script reads them out of the source file, so the frame can be reproduced and cannot be quietly reshaped to produce a better headline.

Criterion five was added to that page rule before any finding was read, for one reason. Without it the frame left out the publisher’s own statistics page, which is where Praxis’s quantified figures actually live, while keeping other firms’ article pages in. A frame that spares the auditor is the failure this artefact exists to measure. The rule is applied to all 23 rows identically.

That gives 89 pages, of which 81 were served and read and 8 were not. Reading them produced 81 candidate sentences, of which 43 are in the table and the rest are listed below with the scope condition each failed.

The half of the frame that could not be built. The scope approved on 2026-10-04 asked for the organic top twenty on a fixed list of twelve mid-market AI buyer queries. That capture could not be made on 2026-10-06 at zero cost without a browser, and it is recorded rather than worked around: one search engine answered the automated request with unrelated results, another returned a challenge page, and the two paid tools had no plan and no units. Each was abandoned on the first refusal. Executing that half is the first widening of the next edition, and until it happens this frame is an incumbents’ frame and the page says so.

Publishers with at least one claim in the table: Auxis, BairesDev, Genpact, HatchWorks AI, LeewayHertz, LOW/CODE Agency, Markovate, Praxis Consulting Company, SeidrLab, Velsof (Velocity Software Solutions), Vention. The other firms in the frame published no sentence meeting all four scope conditions on the pages sampled, or could not be read at all.

What the rows show

Findings, stated without generalising past the sample

  1. Most of these figures are not sourced at all, rather than sourced badly.

    The commonest shape in this edition is not a misquoted document. It is a number with nothing behind it: no document, no link, no sample, no period, not even a firm's name. That is why it is named as its own mode rather than folded into the one R.01 recorded for a firm's name used as a citation.

  2. The one absence proven against a document belongs to the publisher of this page.

    Praxis Consulting Company's own page carries a figure credited to a named, linked report in which the lower bound appears nowhere and the upper bound appears only as chart axis labels. It is this edition's only NOT FOUND, it is shown in the same table under the same rule, and it is the row a reader should check first.

  3. A stated sample size is rare, and it is what separates a checkable figure from an uncheckable one.

    Two publishers in this frame state a method and a sample for their own survey on the same page as the figures. Every figure traceable to one of those two statements is marked verified here. Everything else about an own-survey figure, however precise, could not be checked by anyone.

  4. A figure restated four times and attributed once reads as four facts and one citation.

    One report page states the same percentage four times. Three of those appearances carry nothing. The fourth names a company, roughly 30,000 characters after the first, with no document title and no link. A reader who meets the figure early has nothing to check, which is a shape worth naming because it is invisible to any count of citations per page.

  5. Labelling an estimate as an estimate was the clearest sourcing practice observed here, and it took the claim out of scope.

    One publisher writes "could", "estimations" and "potentially" around six figures and links the document each one comes from. Those six are out of scope by the audit's own fourth condition, because they are not presented as fact. That is the correct outcome and it is worth stating plainly: the way to stay out of this table is to say what a number is.

  6. Saying a figure is not a statistic is better than pretending it is.

    One page states that it is sharing patterns from client engagements and not statistical claims from a controlled sample. The figure is still uncheckable and still appears in the table as such, but the disclosure is the most honest handling of an uncheckable figure in this edition.

  7. Blocks are the reason the rest can be trusted.

    Eight pages in the frame could not be read and are listed with their statuses, and the document behind six linked claims refused the request outright. Those are recorded rather than guessed at, and nothing in the table rests on a page that was not served.

Distribution by verdict

Verdict distribution across the 43 claims checked
VerdictClaims
Verified9
Verified with correction0
Not found1
Unverifiable33

Distribution by transformation

How the 43 claims depart from the document they are credited to
ShapeClaims
none: the figure survives9
no attribution printed at all24
a name used as a citation, with no document behind it7
a real document, and the number is not in it1
a forecast wearing the tense of a present-day result1
an attribution manufactured in the retrieval layer1

The first four shapes below the surviving row are the ones the 2026-09-01 ledger named, carried forward word for word. The commonest shape in this edition, a number with no attribution printed at all, is named here for the first time rather than folded into one of them, because folding it in would have hidden how common it is.

The table

Every claim, its source and its verdict

Rows are grouped by publisher, alphabetically, and the publisher of this page sorts in with the rest. Nothing is ranked. The full record for each row, including the document, what it says, and the positive control where one applies, is in the section below the table.

The 43 quantified AI claims checked in Q4 2026, with publisher, figure, attribution as printed, verdict and transformation
RowPublisherFigureAttribution as printedVerdictShape
Q4-01Praxis Consulting Companypublisher of this page20% to 30%.Accenture, "The Front-Runners' Guide to Scaling AI", named and linked on the page.Not founda real document, and the number is not in it.
Q4-02Praxis Consulting Companypublisher of this page11%.Accenture, "The Front-Runners' Guide to Scaling AI", named and linked on the page.Verifiednone: the figure survives.
Q4-03Praxis Consulting Companypublisher of this page53%.McKinsey, a firm name in parentheses, with no document title, no date and no link.Unverifiablea name used as a citation, with no document behind it.
Q4-04Praxis Consulting Companypublisher of this page30%.IBM, a firm name in the sentence, with no document title and no link.Unverifiablea forecast wearing the tense of a present-day result.
Q4-05Praxis Consulting Companypublisher of this page68% and 31%.no primary attribution. The figures trace to one marketing blog that credits "a 2026 BCG analysis of 1,200 companies" and "a parallel 2026 McKinsey survey of 580 operators", with no URL, no report title and no method for either.Unverifiablean attribution manufactured in the retrieval layer.
Q4-06Praxis Consulting Companypublisher of this page18 percent, 41 percent, 78 percent.Board of Governors of the Federal Reserve System, FEDS Notes, "Monitoring AI Adoption in the US Economy", Jeffrey S. Allen, 2026-04-03, named and linked on the page.Verifiednone: the figure survives.
Q4-07Praxis Consulting Companypublisher of this page15% across 5,172 agents.Brynjolfsson, Li and Raymond, "Generative AI at Work", The Quarterly Journal of Economics vol. 140 issue 2 p. 889, DOI 10.1093/qje/qjae044, named and linked on the page in two independent versions.Verifiednone: the figure survives.
Q4-08Auxis34% and 66%.Grant Thornton's Q1 CFO survey, named by title, with no link.Unverifiablea name used as a citation, with no document behind it.
Q4-09Auxismore than 90%.none printed. The figure sits inside a quoted statement by the firm's own executive.Unverifiableno attribution printed at all.
Q4-10BairesDev13 hours a week.the publisher's own survey, with the method on the same page: "The Q3 2026 edition of the Dev Barometer was fielded in August 2026 among 705 software developers across 60+ countries".Verifiednone: the figure survives.
Q4-11BairesDev51% and 67%.the publisher's own survey, method and sample of 705 stated on the same page.Verifiednone: the figure survives.
Q4-12BairesDev9 hours a week.the publisher's own survey, method and sample of 705 stated on the same page.Verifiednone: the figure survives.
Q4-13BairesDev59% and 52%.the publisher's own survey, method and sample of 705 stated on the same page.Verifiednone: the figure survives.
Q4-14BairesDev58%.the publisher's own survey, method and sample of 705 stated on the same page.Verifiednone: the figure survives.
Q4-15Genpact$18 trillion, 8%, 16%."our research", with no document title, no sample, no method and no link.Unverifiablea name used as a citation, with no document behind it.
Q4-16HatchWorks AIalmost 300%.none printed. The figure sits inside a quoted client statement with no measurement period, baseline or method.Unverifiableno attribution printed at all.
Q4-17LOW/CODE Agency40%.none printed. A result card on the firm's own service page, with no baseline, period or method.Unverifiableno attribution printed at all.
Q4-18Markovate70%.none printed. A homepage result card, with no baseline, period or method.Unverifiableno attribution printed at all.
Q4-19SeidrLab50 to 80 percent.none printed, and the page says so itself: "We're sharing the patterns we see consistently in client engagements, not statistical claims from a controlled sample".Unverifiableno attribution printed at all.
Q4-20Velsof (Velocity Software Solutions)30-40%.none printed.Unverifiableno attribution printed at all.
Q4-21Velsof (Velocity Software Solutions)1-2% and 2-3%.none printed.Unverifiableno attribution printed at all.
Q4-22Velsof (Velocity Software Solutions)6-12 months.none printed.Unverifiableno attribution printed at all.
Q4-23Velsof (Velocity Software Solutions)40%.none printed. An article title in the page's own related-reading list.Unverifiableno attribution printed at all.
Q4-24Velsof (Velocity Software Solutions)$2M.none printed. An anonymised client result on the firm's own service page.Unverifiableno attribution printed at all.
Q4-25Velsof (Velocity Software Solutions)$4.5M.none printed. An anonymised client result on the firm's own service page.Unverifiableno attribution printed at all.
Q4-26Vention98%.none printed, and none anywhere in the surrounding text.Unverifiableno attribution printed at all.
Q4-27Vention$500 million.none printed.Unverifiableno attribution printed at all.
Q4-28Vention20-60x.none printed.Unverifiableno attribution printed at all.
Q4-29Ventionup to 50%.none printed, and no provider is named.Unverifiableno attribution printed at all.
Q4-30Vention95.47% and 100%.the publisher's own engagement, described as "provider-attested", with no document, baseline or measurement period.Unverifiableno attribution printed at all.
Q4-31Vention56% and 57%.the publisher's own survey, with the method on the same page: "The Vention BIXA research methodology was based on surveying 480 qualified decision-makers from the US, UK, and DACH, actively investing in software development".Verifiednone: the figure survives.
Q4-32Vention68%.nothing on three of its four appearances. The fourth reads "According to Atlassian, 68% of developers reported significant time savings of more than 10 hours per week from using AI", naming Atlassian with no document title and no link.Unverifiablea name used as a citation, with no document behind it.
Q4-33Vention54%.none printed.Unverifiableno attribution printed at all.
Q4-34Vention15% to 20%, and 2x to 3x.none printed for either figure.Unverifiableno attribution printed at all.
Q4-35Vention63%.none printed. A statistic card with no source marker.Unverifiableno attribution printed at all.
Q4-36Vention4 times and 10 times."another study". No publisher, no title, no date, no link.Unverifiablea name used as a citation, with no document behind it.
Q4-37Vention16% to 30%, and 31% to 45%.McKinsey, named twice on the page, with no document title and no link either time.Unverifiablea name used as a citation, with no document behind it.
Q4-38Vention10% to 15%.none printed.Unverifiableno attribution printed at all.
Q4-39Vention73%.none printed.Unverifiableno attribution printed at all.
Q4-40LeewayHertz56 percent."one study". No publisher, no title, no date, no link.Unverifiablea name used as a citation, with no document behind it.
Q4-41LeewayHertz60%.none printed. A result card with no baseline, period or method.Unverifiableno attribution printed at all.
Q4-42LeewayHertz2x.none printed. A result card with no baseline, period or method.Unverifiableno attribution printed at all.
Q4-43LeewayHertz60%.none printed. A result card with no baseline, period or method.Unverifiableno attribution printed at all.

The table scrolls sideways inside its own region on a narrow screen. Every value in it is repeated in the row records below, which do not need horizontal scrolling.

The record

Each row in full, with its document and its control

  • Q4-01Not foundcandidate Q001publisher of this page

    The claim, word for word

    Published by Praxis Consulting Company and quoted here. Not a Praxis result, and not a Praxis client figure.

    "Companies that scale AI report 20% to 30% cost savings in automated functions."

    Publisher
    Praxis Consulting Company (consultancy)
    Figure
    20% to 30%.
    Attribution as printed
    Accenture, "The Front-Runners' Guide to Scaling AI", named and linked on the page.
    Primary document
    Open Accenture, The Front-Runners' Guide to Scaling AIhttps://www.accenture.com/content/dam/accenture/final/accenture-com/document-3/Accenture-Front-Runners-Guide-Scaling-AI-2025-POV.pdfHTTP 200, application/pdf, 39 pages read with pypdf
    Figure in the document
    searched for and not in the document.
    What the document says
    20% appears nowhere in the 39-page document. 30% appears only as chart axis labels and as unrelated data-source percentages, never as a cost saving. The document's own cost figure is an 11% decrease in costs, and it is what surveyed organisations expect within 18 months, not a measured result.Those are the document’s own words. Not a Praxis result.
    Positive control on the probe
    the same extraction returned 68,858 characters and 322 quantified tokens, 80 of them distinct, including 30%, 11%, 13% and 16%.This is the evidence that the search would have found the figure had it been there. Without it the absence would not be published.
    Transformation
    a real document, and the number is not in it.
    Note
    The edition's only proven absence, and it belongs to the publisher of this page. R.01 reached the same finding on 2026-09-01; this verdict was re-derived by this edition's own probe rather than inherited. Corrected on the publisher's page on 2026-10-06: R.01 entry A1 now reads Not found, do not publish, with the claim struck and a dated correction notice above its tally.
    Read on
    The Praxis Consulting Company page this was read onhttps://praxisconsultingco.com/research/ai-savings-statisticsretrieved 2026-10-06.
  • Q4-02Verifiedcandidate Q002publisher of this page

    The claim, word for word

    Published by Praxis Consulting Company and quoted here. Not a Praxis result, and not a Praxis client figure.

    "An 11% improvement in customer experience and an 11% decrease in costs."

    Publisher
    Praxis Consulting Company (consultancy)
    Figure
    11%.
    Attribution as printed
    Accenture, "The Front-Runners' Guide to Scaling AI", named and linked on the page.
    Primary document
    Open Accenture, The Front-Runners' Guide to Scaling AIhttps://www.accenture.com/content/dam/accenture/final/accenture-com/document-3/Accenture-Front-Runners-Guide-Scaling-AI-2025-POV.pdfHTTP 200, application/pdf, 39 pages read with pypdf
    Figure in the document
    found in the document.
    What the document says
    Present and read in context: "these organizations expect a 13% increase in productivity, a 12% increase in revenue growth, an 11% improvement in customer experience and an 11% decrease in costs within 18 months".Those are the document’s own words. Not a Praxis result.
    Transformation
    none: the figure survives.
    Note
    Carries its caveat wherever it is used: this is what surveyed executives expect within 18 months, not what was measured.
    Read on
    The Praxis Consulting Company page this was read onhttps://praxisconsultingco.com/research/ai-savings-statisticsretrieved 2026-10-06.
  • Q4-03Unverifiablecandidate Q004publisher of this page

    The claim, word for word

    Published by Praxis Consulting Company and quoted here. Not a Praxis result, and not a Praxis client figure.

    "Top-quartile AI customer-support deployments cut costs 53% (McKinsey)."

    Publisher
    Praxis Consulting Company (consultancy)
    Figure
    53%.
    Attribution as printed
    McKinsey, a firm name in parentheses, with no document title, no date and no link.
    Primary document
    none located
    Figure in the document
    no document to check.
    Transformation
    a name used as a citation, with no document behind it.
    Note
    A name is not a citation. No document was located, so no absence is asserted and no probe control is claimed. R.01 recorded NOT FOUND for this claim on 2026-09-01 with a site-restricted liveness control on mckinsey.com. That control was NOT re-run on 2026-10-06, so this edition records UNVERIFIABLE rather than carrying an absence on a control it did not reproduce; R.01's entry A3 says the same from its side.
    Read on
    The Praxis Consulting Company page this was read onhttps://praxisconsultingco.com/research/ai-savings-statisticsretrieved 2026-10-06.
  • Q4-04Unverifiablecandidate Q005publisher of this page

    The claim, word for word

    Published by Praxis Consulting Company and quoted here. Not a Praxis result, and not a Praxis client figure.

    "IBM says AI customer service reduces cost by roughly 30%."

    Publisher
    Praxis Consulting Company (consultancy)
    Figure
    30%.
    Attribution as printed
    IBM, a firm name in the sentence, with no document title and no link.
    Primary document
    none located
    Figure in the document
    no document to check.
    Transformation
    a forecast wearing the tense of a present-day result.
    Note
    R.01 traced the figure to a Gartner prediction about 2029 re-attributed to IBM and re-tensed as a realised saving. The Gartner release answers automated requests with HTTP 403, so it could not be read on 2026-10-06 either and no absence is asserted.
    Read on
    The Praxis Consulting Company page this was read onhttps://praxisconsultingco.com/research/ai-savings-statisticsretrieved 2026-10-06.
  • Q4-05Unverifiablecandidate Q007publisher of this page

    The claim, word for word

    Published by Praxis Consulting Company and quoted here. Not a Praxis result, and not a Praxis client figure.

    "68% of mid-market vs 31% of enterprise AI reaches production; median time 4.2 vs 13.6 months."

    Publisher
    Praxis Consulting Company (consultancy)
    Figure
    68% and 31%.
    Attribution as printed
    no primary attribution. The figures trace to one marketing blog that credits "a 2026 BCG analysis of 1,200 companies" and "a parallel 2026 McKinsey survey of 580 operators", with no URL, no report title and no method for either.
    Primary document
    none located
    Figure in the document
    no document to check.
    Transformation
    an attribution manufactured in the retrieval layer.
    Note
    R.01 recorded NOT FOUND here on 2026-09-01 with a site-restricted liveness control. That control was NOT re-run on 2026-10-06, so this edition downgrades the verdict to UNVERIFIABLE rather than carrying an absence on a control it did not reproduce. R.01's entry A7 says the same from its side.
    Read on
    The Praxis Consulting Company page this was read onhttps://praxisconsultingco.com/research/ai-savings-statisticsretrieved 2026-10-06.
  • Q4-06Verifiedcandidate Q008publisher of this page

    The claim, word for word

    Published by Praxis Consulting Company and quoted here. Not a Praxis result, and not a Praxis client figure.

    "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)."

    Publisher
    Praxis Consulting Company (consultancy)
    Figure
    18 percent, 41 percent, 78 percent.
    Attribution as printed
    Board of Governors of the Federal Reserve System, FEDS Notes, "Monitoring AI Adoption in the US Economy", Jeffrey S. Allen, 2026-04-03, named and linked on the page.
    Primary document
    Open Federal Reserve FEDS Notes, Monitoring AI Adoption in the US Economyhttps://www.federalreserve.gov/econres/notes/feds-notes/monitoring-ai-adoption-in-the-u-s-economy-20260403.htmlHTTP 200, text/html
    Figure in the document
    found in the document.
    What the document says
    All three present and read in context: the note reports firm-level BTOS adoption, work-related generative AI adoption of about 41 percent as of November in the Real-Time Population Survey, and an employment-weighted figure from the Survey of Business Uncertainty.Those are the document’s own words. Not a Praxis result.
    Transformation
    none: the figure survives.
    Note
    Carries its caveat: the note measures adoption only. It contains no measured productivity or cost effect and must not be allowed to imply savings.
    Read on
    The Praxis Consulting Company page this was read onhttps://praxisconsultingco.com/research/ai-savings-statisticsretrieved 2026-10-06.
  • Q4-07Verifiedcandidate Q009publisher of this page

    The claim, word for word

    Published by Praxis Consulting Company and quoted here. Not a Praxis result, and not a Praxis client figure.

    "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."

    Publisher
    Praxis Consulting Company (consultancy)
    Figure
    15% across 5,172 agents.
    Attribution as printed
    Brynjolfsson, Li and Raymond, "Generative AI at Work", The Quarterly Journal of Economics vol. 140 issue 2 p. 889, DOI 10.1093/qje/qjae044, named and linked on the page in two independent versions.
    Primary document
    Open Brynjolfsson, Li and Raymond, Generative AI at Work (author-hosted published version)https://danielle.li/assets/docs/GenerativeAIatWork.pdfHTTP 200, application/pdf, 54 pages read with pypdf
    Figure in the document
    found in the document.
    What the document says
    Present in the abstract: "We study the staggered introduction of a generative AI-based conversational assistant using data from 5,172 customer-support agents. Access to AI assistance increases worker productivity", with the 15 percent figure stated for issues resolved per hour.Those are the document’s own words. Not a Praxis result.
    Transformation
    none: the figure survives.
    Note
    The strongest figure in either edition: peer reviewed, with the microdata described and the sample stated.
    Read on
    The Praxis Consulting Company page this was read onhttps://praxisconsultingco.com/research/ai-savings-statisticsretrieved 2026-10-06.
  • Q4-08Unverifiablecandidate Q056

    The claim, word for word

    Published by Auxis and quoted here. Not a Praxis result, and not a Praxis client figure.

    "Only 34% of CFOs now rank cost reduction as the primary benefit of outsourcing, while 66% prioritize more strategic outcomes such as access to talent, AI enablement and scalability/process standardization, according to Grant Thornton's Q1 CFO survey."

    Publisher
    Auxis (outsourcer)
    Figure
    34% and 66%.
    Attribution as printed
    Grant Thornton's Q1 CFO survey, named by title, with no link.
    Primary document
    Open Grant Thornton CFO survey hub page (not demonstrably the Q1 edition named)https://www.grantthornton.com/insights/survey-reports/cfo-surveyHTTP 200, text/html
    Figure in the document
    no document to check.
    What the document says
    Neither figure appears on the hub page reached. That page is live and carries 44 quantified tokens, but it is not established to be the Q1 edition the claim names, and the publisher printed no link, so no absence is asserted.Those are the document’s own words. Not a Praxis result.
    Transformation
    a name used as a citation, with no document behind it.
    Note
    A named document with no link is better practice than a bare firm name and still not checkable in one click. Observed on the same page: the quoted executive's title is printed as "Jose Alvarez, Auxis Grant Thornton Partner, IT Modernization & ITO". The words are recorded as printed and no inference is drawn from them.
    Read on
    The Auxis page this was read onhttps://www.auxis.com/key-benefits-of-nearshoring-to-latin-america-for-the-midmarket/retrieved 2026-10-06.
  • Q4-09Unverifiablecandidate Q057

    The claim, word for word

    Published by Auxis and quoted here. Not a Praxis result, and not a Praxis client figure.

    "With more than 90% of AI initiatives falling short of expectations."

    Publisher
    Auxis (outsourcer)
    Figure
    more than 90%.
    Attribution as printed
    none printed. The figure sits inside a quoted statement by the firm's own executive.
    Primary document
    none located
    Figure in the document
    no document to check.
    Transformation
    no attribution printed at all.
    Read on
    The Auxis page this was read onhttps://www.auxis.com/key-benefits-of-nearshoring-to-latin-america-for-the-midmarket/retrieved 2026-10-06.
  • Q4-10Verifiedcandidate Q059

    The claim, word for word

    Published by BairesDev and quoted here. Not a Praxis result, and not a Praxis client figure.

    "Developers report saving an average of 13 hours a week on coding with AI, up from about 7 hours in Q3 2025."

    Publisher
    BairesDev (outsourcer)
    Figure
    13 hours a week.
    Attribution as printed
    the publisher's own survey, with the method on the same page: "The Q3 2026 edition of the Dev Barometer was fielded in August 2026 among 705 software developers across 60+ countries".
    Primary document
    Open BairesDev Dev Barometer, Q3 2026 edition (the page is the report)https://www.bairesdev.com/blog/dev-barometer-q3-2026-devs-answering-for-code/HTTP 200, text/html
    Figure in the document
    found in the document.
    What the document says
    The page states the fielding month, the sample of 705 developers, the 60-plus countries, a separate base of 41 enterprise CTOs, and offers the full survey dataset. It also publishes the median alongside the average and says the average is pulled upward by a minority reporting much larger gains.Those are the document’s own words. Not a Praxis result.
    Transformation
    none: the figure survives.
    Note
    Self-asserted, like any publisher's own survey, and the only publisher in the frame that printed both the sample size and the median next to the average.
    Read on
    The BairesDev page this was read onhttps://www.bairesdev.com/blog/dev-barometer-q3-2026-devs-answering-for-code/retrieved 2026-10-06.
  • Q4-11Verifiedcandidate Q060

    The claim, word for word

    Published by BairesDev and quoted here. Not a Praxis result, and not a Praxis client figure.

    "Just over half of developers, 51%, now spend more than a quarter of their working week reviewing AI-generated work, and 67% report spending more time on that review than they did a year ago."

    Publisher
    BairesDev (outsourcer)
    Figure
    51% and 67%.
    Attribution as printed
    the publisher's own survey, method and sample of 705 stated on the same page.
    Primary document
    Open BairesDev Dev Barometer, Q3 2026 edition (the page is the report)https://www.bairesdev.com/blog/dev-barometer-q3-2026-devs-answering-for-code/HTTP 200, text/html
    Figure in the document
    found in the document.
    Transformation
    none: the figure survives.
    Read on
    The BairesDev page this was read onhttps://www.bairesdev.com/blog/dev-barometer-q3-2026-devs-answering-for-code/retrieved 2026-10-06.
  • Q4-12Verifiedcandidate Q061

    The claim, word for word

    Published by BairesDev and quoted here. Not a Praxis result, and not a Praxis client figure.

    "Developers report an average of 9 hours a week learning AI tools and new technologies, up from 4 hours a year ago."

    Publisher
    BairesDev (outsourcer)
    Figure
    9 hours a week.
    Attribution as printed
    the publisher's own survey, method and sample of 705 stated on the same page.
    Primary document
    Open BairesDev Dev Barometer, Q3 2026 edition (the page is the report)https://www.bairesdev.com/blog/dev-barometer-q3-2026-devs-answering-for-code/HTTP 200, text/html
    Figure in the document
    found in the document.
    Transformation
    none: the figure survives.
    Read on
    The BairesDev page this was read onhttps://www.bairesdev.com/blog/dev-barometer-q3-2026-devs-answering-for-code/retrieved 2026-10-06.
  • Q4-13Verifiedcandidate Q063

    The claim, word for word

    Published by BairesDev and quoted here. Not a Praxis result, and not a Praxis client figure.

    "59% report greater mental effort to catch subtle bugs in AI-generated code, the largest increase of any item we tested, ahead of the 52% reporting more time spent debugging problems AI introduced."

    Publisher
    BairesDev (outsourcer)
    Figure
    59% and 52%.
    Attribution as printed
    the publisher's own survey, method and sample of 705 stated on the same page.
    Primary document
    Open BairesDev Dev Barometer, Q3 2026 edition (the page is the report)https://www.bairesdev.com/blog/dev-barometer-q3-2026-devs-answering-for-code/HTTP 200, text/html
    Figure in the document
    found in the document.
    Transformation
    none: the figure survives.
    Note
    A figure that runs against the publisher's commercial interest, published anyway.
    Read on
    The BairesDev page this was read onhttps://www.bairesdev.com/blog/dev-barometer-q3-2026-devs-answering-for-code/retrieved 2026-10-06.
  • Q4-14Verifiedcandidate Q064

    The claim, word for word

    Published by BairesDev and quoted here. Not a Praxis result, and not a Praxis client figure.

    "Developers who barely use AI report a more fulfilling role 58% of the time."

    Publisher
    BairesDev (outsourcer)
    Figure
    58%.
    Attribution as printed
    the publisher's own survey, method and sample of 705 stated on the same page.
    Primary document
    Open BairesDev Dev Barometer, Q3 2026 edition (the page is the report)https://www.bairesdev.com/blog/dev-barometer-q3-2026-devs-answering-for-code/HTTP 200, text/html
    Figure in the document
    found in the document.
    Transformation
    none: the figure survives.
    Read on
    The BairesDev page this was read onhttps://www.bairesdev.com/blog/dev-barometer-q3-2026-devs-answering-for-code/retrieved 2026-10-06.
  • Q4-15Unverifiablecandidate Q068

    The claim, word for word

    Published by Genpact and quoted here. Not a Praxis result, and not a Praxis client figure.

    "Resolving enterprise debts represents a significant untapped opportunity, estimated at $18 trillion according to our research, unlocking significant gains in growth (around 8% faster annual revenue growth), cost efficiency (roughly 16% annual cost reduction)."

    Publisher
    Genpact (outsourcer)
    Figure
    $18 trillion, 8%, 16%.
    Attribution as printed
    "our research", with no document title, no sample, no method and no link.
    Primary document
    none located
    Figure in the document
    no document to check.
    Transformation
    a name used as a citation, with no document behind it.
    Note
    A large, well-resourced publisher, so it is named. The page states no method and links no document, and a $18 trillion figure is the largest single number in this edition.
    Read on
    The Genpact page this was read onhttps://www.genpact.com/insight/how-four-enterprise-debts-will-make-or-break-your-ai-futureretrieved 2026-10-06.
  • Q4-16Unverifiablecandidate Q069

    The claim, word for word

    Published by HatchWorks AI and quoted here. Not a Praxis result, and not a Praxis client figure.

    "With HatchWorks AI, we improved our velocity by almost 300% while reducing bugs to near zero."

    Publisher
    HatchWorks AI (vendor)
    Figure
    almost 300%.
    Attribution as printed
    none printed. The figure sits inside a quoted client statement with no measurement period, baseline or method.
    Primary document
    none located
    Figure in the document
    no document to check.
    Transformation
    no attribution printed at all.
    Note
    Stated, unverifiable. The publisher's own client figure, which no reader can check.
    Read on
    The HatchWorks AI page this was read onhttps://hatchworks.com/how-we-engage/retrieved 2026-10-06.
  • Q4-17Unverifiablecandidate Q049

    The claim, word for word

    Published by LOW/CODE Agency and quoted here. Not a Praxis result, and not a Praxis client figure.

    "40% per-project cost reduction."

    Publisher
    LOW/CODE Agency (vendor)
    Figure
    40%.
    Attribution as printed
    none printed. A result card on the firm's own service page, with no baseline, period or method.
    Primary document
    none located
    Figure in the document
    no document to check.
    Transformation
    no attribution printed at all.
    Note
    Stated, unverifiable.
    Read on
    The LOW/CODE Agency page this was read onhttps://www.lowcode.agency/services/ai-app-development/ai-consultingretrieved 2026-10-06.
  • Q4-18Unverifiablecandidate Q050

    The claim, word for word

    Published by Markovate and quoted here. Not a Praxis result, and not a Praxis client figure.

    "70% Faster BOM Extraction with AI Blueprint Classifier."

    Publisher
    Markovate (vendor)
    Figure
    70%.
    Attribution as printed
    none printed. A homepage result card, with no baseline, period or method.
    Primary document
    none located
    Figure in the document
    no document to check.
    Transformation
    no attribution printed at all.
    Note
    Stated, unverifiable.
    Read on
    The Markovate page this was read onhttps://markovate.com/retrieved 2026-10-06.
  • Q4-19Unverifiablecandidate Q012

    The claim, word for word

    Published by SeidrLab and quoted here. Not a Praxis result, and not a Praxis client figure.

    "AI tools that automate data collection, normalisation, and report production are generating measurable returns in the 50 to 80 percent reduction range for the assembly portion of reporting workflows."

    Publisher
    SeidrLab (consultancy)
    Figure
    50 to 80 percent.
    Attribution as printed
    none printed, and the page says so itself: "We're sharing the patterns we see consistently in client engagements, not statistical claims from a controlled sample".
    Primary document
    none located
    Figure in the document
    no document to check.
    Transformation
    no attribution printed at all.
    Note
    The only publisher in the frame that states on the page that its own figures are not statistical claims. The figure is still uncheckable, and the disclosure is the most honest handling of an uncheckable figure observed in this edition.
    Read on
    The SeidrLab page this was read onhttps://seidrlab.com/blog/ai-adoption-benchmark-report/retrieved 2026-10-06.
  • Q4-20Unverifiablecandidate Q041

    The claim, word for word

    Published by Velsof (Velocity Software Solutions) and quoted here. Not a Praxis result, and not a Praxis client figure.

    "Clinical documentation: AI-assisted note-taking and coding that reduces physician administrative burden by 30-40%."

    Publisher
    Velsof (Velocity Software Solutions) (vendor)
    Figure
    30-40%.
    Attribution as printed
    none printed.
    Primary document
    none located
    Figure in the document
    no document to check.
    Transformation
    no attribution printed at all.
    Read on
    The Velsof (Velocity Software Solutions) page this was read onhttps://www.velsof.com/blog/ai-consulting-mid-market-companies/retrieved 2026-10-06.
  • Q4-21Unverifiablecandidate Q042

    The claim, word for word

    Published by Velsof (Velocity Software Solutions) and quoted here. Not a Praxis result, and not a Praxis client figure.

    "AI-driven personalization that increases conversion rates by 1-2%, dynamic pricing that improves margins by 2-3%."

    Publisher
    Velsof (Velocity Software Solutions) (vendor)
    Figure
    1-2% and 2-3%.
    Attribution as printed
    none printed.
    Primary document
    none located
    Figure in the document
    no document to check.
    Transformation
    no attribution printed at all.
    Read on
    The Velsof (Velocity Software Solutions) page this was read onhttps://www.velsof.com/blog/ai-consulting-mid-market-companies/retrieved 2026-10-06.
  • Q4-22Unverifiablecandidate Q043

    The claim, word for word

    Published by Velsof (Velocity Software Solutions) and quoted here. Not a Praxis result, and not a Praxis client figure.

    "Most well-scoped AI projects achieve positive ROI within 6-12 months of deployment."

    Publisher
    Velsof (Velocity Software Solutions) (vendor)
    Figure
    6-12 months.
    Attribution as printed
    none printed.
    Primary document
    none located
    Figure in the document
    no document to check.
    Transformation
    no attribution printed at all.
    Note
    "Most" is a quantifier over a population the page never defines.
    Read on
    The Velsof (Velocity Software Solutions) page this was read onhttps://www.velsof.com/blog/ai-consulting-mid-market-companies/retrieved 2026-10-06.
  • Q4-23Unverifiablecandidate Q044

    The claim, word for word

    Published by Velsof (Velocity Software Solutions) and quoted here. Not a Praxis result, and not a Praxis client figure.

    "7 Proven Customer Onboarding Automation Patterns That Cut Costs 40%."

    Publisher
    Velsof (Velocity Software Solutions) (vendor)
    Figure
    40%.
    Attribution as printed
    none printed. An article title in the page's own related-reading list.
    Primary document
    none located
    Figure in the document
    no document to check.
    Transformation
    no attribution printed at all.
    Read on
    The Velsof (Velocity Software Solutions) page this was read onhttps://www.velsof.com/blog/ai-consulting-mid-market-companies/retrieved 2026-10-06.
  • Q4-24Unverifiablecandidate Q046

    The claim, word for word

    Published by Velsof (Velocity Software Solutions) and quoted here. Not a Praxis result, and not a Praxis client figure.

    "Banking Group, AI Strategy Saves $2M in Wasted Investment."

    Publisher
    Velsof (Velocity Software Solutions) (vendor)
    Figure
    $2M.
    Attribution as printed
    none printed. An anonymised client result on the firm's own service page.
    Primary document
    none located
    Figure in the document
    no document to check.
    Transformation
    no attribution printed at all.
    Note
    Stated, unverifiable. An anonymised dollar result cannot be checked by anyone.
    Read on
    The Velsof (Velocity Software Solutions) page this was read onhttps://www.velsof.com/ai-training-consulting/retrieved 2026-10-06.
  • Q4-25Unverifiablecandidate Q048

    The claim, word for word

    Published by Velsof (Velocity Software Solutions) and quoted here. Not a Praxis result, and not a Praxis client figure.

    "The top 3 recommendations, demand forecasting, automated reorder points, and customer churn prediction, were all implemented over the following year, delivering $4.5M in annual savings."

    Publisher
    Velsof (Velocity Software Solutions) (vendor)
    Figure
    $4.5M.
    Attribution as printed
    none printed. An anonymised client result on the firm's own service page.
    Primary document
    none located
    Figure in the document
    no document to check.
    Transformation
    no attribution printed at all.
    Note
    Stated, unverifiable.
    Read on
    The Velsof (Velocity Software Solutions) page this was read onhttps://www.velsof.com/ai-training-consulting/retrieved 2026-10-06.
  • Q4-26Unverifiablecandidate Q013

    The claim, word for word

    Published by Vention and quoted here. Not a Praxis result, and not a Praxis client figure.

    "The per-token cost of AI inference has dropped by 98%, yet enterprise AI bills are tripling."

    Publisher
    Vention (vendor)
    Figure
    98%.
    Attribution as printed
    none printed, and none anywhere in the surrounding text.
    Primary document
    none located
    Figure in the document
    no document to check.
    Transformation
    no attribution printed at all.
    Note
    The article's opening sentence and its largest percentage.
    Read on
    The Vention page this was read onhttps://ventionteams.com/blog/reduce-ai-token-costsretrieved 2026-10-06.
  • Q4-27Unverifiablecandidate Q014

    The claim, word for word

    Published by Vention and quoted here. Not a Praxis result, and not a Praxis client figure.

    "Without the right architecture and governance, monthly AI spend in large-scale deployments can reach as high as $500 million."

    Publisher
    Vention (vendor)
    Figure
    $500 million.
    Attribution as printed
    none printed.
    Primary document
    none located
    Figure in the document
    no document to check.
    Transformation
    no attribution printed at all.
    Read on
    The Vention page this was read onhttps://ventionteams.com/blog/reduce-ai-token-costsretrieved 2026-10-06.
  • Q4-28Unverifiablecandidate Q015

    The claim, word for word

    Published by Vention and quoted here. Not a Praxis result, and not a Praxis client figure.

    "Given the 20-60x price difference across model tiers, dynamic routing alone can significantly reduce AI costs without sacrificing output quality."

    Publisher
    Vention (vendor)
    Figure
    20-60x.
    Attribution as printed
    none printed.
    Primary document
    none located
    Figure in the document
    no document to check.
    Transformation
    no attribution printed at all.
    Read on
    The Vention page this was read onhttps://ventionteams.com/blog/reduce-ai-token-costsretrieved 2026-10-06.
  • Q4-29Unverifiablecandidate Q016

    The claim, word for word

    Published by Vention and quoted here. Not a Praxis result, and not a Praxis client figure.

    "Many providers offer discounts of up to 50% for asynchronous jobs processed during off-peak hours, reducing the cost of large-scale AI workloads."

    Publisher
    Vention (vendor)
    Figure
    up to 50%.
    Attribution as printed
    none printed, and no provider is named.
    Primary document
    none located
    Figure in the document
    no document to check.
    Transformation
    no attribution printed at all.
    Read on
    The Vention page this was read onhttps://ventionteams.com/blog/reduce-ai-token-costsretrieved 2026-10-06.
  • Q4-30Unverifiablecandidate Q018

    The claim, word for word

    Published by Vention and quoted here. Not a Praxis result, and not a Praxis client figure.

    "In one engagement, Vention's work to optimize AI-assisted development reduced the context sent to the model by up to 95.47%, cutting the number of input tokens required while maintaining 100% recall of critical anchors in provider-attested scenarios."

    Publisher
    Vention (vendor)
    Figure
    95.47% and 100%.
    Attribution as printed
    the publisher's own engagement, described as "provider-attested", with no document, baseline or measurement period.
    Primary document
    none located
    Figure in the document
    no document to check.
    Transformation
    no attribution printed at all.
    Note
    Stated, unverifiable. Two decimal places on an unchecked figure is the most precise number in this edition.
    Read on
    The Vention page this was read onhttps://ventionteams.com/blog/reduce-ai-token-costsretrieved 2026-10-06.
  • Q4-31Verifiedcandidate Q020

    The claim, word for word

    Published by Vention and quoted here. Not a Praxis result, and not a Praxis client figure.

    "56% of companies cite cost reductions, and 57% report revenue benefits from adopting AI in software engineering."

    Publisher
    Vention (vendor)
    Figure
    56% and 57%.
    Attribution as printed
    the publisher's own survey, with the method on the same page: "The Vention BIXA research methodology was based on surveying 480 qualified decision-makers from the US, UK, and DACH, actively investing in software development".
    Primary document
    Open Vention BIXA research, Q1 2026 (the page is the report)https://ventionteams.com/ai/sdlc/reportHTTP 200, text/html
    Figure in the document
    found in the document.
    What the document says
    The page states the sample of 480 decision-makers and the three regions, and carries a "[Source: Vention Bixa Research, Q1 2026]" marker on some charts.Those are the document’s own words. Not a Praxis result.
    Transformation
    none: the figure survives.
    Note
    Self-asserted. The page states a method and a sample for its own survey but does not mark which of its figures come from that survey and which are third-party, so the reader has to guess. Rows Q4-32 to Q4-37 are the figures on the same page that are not traceable to it.
    Read on
    The Vention page this was read onhttps://ventionteams.com/ai/sdlc/reportretrieved 2026-10-06.
  • Q4-32Unverifiablecandidate Q021

    The claim, word for word

    Published by Vention and quoted here. Not a Praxis result, and not a Praxis client figure.

    "68% of developers save more than 10 hours per week from using AI."

    Publisher
    Vention (vendor)
    Figure
    68%.
    Attribution as printed
    nothing on three of its four appearances. The fourth reads "According to Atlassian, 68% of developers reported significant time savings of more than 10 hours per week from using AI", naming Atlassian with no document title and no link.
    Primary document
    none located
    Figure in the document
    no document to check.
    Transformation
    a name used as a citation, with no document behind it.
    Note
    The same figure is stated four times on one page and attributed once. The attribution is in the fourth restatement, roughly 30,000 characters after the first, so a reader meeting it early has nothing to check.
    Read on
    The Vention page this was read onhttps://ventionteams.com/ai/sdlc/reportretrieved 2026-10-06.
  • Q4-33Unverifiablecandidate Q027

    The claim, word for word

    Published by Vention and quoted here. Not a Praxis result, and not a Praxis client figure.

    "This is a significant jump from last year, where 54% of developers said they had yet to experience significant productivity benefits by using AI."

    Publisher
    Vention (vendor)
    Figure
    54%.
    Attribution as printed
    none printed.
    Primary document
    none located
    Figure in the document
    no document to check.
    Transformation
    no attribution printed at all.
    Read on
    The Vention page this was read onhttps://ventionteams.com/ai/sdlc/reportretrieved 2026-10-06.
  • Q4-34Unverifiablecandidate Q028

    The claim, word for word

    Published by Vention and quoted here. Not a Praxis result, and not a Praxis client figure.

    "AI-enabled development teams achieve productivity gains of 15% to 20%, while Vention clients report efficiency improvements of 2x to 3x after integrating AI into their delivery processes."

    Publisher
    Vention (vendor)
    Figure
    15% to 20%, and 2x to 3x.
    Attribution as printed
    none printed for either figure.
    Primary document
    none located
    Figure in the document
    no document to check.
    Transformation
    no attribution printed at all.
    Note
    One sentence carries an industry figure and the publisher's own client figure, neither sourced.
    Read on
    The Vention page this was read onhttps://ventionteams.com/ai/sdlc/reportretrieved 2026-10-06.
  • Q4-35Unverifiablecandidate Q029

    The claim, word for word

    Published by Vention and quoted here. Not a Praxis result, and not a Praxis client figure.

    "63% additional delivery capacity in AI labor-equivalent terms."

    Publisher
    Vention (vendor)
    Figure
    63%.
    Attribution as printed
    none printed. A statistic card with no source marker.
    Primary document
    none located
    Figure in the document
    no document to check.
    Transformation
    no attribution printed at all.
    Read on
    The Vention page this was read onhttps://ventionteams.com/ai/sdlc/reportretrieved 2026-10-06.
  • Q4-36Unverifiablecandidate Q032

    The claim, word for word

    Published by Vention and quoted here. Not a Praxis result, and not a Praxis client figure.

    "Another study highlights that AI-enabled development teams ship 4 times faster, but also create 10 times as many security issues."

    Publisher
    Vention (vendor)
    Figure
    4 times and 10 times.
    Attribution as printed
    "another study". No publisher, no title, no date, no link.
    Primary document
    none located
    Figure in the document
    no document to check.
    Transformation
    a name used as a citation, with no document behind it.
    Note
    The purest form of the shape: a study invoked without even a name.
    Read on
    The Vention page this was read onhttps://ventionteams.com/ai/sdlc/reportretrieved 2026-10-06.
  • Q4-37Unverifiablecandidate Q036

    The claim, word for word

    Published by Vention and quoted here. Not a Praxis result, and not a Praxis client figure.

    "McKinsey reports that top-performing AI software organizations achieve productivity gains of 16% to 30% and software quality improvements of 31% to 45%, while lower performers see little measurable impact."

    Publisher
    Vention (vendor)
    Figure
    16% to 30%, and 31% to 45%.
    Attribution as printed
    McKinsey, named twice on the page, with no document title and no link either time.
    Primary document
    none located
    Figure in the document
    no document to check.
    Transformation
    a name used as a citation, with no document behind it.
    Read on
    The Vention page this was read onhttps://ventionteams.com/ai/sdlc/reportretrieved 2026-10-06.
  • Q4-38Unverifiablecandidate Q035

    The claim, word for word

    Published by Vention and quoted here. Not a Praxis result, and not a Praxis client figure.

    "AI coding assistants can improve individual developer productivity by 10% to 15%, but those gains often disappear if teams don't redirect the saved time toward higher-value work."

    Publisher
    Vention (vendor)
    Figure
    10% to 15%.
    Attribution as printed
    none printed.
    Primary document
    none located
    Figure in the document
    no document to check.
    Transformation
    no attribution printed at all.
    Read on
    The Vention page this was read onhttps://ventionteams.com/ai/sdlc/reportretrieved 2026-10-06.
  • Q4-39Unverifiablecandidate Q037

    The claim, word for word

    Published by Vention and quoted here. Not a Praxis result, and not a Praxis client figure.

    "73% of engineers report faster code delivery."

    Publisher
    Vention (vendor)
    Figure
    73%.
    Attribution as printed
    none printed.
    Primary document
    none located
    Figure in the document
    no document to check.
    Transformation
    no attribution printed at all.
    Read on
    The Vention page this was read onhttps://ventionteams.com/ai/sdlc/reportretrieved 2026-10-06.
  • Q4-40Unverifiablecandidate Q076

    The claim, word for word

    Published by LeewayHertz and quoted here. Not a Praxis result, and not a Praxis client figure.

    "One study discovered that software developers who used Microsoft's GitHub Copilot completed tasks 56 percent faster than those who did not use the tool."

    Publisher
    LeewayHertz (vendor)
    Figure
    56 percent.
    Attribution as printed
    "one study". No publisher, no title, no date, no link.
    Primary document
    none located
    Figure in the document
    no document to check.
    Transformation
    a name used as a citation, with no document behind it.
    Read on
    The LeewayHertz page this was read onhttps://www.leewayhertz.com/current-state-of-generative-ai/retrieved 2026-10-06.
  • Q4-41Unverifiablecandidate Q079

    The claim, word for word

    Published by LeewayHertz and quoted here. Not a Praxis result, and not a Praxis client figure.

    "LLM App for Wine Recommendation, achieved a 60% increase in accuracy for personalized wine recommendations."

    Publisher
    LeewayHertz (vendor)
    Figure
    60%.
    Attribution as printed
    none printed. A result card with no baseline, period or method.
    Primary document
    none located
    Figure in the document
    no document to check.
    Transformation
    no attribution printed at all.
    Note
    Stated, unverifiable.
    Read on
    The LeewayHertz page this was read onhttps://www.leewayhertz.com/ai-consulting-services-company/retrieved 2026-10-06.
  • Q4-42Unverifiablecandidate Q080

    The claim, word for word

    Published by LeewayHertz and quoted here. Not a Praxis result, and not a Praxis client figure.

    "AI-powered Medical Assistant, 2x faster decision-making."

    Publisher
    LeewayHertz (vendor)
    Figure
    2x.
    Attribution as printed
    none printed. A result card with no baseline, period or method.
    Primary document
    none located
    Figure in the document
    no document to check.
    Transformation
    no attribution printed at all.
    Note
    Stated, unverifiable.
    Read on
    The LeewayHertz page this was read onhttps://www.leewayhertz.com/ai-consulting-services-company/retrieved 2026-10-06.
  • Q4-43Unverifiablecandidate Q081

    The claim, word for word

    Published by LeewayHertz and quoted here. Not a Praxis result, and not a Praxis client figure.

    "LLM-powered App for Compliance and Security Access, 60% faster information access."

    Publisher
    LeewayHertz (vendor)
    Figure
    60%.
    Attribution as printed
    none printed. A result card with no baseline, period or method.
    Primary document
    none located
    Figure in the document
    no document to check.
    Transformation
    no attribution printed at all.
    Note
    Stated, unverifiable.
    Read on
    The LeewayHertz page this was read onhttps://www.leewayhertz.com/ai-consulting-services-company/retrieved 2026-10-06.

Appendix

What was dropped, and what could not be read

Reading 81 pages produced 81 candidate sentences. 43 reached the table. Everything else is here, so the difference between the two numbers is visible rather than implied. Candidates are listed by figure rather than by sentence, because reprinting other firms’ copy at length is not the point of the page.

Out of scope (31 candidates)

Candidates that failed a scope condition, with the reason
CandidatePublisherFigureConditionWhy it was dropped
Q003Praxis Consulting Company.16%.condition 4.labelled a forecast on the page.
Q006Praxis Consulting Company.30%, 80%.condition 4.the page itself identifies the figure as a forecast about 2029 and says so.
Q010Praxis Consulting Company.11%, 13%, 18 months.condition 4.labelled an expectation on the page.
Q011SeidrLab.$200K, $500K.condition 2.a buyer's consulting budget band, not an effect of AI.
Q017Vention.80%.condition 2.a utilisation threshold stated as a condition, not an outcome.
Q019Vention.80%.condition 2.a utilisation threshold stated as a condition, not an outcome.
Q022Vention.45%.condition 2.developers' trust in AI output, not an effect on cost, time, headcount or output.
Q023Vention.45%.condition 2.developers' trust in AI output, not an effect on cost, time, headcount or output.
Q030Vention.38%, 53%.condition 2.buyers' perception of cost as a barrier, not an effect of AI on cost.
Q031Vention.24%.condition 4.an expectation measured in an experiment, labelled as an expectation.
Q033Vention.70%.condition 4.a forecast about 2027, labelled "expected".
Q040Velsof (Velocity Software Solutions).$1B, $50M, 2,000.condition 2.a definition of the mid-market by employees and revenue.
Q045Velsof (Velocity Software Solutions).18 month.condition 2.the length of a roadmap the firm delivers, which is a deliverable, not an effect.
Q047Velsof (Velocity Software Solutions).$3M, 99%.condition 2.a vendor's accuracy promise narrated inside a story the page then questions.
Q051Phos AI Labs.$100M, $10M.condition 2.the revenue band of the audience the page addresses.
Q052Phos AI Labs.$2M, $50M.condition 2.the revenue band of the audience the page addresses.
Q053Phos AI Labs.90 days.condition 2.an instruction to ask a firm for its own figures, not a figure.
Q054Phos AI Labs.$250K.condition 2.the cost of hiring a person, not an effect of AI.
Q055Phos AI Labs.$5M.condition 2.the revenue band of the audience the page addresses.
Q058Auxis.25%.condition 4.a forecast about 2030, labelled "expected".
Q065Boldr AI.$10K, $150K.condition 2.a price for a service, which is an offer rather than a statistic.
Q066Boldr AI.$10M, $125M.condition 2.the revenue band of the audience the page addresses.
Q067DataArt.20%.condition 2.how a contract apportions spend, not an outcome.
Q070LeewayHertz.67.5%.condition 2.a language-model benchmark accuracy score, not a business cost, time, headcount or output effect.
Q071LeewayHertz.50 percent.condition 4.labelled a projection by the publisher: "We project that ... could".
Q072LeewayHertz.45 percent.condition 4.labelled an estimate: "an estimation report by McKinsey ... could", and the document is linked.
Q073LeewayHertz.15 percent.condition 4.labelled an estimate: "McKinsey's estimations indicate ... could", and the document is linked.
Q074LeewayHertz.5 percent.condition 4.labelled an estimate: "McKinsey's analysis proposes ... could", and the document is linked.
Q075LeewayHertz.45 percent.condition 4.labelled a possibility: "could be anywhere from", and the document is linked.
Q077LeewayHertz.15 percent.condition 4.labelled a possibility: "studies suggest ... could", and the document is linked.
Q078LeewayHertz.$200 billion, $340 billion, 4.7 percent.condition 4.labelled a possibility: "poised to ... potentially", and the document is linked.

Restatements of a figure already in the table (7 candidates)

The unit is one claim, so a figure restated on the same page is counted once. Each of these points at the row that carries it.

  • Q024Ventioncounted in Q4-31
  • Q025Ventioncounted in Q4-31
  • Q026Ventioncounted in Q4-32
  • Q034Ventioncounted in Q4-37
  • Q038Ventioncounted in Q4-31
  • Q039Ventioncounted in Q4-32
  • Q062BairesDevcounted in Q4-10

Pages and documents that could not be read (8 pages)

Each was abandoned on its first refusal. No retry, no change of header or address, nothing solved, no paid tool. Nothing in the table rests on any of them, and listing them is what makes the rest of the page checkable.

Pages in the frame that were not served, with the status that stopped each
PublisherPageWhat the server returnedObserved
Quantum RiseThe Quantum Rise page at quantumrise.comhttps://quantumrise.com/HTTP 200, 0 chars of served text (JavaScript-only shell).2026-10-06.
Quantum RiseThe Quantum Rise page at quantumrise.comhttps://quantumrise.com/insightsHTTP 200, 0 chars of served text (JavaScript-only shell).2026-10-06.
Quantum RiseThe Quantum Rise page at quantumrise.comhttps://quantumrise.com/what-we-doHTTP 200, 0 chars of served text (JavaScript-only shell).2026-10-06.
Opinosis AnalyticsThe Opinosis Analytics page at opinosis-analytics.comhttps://www.opinosis-analytics.com/HTTP 202.2026-10-06.
BCG XThe BCG X page at bcg.comhttps://www.bcg.com/xHTTP 403.2026-10-06.
Brainpool AIThe Brainpool AI page at brainpool.aihttps://brainpool.ai/blogHTTP 403.2026-10-06.
HatchWorks AIThe HatchWorks AI page at hatchworks.comhttps://hatchworks.com/state-of-ai-2026-mid-year-reality-check/HTTP 200, 126 chars of served text (JavaScript-only shell).2026-10-06.
HatchWorks AIThe HatchWorks AI page at hatchworks.comhttps://hatchworks.com/HTTP 200, 163 chars of served text (JavaScript-only shell).2026-10-06.

One document refused the request too. The report six of the audited claims link to answered a single plain request by closing the connection, so those six could not be read here and none of them is reported as an absence. A reader with a browser can open it, which is the point of the publisher having printed the link.

Limits

What this page will not say

Written down so the refusal is checkable against the page, and so the next edition inherits it.

  • No ranking, no scorecard and no league table. The page reports claims, not firms, and reaching a verdict about a firm from its rows is not something the method supports.
  • No claim about a rate of fabrication in any AI retrieval system. R.01 refused this and the refusal carries forward.
  • Nothing about any firm's revenue, headcount, quality, competence or client outcomes.
  • No generalisation past the sample. Forty-three claims from one day's reading of 81 pages is a small sample with a published method, and the numbers here describe those claims and nothing else.
  • No inference where a document could not be read. A page that refused the request produces UNVERIFIABLE, never NOT FOUND.

Cadence

The next edition

Q1 2027, retrieval window 2027-03-01 to 2027-03-10, published by 2027-04-15

A named index with a stated method published on a date is worth citing because it is a stable thing to point at. A missed quarter removes the only property that makes it so. The frame, the verdict words and the transformation shapes carry forward unchanged, so edition two is comparable to this one; a new shape gets named rather than folded into an old one, and the change is written in the record below.

Every row about a named third party is re-observed before each revision and the new date is written on the page. A stale fact about a named firm is worse than no fact. A publisher that believes a row misreads its page can write to the address on the contact page; the row is re-observed, not negotiated.

  1. 2026-10-06Later the same day: the publisher's own R.01 ledger was corrected in response to row Q4-01 (entry A1 relabelled Not found, its claim struck, a dated notice above its tally). Row Q4-03's note now states, as Q4-05's already did, that R.01's 2026-09-01 liveness control was not re-run here. The disclosure's count of Praxis rows with no document corrected from four to three; Q4-01 names a document that does not contain the figure, which is a different state. No verdict changed.
  2. 2026-10-06First edition, Q4 2026. Frame derived by script from the published cells of the eight-criteria comparison, 23 firms and 89 pages. 81 pages read, 8 recorded unread with their statuses, 81 candidate claims, 43 in the table. The organic-top-20 half of the approved frame could not be captured at zero cost without a browser; recorded, not worked around. New transformation shape named this edition: no attribution printed at all.

Cite this edition

Praxis Consulting Company, "The Mid-Market AI Claim Audit, Q4 2026", retrieved 2026-10-06, https://praxisconsultingco.com/research/mid-market-ai-claim-audit

Cite a single claim by its row id, the publisher and the retrieval date, and link the page named beside it rather than this page alone. Because the rows carry no order, one row can be quoted without implying a verdict about any firm.

https://praxisconsultingco.com/research/mid-market-ai-claim-audit

Bring a number to the first conversation

If a firm has quoted you a figure for what AI will save you, ask it for the document the figure is in and the page it is on. Praxis answers that question about its own numbers, including the one on this page that failed, and the first conversation is free.

Request a free quoteEmail usCall (415) 634-5434

No obligation · a scoping conversation first