The Signal.

The Signal — Issue 16

McKinsey’s “The AI value curve” works out, task by task across 831 occupations, what SI could save on existing work, and states that it does not model the organizational, regulatory or workforce constraints that turn potential into results; a revised NBER paper by economists from seven central banks and five universities finds 78% of nearly 9,000 surveyed firms use SI while 83% report no productivity impact over the past three years. Tuesday, October 27, 2026.

Current issue · Issue 17 · November 3 · Issue 15 · October 20 · Issue 14 · October 14 · Issue 13 · October 7 · Issue 12 · September 29 · Issue 11 · September 22 · Issue 10 · September 15 · Issue 9 · September 8 · Issue 8 · September 1 · Issue 7 · August 26 · Issue 6 · August 19 · Issue 5 · August 11 · Issue 4 · August 5 · Issue 3 · July 28 · Issue 2 · July 21 · Issue 1 · July 14

McKinsey priced what SI could save on existing work, and said plainly that it does not model the organizational constraints that decide whether anyone gets it. The same month, economists from seven central banks and five universities reported what executives at nearly 9,000 firms say SI has done to their numbers: for most of them, nothing yet. The ceiling is sized and the capture is near zero. Between them sits a list of decisions with no owner’s name on it.

• McKinsey’s “The AI value curve” sizes what SI could save on existing work and says it does not model organizational, regulatory or workforce constraints. (McKinsey / QuantumBlack, October 2026 · modeled estimate, not measured results · no value figure from it printed here · read from the PDF; no public link claimed.)
• Central-bank surveys across seven countries find 78% of firms use SI, yet 83% report no productivity impact over the past three years. (NBER, “Firm Data on AI”, revised October 2026 · executive survey, sales per employee, response-weighted across country panels.)
• 87% of firms surveyed report no employment impact over the past three years, and executives expect, on average, a productivity gain of about 0.6% a year over the next three. (NBER, “Firm Data on AI” · survey of nearly 9,000 CFOs, CEOs and senior executives, firms with 10 or more employees · all firms, users and non-users · a forward expectation, not a measurement.)

Hold the two findings together. McKinsey sized the ceiling: what the work would be worth if every freed hour turned into value. The central banks measured the capture: most firms surveyed, users and non-users alike, say SI has not moved sales per employee yet. Between the two sits every decision the model does not price — who owns the redesigned work, and where the freed hours go.

McKinsey’s QuantumBlack and its Tech & AI group published “The AI value curve” in October 2026. It asks a narrower question than an adoption count, and a better one. It doesn’t ask how many people use a copilot. It asks whether superintelligence (SI, the term the White House made official on 29 September) can do a given task to acceptable quality at a lower full cost, with human review, validation and rework counted in. The authors’ complaint is that organizations “are still measuring AI adoption instead of value.”

The method works from the bottom up. For each region, industry and occupation, across 831 occupations, it takes the fully loaded labor cost of an hour, subtracts what SI costs to do that hour, and multiplies by the hours the work could hand over. Add those up and you get the curve. The strongest economics come from augmenting senior, high-wage knowledge work: legal, management, healthcare practitioners, chief executives. High-volume routine work is the other large opportunity. The authors say the senior roles are not likely to be replaced; the value comes from augmenting them. The curve is also local. A use case that pays at US wages can lose money in a lower-wage market.

Then the paper states what it is not measuring. It does not model organizational, regulatory or workforce constraints, and it says “economic potential does not necessarily translate into realizable value” (p. 11). One-time transformation and enablement investment (compute, cloud, platforms, training, change management) “would pull the net pool even lower” (p. 21). Its fifth recommendation to leaders is to plan redeployment before deployment, because otherwise “productivity can disappear into lower utilization and unmeasured slack” (p. 22).

Population and screen. A modeled estimate, not a measurement of any company’s results. The model compares SI cost with labor cost and nothing else. It holds model pricing constant worldwide and does not show its token prices, model mix or review-hour inputs, so nobody outside McKinsey can reproduce it. No sector or country figure from it is used here. In several places the paper’s text and its own exhibits disagree, so this issue cites the framework and the paper’s stated limits.

Read it as two numbers that are not on the same page. McKinsey has sized the ceiling: what the work would be worth if every freed hour turned into value. The capture is how much of that ceiling a given company actually banks. That depends on who owns the redesigned work, where the freed hours go, and whether the people around the task agree on what “acceptable quality” means. The paper names that gap again and again, from its summary to its closing paragraph, and never puts a number on it. AJ Maxwell calls the cost of running work without those decisions the coordination tax. McKinsey’s stated limits leave that cost outside the model, between the curve and the P&L.

The paper’s first recommendation fits on one line: audit the tasks in your top five cost centers. Before you do, ask who will own the hours you free up.

Economists from the Atlanta Fed, the Bank of England, the Bundesbank and four other central banks, with Stanford’s Nicholas Bloom among the authors, revised their NBER paper “Firm Data on AI” in October 2026. SI is in wide use at these firms: 78% of them use it. It asks executives what the tools have done to their sales per employee and their headcount.

So far the answer is: not much. Looking back three years, 83% of firms report no impact on productivity and 87% report no impact on employment. Looking forward, executives expect, on average, a productivity gain of about 0.6% a year over the next three years. That’s real money at the scale of an economy. For any single company it’s a modest number.

Our read: the gap isn’t hesitation. Most of these firms already use the tools. What hasn’t happened yet is the conversion: the freed hours, the redesigned roles, the work that’s now done differently and shows up in the accounts. Wide use and almost no reported effect at the same time is, on our reading, what it looks like when tools are bought but work isn’t redesigned.

Population and screen. A survey of executives, not a measurement from company accounts. The data comes from existing firm panels, most of them run by central banks: the Atlanta Fed’s Survey of Business Uncertainty in the US, the Bank of England’s Decision Maker Panel in the UK, and matching panels in Germany, Canada, Australia, Mexico and South Africa. Nearly 9,000 CFOs, CEOs and senior executives, at firms with 10 or more employees. They were recruited by phone and answered online. Fielded May to July 2026, except Germany, fielded January 2026. “Productivity” means sales per employee as the executives report it. The cross-country figures are weighted by number of responses, not by size of economy, so they reflect the mix of panels rather than a global average. Country figures are not used here.

Read it beside the McKinsey segment above. McKinsey sized the ceiling: what SI could be worth if every freed hour turned into value. This paper measures the capture: what executives say has reached their numbers so far. In the space between the two, the decisions about who owns the redesigned work and where the hours go still haven’t been made. The cost of that gap is the coordination tax. McKinsey says outright that it does not model organizational constraints. The central banks’ survey shows what the accounts look like while that cost goes unaddressed.

The tools are bought. The value still has to reach the accounts.

Put the two documents on one desk. They are readings of the same machine, taken at opposite ends. McKinsey reads the input side: task by task, where handing work over costs less than doing it, the strongest economics sit at the senior end of the payroll. The central banks read the output side: across the firms surveyed, most of which already use SI, sales per employee and headcount have barely moved. Neither document contradicts the other, because they are not measuring the same thing. One prices the hours. The other asks whether anyone banked them.

What sits between an hour freed and a number moved is a short list of decisions, and every one of them has to be made by someone with a name. Who owns the redesigned task. Where the freed hours are reassigned, and on whose budget. Who signs off that the output quality is acceptable, and against what standard. McKinsey’s paper says it does not model organizational, regulatory or workforce constraints, and on our reading every item on that list is one. The central banks’ survey is what the accounts look like while that list stays unassigned.

So read the two headline findings together. The ceiling is real, on the model’s own terms. The capture is near zero, on the executives’ own reports. What separates them is a list of unmade decisions with no owner. Start where McKinsey starts: the tasks in your top five cost centers. Then do the part neither document can do for you, and put a name next to the hours.

No McKinsey value figure is printed in this issue. The only number taken from the paper is its count of 831 occupations. The paper carries a global total and sector and country pools; in several places its text and its own exhibits disagree, and its sensitivity comparisons are not internally consistent, so the issue cites the framework and the paper’s stated limits and refuses the numbers. A different McKinsey finding from April 2026, built on a different threshold, is not brought in as context; fusing two McKinsey publications is a defect this publication has logged before. There is no public link to the paper here: mckinsey.com does not answer from this host, and the issue read an AJ-sent copy of the PDF, which is said plainly in Verification rather than linking a page nobody fetched.

From the NBER paper, four figures run and nothing else. Refused: the per-employee and aggregate spending figures, the spending-concentration split, executives’ own usage hours, the bucket-midpoint averages behind the headline shares, and every single-country bar. The conference deck’s own per-annum headline disagrees with the paper’s abstract and with the deck’s own three-year chart, and is never cited; the figure here is the abstract’s. An earlier, February version of this paper carried different headline shares; that version is superseded, and its figures are retired on this publication’s own list.

Held. The survey’s forward-looking employment churn finding — offsetting hiring and cuts under a flat net number — is real and is not run, because its headline share collides with a figure this publication has reserved for another source.

Named gap. Nobody in this issue measures capture from company accounts. McKinsey models; the central banks ask executives. The number that would settle it — audited results at firms that redesigned the work — does not exist in either document.

BUILD AND AUDIT, 27 October 2026. §02 and §03 arrive from banked, AJ-greenlit segments; the copy pass ran first and the claim audit second (R19b), with both local PDFs opened in full by the audit seat on 27 October 2026 and its fixes applied to the text above (reviews/signal-issue-16-claim-audit-2026-10-27.md).

McKinsey / QuantumBlack and Tech & AI, “The AI value curve: The economics of applying AI to existing work” (AJ-sent PDF, 22 pp, October 2026). All 22 pages read by the audit seat 27 October 2026. Quoted spans verified character by character: pp. 1-2, p. 11, p. 21 (antecedent: one-time transformation and enablement investment), p. 22 (action 5). Method facts p. 4. Negative claims read off the full document: does not model organizational, regulatory or workforce constraints (p. 11); excludes one-time transformation costs (p. 21); shows no token prices, model mix or review-hour inputs (all pages). No public link: mckinsey.com did not answer from this host on 27 October 2026; the paper is named in prose. Carries §02, signal 1.

NBER working paper 34836, “Firm Data on AI”. FETCHED 27 October 2026 — CONFIRMED, PRIMARY PAGE. Page shows: title “Firm Data on AI”; authors Yotzov, Barrero, Bloom, Bunn, Chawla, Davis, Foster, Jalca, Meyer, Mizen, Navarrete, Oh, Garcia Loredo, Reid, Smietanka, Thwaites, Wang et al.; Issue Date February 2026, Revision Date October 2026. The page prints no day; this issue cites October 2026. Abstract confirms verbatim: “We survey nearly 9,000 CFOs, CEOs and senior executives at firms in the US, UK, Germany, Australia, Canada, Mexico and South Africa in Summer 2026”; “78% of firms use AI”; “executives report that AI has had essentially no net effect on firm employment over the last 3 years”; “expect a substantially larger 0.6% per annum boost over the next 3 years”. The abstract’s spending, usage-hours, churn and realized-gain figures are not used. Carries §03, signals 2 and 3.

NBER conference deck (AJ-sent PDF, 32 slides, dated 20 October 2026), all slides read by the audit seat 27 October 2026. 83%: p23, all firms, N=8,438, sales per employee, past three years. 87%: p28, all firms, N=8,744. Survey design pp. 6-8. Seven central banks among the author affiliations, p1. The deck’s summary-slide forecast and its slide-4 headcount and bank count conflict with the abstract and are not used. Superseded February figures appear nowhere in this issue. The paper PDF itself was not opened. Carries §03.