The Claims Register

← Back

Updates, corrections and new evidence for the claims in my own work.

The argument in the book is that a claim is only worth as much as the record behind it. Holding sellers to that standard while exempting myself would be indefensible, so this is where I keep my own record.

How this works ▾

I wrote the book and I write everything on this site, which means I am also the person reviewing it. That is not an ideal arrangement, and I will not pretend otherwise. The best answer available to me is to publish the review rather than ask you to take my word for it.

In practice it works as follows. When a source is updated or superseded, the post follows it, and this page says what changed in the source. Where I find stronger evidence for an argument I have already made, I add the source. If events overtake a post, I rewrite it rather than leave it standing.

Each entry states which of those applied.

  • Holds. I returned to the source and the claim still stands.
  • Corrected. A claim widely repeated in the field turns out to trace somewhere else. The entry names the source that settles it.
  • Strengthened. The argument was sound but the supporting evidence was thin, so I found better and cited it.
  • Revised. Circumstances changed and the post changed with them. The URL is unchanged.

Where I could not trace a claim all the way back to a primary source, the entry says corroborated rather than verified. I would rather state the difference than obscure it.

The posts: the book, extended

I re-read published posts and ask whether they remain true, whether they still carry weight, and whether the sourcing holds. Anything that no longer holds is brought up to date in the post.

  • ▾

    This post reads the Commission’s procurement proposal and finds that quality is weighted at award, while nothing compares the delivered result with what those points were given for. Since it went up, practitioners reading the proposal have pointed to how much it leaves to each contracting authority: buyers would choose their own quality criteria and could negotiate within the main procedures without first having to justify it.

    I added a section to this post on why that discretion makes the buyer’s own written result matter more, along with a graphic that sets award, delivery and completion against that record. No figure and no conclusion changed. The proposal is open for comment until November 11.

  • ▾

    Harvard Business School and the Burning Glass Institute studied 11,300 roles at large US employers that had removed a degree requirement from their job postings. In those roles the share of new hires without a bachelor’s degree rose about 3.5 percentage points. Nearly all of that change came from 37 percent of the firms. Around 45 percent changed the posting and made no meaningful difference to who they actually hired, and about a fifth made short-term gains that did not stick and in the long run hired a smaller share of people without degrees.

    I added the study to this post because it shows, across thousands of roles, what a stated standard does when nothing checks it afterward. The report is Skills-Based Hiring: The Long Road from Pronouncements to Practice by Matt Sigelman, Joseph Fuller and Alex Martin, February 2024.

  • ▾

    New Orleans now has a public record of what this post describes. On October 15 the city’s inspector general reported that its fleet division has been paying about $1.4 million a year since 2025 for vehicle tracking software, covering 2,481 licenses, and that only 62 percent of the city’s 2,577 vehicles were connected to it.

    City employees and the police unions had raised privacy concerns. The city’s vehicle policy was silent on tracking software, so the division had no authority to require it. The inspector general estimated that about $498,000 went to unused licenses and recommended the policy be amended by April 30, 2027. I added the case to the post because the purchase depended on people accepting a change, and nothing in the city’s policy required that change. The source is the inspector general’s October 15, 2026 report on fleet operations in the Equipment Maintenance Division.

  • ▾

    Wharton Human-AI Research and GBK Collective published a study in October that finds the divide at the center of this post. About 1,400 people at US companies with at least 1,000 employees answered it in July. Forty percent of the managers and executives said they were very confident their company could track its total AI spend once costs bundled into other software and cloud services were counted. Of that group, 56 percent reported strongly positive returns, against 10 percent of those less confident.

    The authors note that the confident group is also more enthusiastic about AI in general, so some of the gap may be “conviction rather than capability.” I put the study in the post next to the KPMG figures because the two draw the line in the same place. The report is The Adaptive Advantage: Who’s Winning with AI and How, October 2026.

  • ▾

    In February the FinOps Foundation changed its mission from “Advancing the People who manage the value of Cloud” to “Advancing the People who manage the Value of Technology.” This post uses that change as the plainest sign that the market now wants value looked after once a system is live, so I put the Foundation’s own words into it, along with its survey numbers.

    The same release says 98 percent of the 1,192 people who answered its 2026 survey now manage AI spend, up from 31 percent two years earlier. The post’s argument holds. Tools that read cost and usage month by month still can’t show whether value is building from one year to the next. The source is the Foundation’s February 19, 2026 release.

  • ▾

    Since March 2024 the IASB has been working on a draft amendment to IFRS 3 that would make a company report, each year after an acquisition, how far the targets it set on the day of the deal are being met. This post follows what has happened to it. I went back to the Board’s own record this month. It last discussed the project on September 23, 2026, and its project page says it “was not asked to make any decisions.” The draft is still a draft, as the post says.

    The vote the post leans on holds as well. On May 20, 2026, seven of thirteen members agreed the revised package was worth its cost, the narrowest vote in that month’s Update. Every document the post relies on is now linked from it, from the 2024 exposure draft to the European Securities and Markets Authority’s July 2024 letter.

  • ▾

    Finance executives told The CFO Survey that AI raised their labor productivity by 1.8 percent on average in 2025. Their own answers about revenue and headcount put it at 0.6 percent. The survey is run by Duke University with the Federal Reserve Banks of Richmond and Atlanta, and nearly 750 executives answered it between November 2025 and January 2026. I added it to this post because it shows the gap the post is about inside one set of respondents, without comparing two surveys.

    The economists who studied the answers think most of the difference is timing, with revenue and quality gains still to show up, and both numbers are self-reported. The post says that too. Even allowing for it, the gain they reported is three times what their own figures support. The paper is NBER Working Paper 34984 by Baslandze and colleagues, March 2026.

  • ▾

    When GAO reviewed category management at Veterans Affairs in September 2025, it made six recommendations. The one this post turns on asks VA to find out whether item-level data on prices paid, utilization and cost avoidance is even available.

    VA agreed with all six. I checked GAO’s page this month to see where they stand, and all six are still listed as open, the data recommendation included. The post now says so.

  • ▾

    Pay by the token and the seller decides how many tokens the work takes. This post’s point is that almost every number a buyer uses to judge that comes from the seller. I added the federal government’s version of the same thing.

    On 31 March 2026 the Office of Management and Budget issued memorandum M-26-10. It tells agencies to ask their current IT vendors for “information about the agency’s utilization rates and prices paid,” and to have new contracts require IT vendors “to disclose utilization and pricing information to the Government.” That’s a sensible way to see what agencies are using and paying for. It also means the government’s count of its own consumption starts with the seller’s meter, and the memo never says what utilization means.

  • ▾

    The post sets a test for any capability claim: show the outcome landing more than once, without the two or three people who make it work today. I added a section on what happens when somebody actually runs a test like that.

    SkillsBench, an academic benchmark, ran 87 tasks with and without curated skills for AI agents across 18 setups, and a fixed check decided whether each task was done. The average pass rate rose from 33.9 percent to 50.5 percent, but the gain ran anywhere from 4.1 to 25.7 points depending on the setup, so the number that matters is the one from your own setup. A separate study of 49,943 published agent skills found that 80.0 percent behave differently from what they declare, mostly through oversight rather than intent. And Malaysia’s National Audit Department found that the navy’s first Littoral Combat Ship “will be handed over without full operational capability,” as Bernama reported on 5 October 2026, and recommended that any conditional acceptance spell out what the contractor still owes.

  • ▾

    One vendor objection this post answers is “We can’t commit to outcomes we don’t control.” School districts that buy education technology on outcomes-based contracts already have a working answer, so I added their example. The district writes its own obligations into the same contract the vendor signs.

    Digital Promise studied the first cohort of these districts and published its findings on 16 March 2026. In its words, “Districts are accountable for implementing the tool at the intended dosage,” meaning how much and how often students use it. Usage ran ten times higher than under traditional contracts, and districts met the dosage requirement for 50 to 95 percent of participating students. These are early results from one cohort, and the post says so.

  • ▾

    This post argues that GSA’s draft AI clause, GSAR 552.239-7001, hands the buyer real power to test a model, then ties the penalty to how the model behaves rather than to whether the agency got the result it paid for. I wrote it from the June draft. GSA issued the final clause on 28 September 2026, effective 19 October, so I read the final text against the argument.

    A good deal changed. The “unbiased AI principles” are mostly gone, replaced by a duty to use reasonable efforts to make the model accurate and objective. The Government keeps its right to run automated tests of the model at any time with its own benchmarks, and decommissioning costs after a termination for cause are now capped at 25 percent of the affected order. The word “outcome” still doesn’t appear anywhere in the clause, so the penalty still turns on the model’s behavior and the argument holds.

    The post covers the final clause and links GSA’s deviation memo and FedScoop’s report, so anyone can check it.

  • ▾

    This post follows the industry pushback on GSA’s draft AI clause, GSAR 552.239-7001. My argument was that the fight that matters is over order of precedence, meaning which paper wins when the buyer’s clause and the seller’s commercial terms disagree. GSA has now issued the final clause, and the post covers it.

    GSA issued it on 28 September 2026 as a deviation, effective 19 October. It narrowed where the clause applies and replaced most of the “unbiased AI principles” language with a duty to use reasonable efforts, so industry got a good part of what it asked for. The precedence rule didn’t move. Paragraph (c) of the final text still says that where the clause conflicts with the contractor’s commercial agreements, “this clause takes precedence.” Crowell & Moring read it the same way in their 2 October alert.

    The post links GSA’s deviation memo and the Crowell alert, so anyone can read paragraph (c) for themselves.

  • ▾

    This post argues that most commitments don’t fail outright. They drift, one reasonable review at a time, and nothing in the usual governance sets off an alarm. I’ve added a real program that shows that shape, from an Australian audit published in July.

    Australia’s Department of Foreign Affairs and Trade received A$394.3 million for a program to strengthen security at its embassies. Its plan for measuring the benefits was approved in October 2019, eleven months after the government agreed the policy. In February 2021 one target was cut from 90 alarms installed to 50. By August 2022 the program had gone from 50 metrics to 74, partly because new projects were added. The auditors put it plainly: “There was not a clear baseline established to measure program benefits against: metrics and targets changed throughout the life of the program.”

    The department closed the program in December 2023, calling it an overall success, and the audit office’s own conclusion is that the benefits were “partly realised.” The report doesn’t suggest anyone acted in bad faith, and neither does the post. It links the report, so anyone can read paragraphs 1.19 to 1.21 for themselves.

  • ▾

    This post is about the performance notices British public buyers now have to publish on their suppliers, and how much of what goes into them the buyer gets to choose. Britain already keeps a public record of what departments choose to measure when the choice is theirs, and the post now uses it.

    The Cabinet Office publishes the indicators for central government’s most important contracts let under the 2015 regulations, along with each supplier’s rating. The file for April to June 2026, released at the end of September, lists 4,689 indicators across 1,473 contracts in 17 departments. Of the 4,455 indicators given one of the four performance ratings, 4,051 are Good, a little under 91 percent. Not one of the 4,689 descriptions mentions a business case.

    The government is plain about what the file is. Its notes call the data a snapshot that “does not represent a comprehensive assessment of the performance of the service, the contract or the supplier.” The post links the collection, so anyone can open the spreadsheet and count for themselves.

  • ▾

    I have two posts about a term that spread through finance and procurement conversations before anybody agreed what it measured. One of them is about the Linux Foundation’s Tokenomics Foundation, which launched on 4 August 2026 with thirty members and a charter to set standards for the economics of AI. Both were written while that body had the charter and no definition.

    The definition exists now. The Foundation’s own documentation carries it as a draft, version 0.5.2, dated 1 September 2026. AI Tokenomics is the discipline of converting energy and capital into AI capabilities and efficiently consuming that intelligence across the organization, to realize measurable business value.

    The closing clause is the part a buyer should read. The group setting standards for how AI spend gets counted has now written down that the counting is for business value. That is the right ambition, and it is more than the category had in August.

    It still leaves the harder half unwritten. The definition was drafted where the meters are, by members who sell the capacity being metered. What the money was for is recorded inside the buying organization, and no member’s telemetry reaches that record. Naming business value as the destination does not settle who says what the value was.

    Both posts now name the definition, its version and its date. The reading in them did not change. It rests on the Foundation’s current words instead of on its launch release. The number says 0.5.2, so I expect to read it again.

  • ▾

    Two posts on this site carry the new application dates for the EU’s high-risk AI rules. I had those dates from a Council press release and from law firm notes written about the regulation. I have now read the regulation.

    It is Regulation (EU) 2026/1744, agreed 8 July 2026 and published in the Official Journal on 24 July 2026, the Digital Omnibus on AI. Its own text sets the application of Sections 1, 2 and 3 of Chapter III to 2 December 2027, and to 2 August 2028 for AI systems classified as high-risk under Article 6(1) and Annex I.

    Both dates match what the posts already say. 2 December 2027 for the stand-alone high-risk systems under Annex III, which is hiring, credit, education and decisions of that kind. 2 August 2028 for AI built into regulated products. Nothing on either page needed changing.

    What changed is where the claim rests. A press release describes a law. The Official Journal carries it, and that is the address a reader should be able to go to. Both dates now trace to the instrument that set them.

  • ▾

    The sentence people reach for when metrics go wrong is this one. When a measure becomes a target, it ceases to be a good measure. It gets called Goodhart’s law. The wording is not his.

    It belongs to Marilyn Strathern, who wrote it in 1997 in a paper about audit in British universities, “Improving ratings”: audit in the British University system, European Review, volume 5, issue 3, pages 305 to 321. She was compressing Goodhart, and the compression is the version that traveled.

    What Charles Goodhart published in 1975 is longer and more exact. Any observed statistical regularity will tend to collapse once pressure is placed upon it for control purposes. It sits in Problems of Monetary Management: The UK Experience, in the Reserve Bank of Australia’s Papers in Monetary Economics.

    My post on value drift now credits each of them with what they wrote, and Goodhart’s own sentence is the one carrying the argument, because it names the mechanism instead of the symptom. Substitution happens when a regularity collapses under the weight of being managed. An organization counts customer touches because touches can be counted this week, and the thing it meant to measure quietly stops being measured at all.

    Strathern’s paper is verified at its publisher of record for author, title, journal, volume, issue, year and pages, and it is linked above. That the two sentences read exactly as written here is corroborated rather than verified. Her article sits behind the publisher’s paywall and Goodhart’s 1975 paper is in a conference volume I have not held.

  • ▾
    Every number in this post comes off a page that keeps moving. Artificial Analysis re-runs its index as new models arrive, so a comparison taken from it is true on the day you read it. The figures in the post now say which day that was, 25 September 2026, in the text and in the description of the chart.

    One of them has moved since. The median output tokens for comparable models was 88 million then. Today it is 82 million. The model still generated 260 million tokens running the same evaluations, so the gap I described as roughly three times the output is now a little more than three times rather than a little less.

    The rest holds. The score of 58 against a median of 26 for comparable models, the measured $5.98 to run one task on the index, the four dollars and twenty dollars per million tokens, and the page’s own line calling the model expensive next to others at its price.

    None of this weakens the post. A buyer pricing the purchase off the announcement got a list price that was accurate for about a day, and no figure at all for how many tokens the work would take. The second number is the one that moved again in the week after I published, and the only reason anyone can watch it move is that a third party keeps measuring it.

  • ▾
    Both halves of this post are statutes, and the European half was carried on the page by two law firm alerts, which is how most people meet a new regulation. I went to the Regulation itself this time.

    It is Regulation (EU) 2026/1744, the Digital Omnibus on AI, adopted 8 July 2026 and published in the Official Journal on 24 July. My reading of it holds. The date of application for Chapter III, Sections 1, 2 and 3 moves to 2 December 2027 for high-risk systems caught by Article 6(2) and Annex III, and to 2 August 2028 for those caught by Article 6(1) and Annex I. The deployer obligations, impact assessment included, sit inside Section 3, so they went with it. Article 6(5) is the single carve-out left where it was.

    Article 50 stayed put, and the Regulation is more specific than my post was about why December 2026 matters. It grants a four-month transitional period to providers who had already put generative systems on the market before 2 August 2026, so those providers have until early December to meet the marking duty. The post now says that plainly rather than leaving a reader to work back from the date.

    Colorado was easier to check, because the legislature publishes the bill. SB26-189, Automated Decision-Making Technology, signed 14 May 2026. The state’s own summary of it lists what a developer owes a deployer starting 1 January 2027: intended uses, categories of training data, known limitations, and instructions for appropriate use and human review. Every item on that list is something a developer has to say, and not one of them requires anybody to examine what the technology does.

    The sources line now carries the Regulation’s address at EUR-Lex and the bill’s address at the Colorado General Assembly, next to the two alerts. The reading did not change. It sits on the statutes now, so anyone can check it against them.

  • ▾
    This post argues that the federal fixed-price order settles what the government will pay and leaves the proof of what it got unassigned. That argument now rests on the government’s own words instead of on mine. Federal Acquisition Regulation 2.101 carries the definition: “Best value means the expected outcome of an acquisition that, in the Government’s estimation, provides the greatest overall benefit in response to the requirement.” Expected, and in the Government’s estimation. Both of those describe a forecast made before the work starts.

    So I went looking for where that forecast gets revisited. Part 42 covers Contract Administration and Audit Services, and Part 46 covers Quality Assurance, and those are the two parts of the regulation that govern a contract after it has been awarded. I read both and counted. Best value appears zero times in either one. So do outcome and estimation.

    I had been making this point from the shape of the order. The post now makes it from the regulation’s own definition and from two counts anyone can repeat in about five minutes.

  • ▾
    The whole post turns on two federal clauses, so I went back and read both of them through rather than trusting what I had already written about them.

    Federal Acquisition Regulation 52.215-2 defines records about as widely as the word can be defined, reaches the proposal and the discussions on it and the pricing and the material behind any cost, funding or performance report the contractor had to furnish, and stays open for three years after final payment. The word outcome does not appear in the clause once. Federal Acquisition Regulation 52.246-4 lets the government inspect and test all the services the contract calls for, require them performed again at no increase in contract amount, and where that will not fix the defect, reduce the contract price to reflect the reduced value of the services performed.

    Everything the post says about either clause is in the text of the clause. The references carry the full addresses now, so anyone can open both and read them the way I did.

  • ▾
    This post says the fixed price order asks agencies for a count of exceptions and nothing about whether the purchase returned anything. The hardest test of that is the regulation written under the order, so I obtained the proposed text of Federal Acquisition Regulation 16.104 as published in the Federal Register and read all ten pages of it.

    The claim holds. The proposed section requires an agency head to approve a written justification before a covered contract or order is used, and it says exactly what that justification has to contain: the information in section 16.103, which is titled Documenting contract type, plus any determination and findings the contract type already requires. Nothing in the ten pages asks an agency to measure, verify or report what was delivered. The only two reporting duties in the document are an establishment notice for a federally funded research center and an invoicing schedule, and neither is about a result.

    I checked the post’s paraphrase of the order’s policy against the order as well. “Fixed-price contracts with performance-based considerations” is the order’s own phrase, word for word.

    The reference now carries the proposed rule as published, so anyone can run the same test on the primary.

  • ▾
    The thirty percent weighting in this post came from Recital 42 of the European Commission’s September proposal. A recital explains a regulation and binds nobody, so I went looking for the same point in the operative articles.

    It is there, in Article 98(2). Quality criteria, the third subparagraph says, “shall refer to any criteria used to assess the degree to which a tender proposes beneficial, efficient or sustainable outcomes in relation to the subject-matter of the contract.” The operative word is proposes. What the thirty percent weighs is a judgment about what a tender offers, and the binding article says so in its own words.

    The post now names the article and quotes that subparagraph. Anyone who wants to test the reading can open page 104 of COM(2026) 590 and do it without me.

  • ▾
    This post rests on three sets of rules that govern the numbers a company reports about itself: revenue recognition under ASC 606 and IFRS 15, officer certification under Sections 302 and 906 of Sarbanes-Oxley, and the statement of cash flows under ASC 230 and IAS 7. I had cited all of them by number alone.

    Five now carry addresses, and I opened each one before I put it on the page. IFRS 15 and IAS 7 at ifrs.org, the Sarbanes-Oxley Act of 2002 at govinfo.gov, the PCAOB auditing standards, and the GAO Government Auditing Standards, where I confirmed the 2024 revision is still the current edition.

    Two stay as names, deliberately. FASB’s Codification sits behind a login, so ASC 606 and ASC 230 are cited by number. A link you cannot open is worse than a name you can search.

    I also checked the post’s central claim again, against the PCAOB standards, the AICPA and IFRS literature, and the GAO standards. No rule there obliges an organization to verify whether the benefits claimed in a business case were delivered. The claim holds.

    Nothing the post asserts has changed. Now you can check it without hunting for the sources.

  • ▾
    Every number in this post rests on one UK procurement policy note. I opened the government’s published text again and read it figure by figure against what the post says.

    All of it reproduces. Contracts worth a million pounds or more and under five million carry a minimum ten percent weighting, and contracts at five million and above carry a minimum twenty percent. The million pound threshold counts VAT. The weighting applies to the non price criteria rather than to the whole evaluation, which is the part most summaries leave out, and it applies to procurements commenced on or after 1 January 2027. The note also says that evidence of poor performance against social value KPIs can be taken into account in considering whether there are grounds to exclude suppliers from bidding for future tenders.

    The reference now points straight at the policy note, so anyone reading the post can open the same document and check the numbers for themselves.

  • ▾
    This post follows one number through three publications, so the argument only works if those three pages still say what I said they say. I opened all three again.

    They hold. The 8 June 2026 release still reports that 84 percent of finance organizations have implemented or are planning to implement AI and that only 7 percent report a high or very high impact, from a survey of 183 CFOs taken in June 2025. The page that restates it still carries the sentence I quoted, that 84 percent of CFOs say they have not yet realized a return on their AI investments in finance, and it is still hyperlinked to the release that says nothing of the kind. The third piece sits at the same address under a new headline.

    One source got added, and it is the same organization talking. Its release of 28 May 2026 tells finance chiefs they must not mistake activity for impact, and says that counts of pilots, tools rolled out or use cases in production show finance is moving but do not prove AI is delivering the value boards now expect. That is the distinction this post draws, published eleven days before the survey was turned into a statement about returns. It is in the post now, quoted and sourced.

    All three references are links rather than plain text, so the check the post asks you to run takes a couple of minutes and no copying and pasting.

  • ▾
    This post is about consulting reports that cited sources which did not say what the reports claimed. The same test belongs on my own numbers, so each of the three figures the argument rests on now points at a document that states it.

    The KPMG number does the most work. It points at GPTZero’s investigation, which says that of the 45 citations in the report, only five accurately point to real sources. The Deloitte contract value and the refund each point at a published report that carries the figure: about AU$440,000 for the work, and about AU$97,000 returned, roughly US$63,000.

    One report I decided not to use says in its own body that Deloitte would refund the full AU$440,000. That is the contract value, not the refund, and the same article has it right in the headline. Citing it would have handed a reader a source contradicting the sentence it was there to support.

  • ▾
    This post argues that the AI Act finally gives a buyer somewhere to complain, and still gives nobody anything that asks whether the capability delivered what was paid for. I opened the statute itself and read the articles the post leans on.

    The phrase the whole argument rests on is there. “Information gathered on its own initiative” sits verbatim in Article 89(2)(c). The penalty figures are verbatim in Article 101, and the post now names Article 101 for them. The six complaint grounds I listed match the Commission’s own six bullets, in the same order. Every source on the post carries an address that opens.

    The best thing I found cost nothing. On the Commission’s complaints page, the words value, outcome, performance, benefit, return on investment and effectiveness each appear zero times. That is my claim, sitting inside their own document, and anyone can check it in under a minute.

    One date moved. The post described that material as last updated 31 July 2026. The page now reads 7 August 2026, so the post reads that. Nothing it asserts changed.

    And here is the part I did not expect to be writing yet. The post said an evidence layer would be on sale within a quarter. It is. A compliance vendor published in September arguing that a buyer needs evidence tied to a named interaction, a user, a policy and a date. On that page the words value, buyer, return and before signature each appear zero times. The evidence being built is for the regulator. The sentence in my post has moved out of the future tense.

  • ▾
    This post argues that the industry’s answer to unrealized value keeps arriving as a new office, and that an office can be defunded. I used PwC’s Value Realization Office as the example. The page itself is linked on the post now, with the date I read it.

    Opening it paid better than that. The word “standard” does not appear anywhere on it. That is the point the post makes, and a reader can check it in about a minute instead of taking my word for it.

    One other thing. A summary I came across said the offering ties fees to realized outcomes through shared risk. Had that been on the page it would have been a fair answer to my post. I read both pages and searched them. The language is not there. It was the summary’s, not PwC’s.

    The claim stands, and it now rests on something a reader can open.

  • ▾
    This post reads a CFO’s four steps for governing enterprise AI and argues that every one of them stops at the buyer’s own systems. I named the column, its author and its date, and gave no link. The link is on the post now, with the date I re-opened it.

    I also went back to see who else is working this ground. Buyer-side evaluation guides already tell buyers to settle what counts as delivered before they sign, and to run the acceptance test on their own data. That is good advice and the ground is well occupied. The post says so now.

    What none of them reaches is the billing unit. On the closest of them the word “unit” never appears. A buyer can define what delivered means and still be billed in a unit the vendor picked. That is the gap this post is about, and it is easier to see now that the rest of the ground is named.

  • ▾
    This post says that an ISO/IEC 42001 certificate tells a buyer the seller manages AI responsibly, and tells that buyer nothing about whether what is being bought will produce a result. That still holds. What it did not give a reader was any way to check the facts underneath it.

    I described an accounting firm’s new accreditation, a certification body’s, an earlier one from a competitor, the ISO standard that governs the certifiers, and the accreditation body that lists them. I named all five and gave the address of none of them. All five are on the post now, so anyone can open them. The accreditation announcement is dated 29 July 2026 and says a certificate “provides independent validation of an organization’s AI management system, supporting RFP responses, security reviews, regulatory preparedness”. That is the announcement’s own account of where a certificate is meant to be used, and it is the claim the post answers.

    One thing has changed since I wrote it. The point that a certificate tests a process rather than a result was mine to make when the post went up. The certifiers make it themselves now, and so do the firms selling compliance work. The post says so and credits them, because a reader should know when the people being written about agree with the premise. The same goes for the advice to read the scope and ask whether it covers the system being bought, which has become standard guidance rather than an observation of mine.

    What is left is the part I have not seen anyone else make. Pressing the seller harder does not reach it. The information that decides whether a purchase was worth making does not sit on the seller’s side of the table at all. Only the buyer can say what would count as value here, and the standard the result gets measured against is the buyer’s to hold.

  • ▾
    Two of the figures in this post come from World Commerce and Contracting. One is average contract value erosion of 8.6 percent. The other is 11 percent of contract value lost once an agreement moves into delivery and ongoing management. I went back to the second one to see who produced it.

    It comes from a report World Commerce and Contracting published with Ironclad, which sells contract software. The report’s own page carries the Ironclad logo under the words “In partnership with”, and the credited authors are the president of World Commerce and Contracting together with Ironclad’s chief executive and its chief technology officer. The post now says so in the sentence that uses the number, and the references line at the foot names the partnership and gives the addresses of both sources.

    I kept the figure. A reader should know when a company with something to sell helped produce a study, and that is a different matter from whether the number is right. My whole argument here is that the party being measured should not be the only party holding the evidence. I am not going to apply that to everyone else and leave my own citations alone.

    One line in the report is worth reading beside the post. It says the real loss is often well above 11 percent, because most organizations never formally track what they lose after signature. That is the same hole the post is about. Nobody keeps the record, so the size of the loss is an estimate made by the people who went looking for it.

  • ▾
    This post rests on three dates. The EU AI Act’s high-risk obligations were due on 2 August 2026. They were pushed back, and the whole argument turns on how far. I opened the Council’s own press release again and read the paragraph that sets the new timeline.

    It is dated 29 June 2026. It says the co-legislators “agreed on a fixed timeline for the delayed application of high-risk rules: the new application dates would be 2 December 2027 for stand-alone high-risk AI systems and 2 August 2028 for high-risk AI systems embedded in products.” That is the deferral this post describes, in the words of the body that adopted it.

    The dates hold and nothing in the argument changed. I added the link to the post, so a reader can see the timeline at the source instead of taking my word for the calendar. The point of the piece is that a buyer should not wait on a regulator’s calendar to learn whether the value arrived. A date sitting inside that argument should be one anyone can check in a minute.

  • ▾
    Section 52(4) of the Procurement Act tells a buyer to publish an assessment of its supplier and then leaves the contents of that assessment almost entirely open. I wrote in this post that the gap is a choice about what the published record is for, and the statute was the only thing I had behind it. Two documents written this year make the point better than I could.

    The first is regulation 38A of the Procurement Regulations 2024, in force since 1 April 2026. It lists what an authority publishes about a payment: the paying authority’s name, a contact postal address and email address, and its unique identifier; the same set again for the authority that published the contract details notice where that is a different body; unique identifiers for the procurement and the contract; the supplier’s name and unique identifier; the value of the payment net of VAT; and the date the payment was made. None of that is left to anybody’s judgment.

    The second is Procurement Policy Note 022 on steel, and the guidance attached to it goes further. Paragraph 13 hands the buyer four strings and the box to type them in, one for UK steel, one for no, one for mixed and one for not known, with a rationale required on the last three. Two model texts are supplied so nobody has to invent one. Contractors report origin on a standard template within ten weeks of the financial year ending, the buying organization consolidates it into an annual return to the Department for Business and Trade, and paragraph 73 asks the contractor whether the mill certificate records where the steel was melted and poured.

    Paragraph 66 of that same guidance is what it has to say about whether the contract worked. In full: “KPIs should drive both a focus on outcomes and continuous improvement (aligning with the Social Value Model where appropriate) and the intended benefits to be realised during contract delivery.” There is no field, no set of permitted values, no template and nobody it goes to.

    No figure in the post changed and no conclusion changed. I added a section carrying both documents, with links, so anyone can read the steel paragraphs and the performance paragraph and see for themselves how differently the same department writes when it decides a thing matters.

  • ▾
    This post says the FAR inspection clause for fixed-price services has carried the same shape since 1983, and its references line names that clause as (Aug 1996). A reader following the citation lands on a page headed 1996 with no route from there back to 1983. I opened the eCFR text of 48 CFR 52.246-4 and read the source note.

    It reads: 48 FR 42478, Sept. 19, 1983, as amended at 56 FR 67136, Dec. 27, 1991; 61 FR 31665, June 20, 1996. September 1983 is when the clause was first published. The 1996 label on the page is a revision designation, arrived at after two amendments, one in December 1991 and one in June 1996. Both dates are right and they answer different questions. I read paragraph (e)(2) on the same page, the price reduction remedy the post leans on, and it is there in the current text.

    The date stands and nothing in the body changed. The references line now carries the Federal Register chain, so anyone can confirm the 1983 date without hunting for it. This post argues that a buyer should hold a right that reaches something it can check for itself. I did not want a date sitting inside that argument that a reader could not check.

  • ▾
    This post tells a buyer to keep its own definition of a performance measure in its own files, so the measure cannot move quietly after a difficult quarter. When I wrote that, it was advice. The regulations underneath the Act turn it into something a buyer can check.

    Regulation 39(4)(d) of the Procurement Regulations 2024 sets out what a contract performance notice has to carry. It is “a description of the three key performance indicators set in accordance with section 52(1) which the contracting authority regards, at the time the contract performance notice is published, as most material to performance of the contract obligations.” Materiality is judged on the day the verdict is written rather than on the day the contract was signed. The Cabinet Office guidance on key performance indicators, updated 20 July 2026, states the consequence at paragraph 14: “These KPIs may be different from those set out in the contract details notice or in previous contract performance notices,” because the ones included must be those regarded as most material when the notice is published, “so may change over time.” Paragraph 17 goes further. Where an authority considers that fewer than three of its indicators are still material, it may introduce a new indicator or amend an existing one, provided that is a permitted modification under section 74, or carry on reporting the old one with an Other rating.

    So a buyer chooses which three of its own measures the public sees, and it chooses on the day it publishes. None of that is concealed and none of it breaks a rule. It does mean the published record can move between one notice and the next, and the only thing that holds the earlier meaning is the buyer’s own file. I added a section carrying both sources, including the section 74 gate in full, and one sentence at the end of the post tying the existing advice back to the regulation. No figure changed and no conclusion changed. What changed is that a reader can now check the advice against the rule that makes it necessary.

  • ▾
    In August I wrote that the open specification for computable value models is, whatever else it is, a vendor catalog. I took that from the specification’s own account of what it produces.

    I read it again this week, governance and licensing included. Those pages say who holds the decisions. A value-modeling software vendor started the project and leads every decision until there are ten production implementations, and the prose license requires permanent attribution to that vendor as originator.

    I added those facts to the post. They put the claim on the specification’s own record rather than on my reading of it. The schemas remain free for anyone to implement, and I said so in the same paragraph, because who may use a schema is a different question from who decided what it measures.

  • ▾
    I published this on 5 September, writing about a discussion draft Senator Warner released on 29 June. The draft is no longer a draft. It was introduced in the Senate on 21 July as S. 5051, the Artificial Intelligence Access, Gatekeeper Exchange, and Nondiscriminatory Transfer Act, and referred to the Committee on Commerce, Science, and Transportation. I read the introduced text at govinfo before changing anything in the post.

    The post now cites the bill, and the citation moved with it. In the text as introduced, the clause putting the agent’s duties beyond the reach of a contract sits at section 3(g)(3), and it is more exact than my summary of it: those duties may not be waived, limited or modified by contract, by terms of service, or by any form of user consent. That is what the post quotes now.

    One provision I had underweighted. Certification is not only registration with the Federal Trade Commission. Section 3(d)(4) has recognized certification bodies conducting periodic assessments of conformity, and it requires those bodies to be independent. That makes the argument easier to hold rather than harder: the text still specifies what the arrangement must have, and it now names who checks.

    Nothing in the argument changed. Every duty I described against the June draft I checked again, line by line, against the introduced text, and all of them hold.

  • ▾
    The page I cited for this post is gone, in the ordinary way research pages go. The publisher rolled it forward to its September 2026 editions. Anyone following my source line last week would have landed on figures that were not the figures I quoted.

    I opened the underlying reports. Every August figure I used is still there, word for word, on the two Edition 2 report pages, and the Sources line now points at those. Nothing I claimed changed.

    The September edition also makes the point harder to argue with. It reports the same shape on bases roughly twice the size: two percent of respondents put accountability for value in the chief executive’s seat, thirty percent in technology. A finding that survives a doubling of the base is a sturdier finding than the one I published, and I added a sentence saying so.

  • ▾
    This post rested on two measurements: EY’s ninety-eight percent and KPMG’s twenty-six percent. Two is enough to carry an argument and thin enough to be waved off as a pair of outliers.

    McKinsey published its State of AI survey on 25 August, and it is a third instrument on the same question, taken independently of the other two. It was fielded between 4 May and 8 June 2026 with 1,719 participants across 97 nations. Thirty-seven percent attribute at least some EBIT impact to their use of AI. Six percent attribute five percent or more and call that impact significant. Eighty percent report improved individual productivity.

    The last two figures are the post’s argument arriving from a source that was not looking for it. Productivity improvement is reported nearly everywhere. A number attached to earnings is reported by six percent. I opened McKinsey’s own page rather than working from the coverage of it, and added a paragraph carrying the figures and the survey’s basis. No figure already in the post changed.

  • ▾
    I published this on 3 September, arguing that the Procurement Act 2023 builds the machinery to publish a buyer’s verdict on a supplier and then leaves the subject of that verdict undefined. Section 52(4) says a key performance indicator is a factor or measure against which a supplier’s performance can be assessed, and that is the whole definition. The obvious objection is that a statute cannot reasonably be expected to specify an outcome, because legislation is written to stay open. I did not answer that objection when I published. Section 70 of the same Act answers it. It requires an authority to publish specified information about any payment of more than thirty thousand pounds made under a public contract, before the end of thirty days beginning with the last day of the quarter in which the payment was made. The threshold is a number, the deadline is fixed, and what has to be published is not left to the authority to decide. The same statute that produced section 52(4) was exact when the subject was money leaving the account. That makes the looseness in the performance duty a decision about what the published record is for, rather than a limit on what legislation can specify. I added a paragraph saying so on 5 September. No figure changed and no conclusion changed. What was missing was the answer to the obvious objection, and it was sitting in the same Act.
  • ▾
    This post opens on two Gartner figures: 84 percent of finance organizations have implemented AI or plan to, and 7 percent report a high or very high impact. Both come from Gartner’s release of 8 June 2026, reporting a survey of 183 CFOs conducted the previous June. A different version of the first figure was circulating, that 84 percent of CFOs had not realized a return on their AI investments. That is not a rounding error, it is an inversion. An adoption count had been converted into a failure verdict, and the page that made the switch linked back to the primary source that contradicts it. Forbes subsequently carried the inverted number and attached it to a different survey of 115 CFOs. My figures came from the primary source, so nothing here required correction. The chain was worth documenting, and on 15 August I set it out in The Business Case Cites What It Cannot Trace and added a reference to it here.
  • ▾
    The central claim here began as my reading of two public filings. Before publishing, I counted the remainder: 79 filings, with a single demand for withdrawal standing apart from the industry. The count and the source both appear in the post.

← Back