Executive Summary

The Monetary Policy Report is the Bank of England's quarterly diagnosis of the British economy. The edition dated 5 November 2026 carried the decision the Monetary Policy Committee had reached the day before: eight of the nine members voted to raise Bank Rate by 0.25 percentage points to 4%, and one argued for a larger rise. In the whole document AI turns up in exactly two places, both of them standalone explainers that the report calls boxes.

Box C deals with what AI has added to growth. Because no statistic counts AI directly, Bank staff built a proxy out of the combined output of three industry classifications. Publishing (SIC 58) covers the development of generative AI agents and application software; computer programming and consulting (SIC 62) covers AI model selection and systems consulting; information services (SIC 63) covers AI data processing and data centre services. The report then strips out the longer-run expansion of technology-related activity and calls what is left "novel AI-related activity".

The same box attaches two caveats of its own. One is that "not all the output of these sectors will capture AI activity". The other is that some of the growth the proxy picks up may itself be the longer-run expansion of technology-related activity. What the report does not say is how large that trend was taken to be. Neither the figure before the subtraction nor the figure after it appears anywhere; only the result, "just under a fifth", survives.

The second place, Box D, is the one this article follows. There the report sets out the productivity effect of AI across three sentences, all resting on a single business survey. The survey is the Decision Maker Panel, a monthly panel of UK company executives the Bank has run since 2016, generally shortened to DMP.

The press handled those three sentences accurately. The two outlets that covered this box on the day of publication and the day after both kept the 0.5% three-year cumulative figure apart from the 0.9 percentage point annual one, and both said in the body of the piece that the source was a survey of company executives. So this article is not aimed at the coverage. The compression happens earlier than that, inside the report.

Box D also supplies one piece of background. In September 2026 the Office for National Statistics changed how it measures labour input. The old method took hours worked from the Labour Force Survey, where a low response rate made those hours look longer than they were. Under the new method, which mixes in administrative data, annual growth in output per hour over 2009–2024 rose from around 0.5% to close to 1%. Output did not move; the denominator did. Quarterly statistics on the new basis are published on 17 November 2026, and industry-level figures follow in February 2027 by the report's own reckoning. Every figure in this article is as of 8 November 2026.

November was not the first time the Bank put this survey question into a report. The February and July 2026 editions each carried a sentence on the same question. Download all three, line the relevant passages up, and three things turn out to have moved over six months: the figure rose, the name of the unit changed, and a citation appeared.

Each sentence is quoted from the full PDF of the edition in question. The word Yotzov does not occur once in either the February or the July text (zero hits in a search of both PDFs).

The rise from 0.6 to 0.9 has a plain explanation: the survey wave changed. February drew on the winter wave collected between November 2025 and January 2026; July and November both draw on the summer wave collected between May and July 2026. Asked the same question six months later, firms wrote down a bigger expectation.

The second thing that moved is the name of the unit. July's "0.9% per year" reads as an annual effect on the level of productivity; November's "0.9 percentage points per year" reads as a contribution to the growth rate. The two sentences do not point at the same quantity. The figure, meanwhile, did not move. A tool that compares numbers alone will never catch this, because 0.9 and 0.9 are the same value.

Third, the citation and the "split roughly equally" decomposition only arrive in November. Read that as evidence in the Bank's favour. Of the three editions, November is the most precise: it names the paper the figure came from and sets out how the figure divides between the employment side and the output side. The point lies elsewhere. A number rounded to one decimal place cannot tell you which step it came out of. The only things that can tell you are the unit name and the citation beside it, and both come and go from edition to edition.

2.1Setting 0.5% beside 0.9 percentage points makes the gap look smaller

Put the first and third sentences of Box D next to each other as written and you get 0.5 and 0.9, a gap of 1.8 times. But the first is a cumulative record over the past three years and the second is an annual contribution over the next three. They are measured with different rulers. The report itself converts the 0.5% into "about 0.1–0.2 percentage points" a year in that same first sentence. Measured on that converted scale, the distance between record and expectation is between 4.5 and 9 times. Matching the units does not shrink the gap; it widens it.

2.2The same report forecasts less productivity growth than that

An annex to the same report carries the Bank's own productivity forecast. Table A1.C projects labour productivity (output per worker) at 0.4% for 2026, 0.7% for 2027 and 0.8% for both 2028 and 2029. What firms expect from AI alone is a contribution of 0.9 percentage points a year, while total productivity growth in the Bank's central projection runs at 0.7–0.8% a year over the next three years. The units line up: the paper's 0.85 percentage points is gross value added per worker, and the annex forecast is output per worker.

A line between fact and reading. The report does not set these two figures side by side. They merely sit in the same document, and placing them side by side is our doing. Nor does the comparison show the report contradicting itself. The Bank passed the survey figure on without putting it into its own forecast, and says as much in its own words.

That passage is what keeps this article balanced. The Bank did not take the survey figure on trust and change policy on it. It quoted the figure, named the source, and kept it apart from its own forecast. The tension here does not come from carelessness at an institution. It comes from the fact that the moment the sentence leaves the report, the marker keeping it apart falls off.

The paper the report cites is Yotzov and co-authors, Firm Data on AI (NBER Working Paper 34836), written in February 2026 and revised in October. Its authors put the same five questions onto business surveys in seven countries, with the DMP carrying the British share. Follow the paper and the aggregate data and you find four transformations standing between the executives' answers and the report's "about 0.9 percentage points". Not one firm wrote that number down. What a firm did was pick one option on a multiple-choice question.

Start with the question those firms were asked, word for word: "How has the adoption of artificial intelligence technologies affected the volume of sales per employee in your business over the past three years?" The paper is explicit that "productivity in this analysis is defined as the volume of sales per employee". Volume rather than value matters here. Volume is what separates real gains from price rises, and seeing the distinction is what makes visible the step the report took when it rendered these answers as "productivity".

An executive has five boxes to choose from: large decrease above 5%, small decrease below 5%, no impact, small increase below 5%, large increase above 5%. There is no field for writing in a number. Four steps then turn that choice into a sentence in the report.

Step three is itself the product of a chosen rule. To collapse scattered answers into a single number you first have to decide how heavily each firm's answer counts, and the rule applied to the UK column is employment weighting, so the answers of large employers carry proportionally more weight. The paper states the rule in the notes to every figure, and it is not the same everywhere: the US, UK, Germany and Mexico are employment-weighted, while Canada, Australia and South Africa are equal-weighted. How far the rule can move a number is something the paper measured directly on a different question — swap the weighting scheme alone and the UK AI adoption rate moves by five percentage points (section 6.3). The percentages in the published aggregate table have already been through this rule, which is why putting the values back in the next section reproduces the same average.

3.1The outermost box is open at the top

The weakest link in the step-two value assignment is the box at either extreme. The label "large increase, above 5%" has no upper bound. Someone thinking of 6% and someone thinking of 40% land in the same box. The researchers decided to put 7.5 there. For a closed box you can use the midpoint; for a box open at the top, no midpoint exists. Whatever goes in is an assumption.

3.2The four assumptions behind step four

Turning the survey average from step three into a contribution to productivity growth requires a picture of the whole economy. The notes to Table 3 of the paper set out the assumptions in four lines. AI spending is taken as £313 per employee per year over the past three years and £537 over the next three; that spending is assumed to raise intermediate inputs in proportion to employment; labour and capital shares of 0.67 and 0.33 are applied; and the capital-deepening component is assumed to come from AI spending equal to 6% of the total capital stock. The three rows of Table 3 are the result.

The UK column of Table 3 in NBER Working Paper 34836. The report's "about 0.9 percentage points" comes from the middle row. Pick a different row and the same survey yields 0.7 instead of 0.9.

3.3Two routes meet at one decimal place

Here is where it becomes hard to say which route the report's 0.9 percentage points took. Divide the step-three survey average of 2.6190% by three years and you get 0.873 a year. Run it through the step-four growth accounting instead and you get 0.85. Two values from two different steps — and rounded to one decimal place, both become 0.9. The sentence in the report cannot tell you which.

Of the two routes, the divide-by-three one leaves tracks in the earlier editions, starting with February's "0.6% per year". The winter wave's UK three-year cumulative figure of +1.8585% divided by three is 0.6195. July's "0.9% per year" is the 0.8730 just derived above, and the "0.4% per year" for employment in the same July sentence comes from dividing the summer wave's employment expectation of −1.3130% by three, which gives −0.4377. All three figures reproduce as the survey average divided by three. What the two earlier editions printed was the survey average itself, with no growth accounting applied.

Because November cites Yotzov and co-authors, the growth-accounting route is the likelier one there. But July's "0.9% per year" was written when there was neither a citation nor a growth-accounting step, so it can only be the survey average divided by three. We record the facts without settling the question. Rounding does not merely shorten a number; it erases the steps the number came through. Where two routes meet at the same point, whoever is standing there can no longer say which one they took.

November's "split roughly equally" checks out the same way. In the aggregate data the UK expectation for productivity over the next three years is +2.619% and for employment −1.313%. Divide each by three and you get +0.873% and −0.438% a year; subtract the employment fall from the productivity figure and what is left on the output side is +0.435% a year. The employment side at 0.438 and the output side at 0.435 part company only at the third decimal. The report's "roughly equally" is an accurate description.

This article could get this far only because the paper's project published its aggregate data as a spreadsheet. For every question and every country it lists the share of responses in each of the five boxes to four decimal places. Given five shares and the value-assignment rule, anyone can rebuild the average. So we rebuilt it.

Two sets of code written independently of each other produced the same numbers both times: a UK cumulative average of +0.5355% over the past three years, +2.6190% over the next three, and an employment expectation of −1.3130% over the next three. All three match the published figures to four decimal places. The claim of this article is not that the numbers are wrong. The arithmetic holds at every step.

What the rebuild exposes is what the average covered up. Asked about the past three years, eight UK firms in ten (82.24%) ticked "no impact". Firms reporting any productivity gain amount to fewer than two in ten — 14.40% small and 2.66% large. The 0.5% in the report is an average that this minority pulled up. Move to expectations for the next three years and the picture changes: "no impact" falls to 35.17%, with small increases at 38.05% and large increases at 23.71% taking up the space.

4.1What moved over six months is the number of people who changed boxes

The published file carries a separate winter-wave block below the summer wave. Put the UK column of the two waves together and the route the average took becomes visible. For the past three years, "no impact" fell from 89.32% to 82.24%; for the next three years, it fell from 53.24% to 35.17%. That many respondents shifted into the increase boxes, and the cumulative averages rose from +0.292% to +0.5355% and from +1.859% to +2.619%.

What deserves attention is the nature of the force that lifted the average. If someone had written in a bigger number, the story would be simple. But there is no number to write in. The values on offer are ±7.5, ±2.5 and 0 — five of them — and a respondent only picks one. What moved is not the values but the number of people who changed boxes. That same shift goes out into the world as the sentence "firms now see a larger effect from AI".

4.2Shaking the 7.5 in the outermost box

How much of the average rests on the 7.5 placed in the open-topped box is something you can see by changing that value alone and rebuilding. Holding the other box values and all the response shares fixed, we swapped out only the number assigned to "large increase, above 5%".

Recalculated on the UK next-three-years distribution in the aggregate data, changing only the assigned value. The right-hand column simply divides the three-year cumulative figure by three, which is a different ruler from the report's growth-accounting figure.

Move the one number nobody picked between 6 and 10 and the survey average swings between 2.27% and 3.20% — 13% below or 22% above the middle case. One step further is where we stop. This sensitivity belongs to the survey average itself, whereas the report's 0.9 percentage points has growth accounting laid on top of that average, so the proportions do not carry across exactly. What can be stated precisely is this much: the 2.6% average the report's figure starts from could sit anywhere between 2.3% and 4.3%.

4.3Publishing the sensitivity alongside is the standard prescription

Recovering a mean from binned categorical data is a problem that income and wealth distribution statistics have worked on for decades, and three conclusions from that literature apply here without modification. First, putting the midpoint into a closed bin usually works well. Second, the error concentrates in the open-ended bin at the end, because no midpoint is defined there, and most of the error in the overall mean originates in it. Third, the standard prescriptions follow from that: estimate the tail and impute it, use a separately published overall mean to back out and constrain the end bin, or decline to pin down a single value and publish the sensitivity instead.

What cannot be carried over needs saying just as plainly. That literature deals with income distributions that really are unbounded above, whereas the end bin here is "volume of sales per employee up by more than 5%". There is no basis for assuming a tail as heavy as income, so the techniques cannot simply be lifted. What does carry over is three facts: the open end bin dominates the error in the mean, a single deterministic value erases the variance inside the bin, and good practice sends the sensitivity to that value out alongside the number.

The same survey also asked about employment. Asked how AI would affect headcount at their own firm over the next three years, almost half of UK firms (46.9%) picked a box on the decrease side. One in eight (12.1%) said the fall would exceed 5%. What survives into the report is an average of −1.3% and the clause "split roughly equally".

Source: aggregate data published with NBER Working Paper 34836, UK DMP responses collected May to July 2026. On the same question for the past three years, "no impact" runs at 87.54% and the cumulative average is −0.0925%, consistent with the report's "close to zero".

The past and the future sit far apart here. Asked about the past three years, close to nine firms in ten said employment had not changed. Asked about the next three, almost half of those same firms pick a decrease. What the Bank's regional Agents hear from companies points the same way: AI adoption "is reducing the likelihood of firms hiring or replacing workers, but not materially increasing the likelihood of firing workers so far".

5.1"Split roughly equally" would have been wrong six months earlier

As the previous section confirmed, November's "split roughly equally" matches the summer-wave figures exactly. Run the same calculation on the winter wave's UK column, though, and a different picture appears. We derived the output-side expectation by subtracting the employment fall from the productivity expectation, then took the ratio between the two.

This split is our calculation. Neither the report nor the paper presents it this way. We pulled the UK column out of both wave blocks in the published aggregate data and did one subtraction; anyone with the same file can reproduce it.

Six months earlier, close to three quarters of what British firms expected came from cutting staff to raise productivity. By summer it was an even split. Almost all of the increase in the productivity expectation came from the output side, while the employment expectation barely moved, from −1.36 to −1.31. Had February written "split roughly equally", the sentence would have been wrong. The line in the report is a photograph of that wave, not a property of the firms.

The paper confirms the same direction by another route. For UK firms the relationship between expected employment effects and expected productivity effects is negative and statistically significant, but the slope is −0.07, well below one in absolute value. The paper's own reading: "Firms therefore expect a substantial share of the productivity gains from AI to arrive through higher output rather than lower headcount."

5.2The report puts three explanations side by side for the same phenomenon

For the recent British pattern of slowing employment alongside rising output, Table D.A in Box D lists three explanations in parallel. The first is cyclical adjustment: firms unwinding the labour they hoarded through the pandemic. The second is the rise in labour costs — as the National Living Wage and employer national insurance contributions push lower-paid jobs out, average productivity rises mechanically. The third is AI. The report marks the first two as having no material persistent effect and reserves "structural and likely persistent" for AI alone. How much of the pattern each of the three accounts for, it does not say.

The Bank of Korea put the same five questions to executives at 749 firms with 50 or more regular employees, online, between 15 and 26 June 2026. Appendix C of the paper says the questions were translated into Korean while preserving their structure, economic concepts, numerical thresholds and response categories, and that industry weights came from employment shares in the 2024 Census on Establishments. When the questions match and the translation is faithful, the results can be set side by side. Do that, and the Korean numbers jump.

Source: Appendix C of NBER Working Paper 34836. The +0.4% seven-country figure in the third row is Appendix C's rendering; the main text and the aggregate data write the same value (0.4647) as "around 0.5%". For the comparison with Korea we kept to Appendix C's ruler.

The realised figure for the past three years is four times higher; the expectation for the next three is twice as high. Better not to blur those two multiples together. The gap on the realised side is far wider than the gap on the expectations side.

6.1The paper attributes the difference to survey context

The authors kept the Korean results out of the seven-country average and put them in an appendix instead, and they wrote down why.

Reading this as evidence that Korean firms are more productive gets it wrong. What the paper is pointing at is not a difference in productivity but a difference in who sat down in front of the questionnaire. And the explanation does not stay at the level of conjecture. A curve of exactly this shape is already drawn inside the main sample's own data.

6.2How much the respondent uses AI predicts the answer

The paper asked separately how many hours a week each respondent personally uses AI, and a footnote records how productivity expectations split across those bands. Respondents who never use AI expect +1.1%; those who use it more than five hours a week expect +3.9%. The employment side widens too, from −0.7% to −2.2%. Korean respondents average 4.1 hours, and the Korean expectation is +3.3%.

Korean respondents land almost exactly on the right-hand end of the curve the main sample drew for itself. The paper does not read this as simple optimism. Its reading is that respondents more directly exposed to AI see both the gains and the displacement as larger, since their expected employment fall grows alongside.

A number from the previous section turns up once more inside this curve. The hours on the horizontal axis are themselves built by assigning values to multiple-choice answers. The note to the paper's figure puts it this way: "values are assigned to each of the options: 0 to 'Not at all'; 0.5 to 'up to 1 hour a week'; 3 to '1 to 5 hours a week'; 7.5 to '>5 hours a week'." The rule that puts 7.5 into an open-topped end box is at work here as well. The ruler used for the check is the same ruler as the thing being checked. That does not make the curve useless. The paper included it to stress-test its own result, and as a diagnostic it does its job.

6.3The paper spends a section measuring why surveys disagree

The point this article is working towards appears, in a section of its own, inside the paper that produced the figure. The authors note that measures of firm AI adoption vary widely across recently published studies, pin part of the gap on three features of survey design, and then measure how wide a swing each feature produces inside the UK DMP itself.

Sources: main text, Table A2 and Figure A7 of NBER Working Paper 34836. All three rows come from the same set of responses to the same question about whether the firm uses AI.

The paper's own conclusion: "Taken together, these three factors each move measured adoption by between five and twelve percentage points, and they compound: a survey of small firms, weighted by business counts and answered by non-executives, will report substantially lower adoption than one of large firms, weighted by employment and answered by CEOs. They therefore narrow the range of estimates in the literature without closing it. Differences in question wording and in survey timing remain important additional explanations."

For why chief executives answer higher, the paper offers two readings and declines to choose between them. Executives may see their organisation's AI use more completely, in which case surveys of junior staff under-report; or they may be more willing to report use they have not seen directly, in which case executive surveys over-report. Either way, the paper concludes, comparing adoption rates across surveys means counting who answered the questionnaire. We decline to choose between the two as well. Another case where answers written down by executives become an industry metric is covered elsewhere on this blog (the Bessemer and Bain survey of hospital AI adoption).

6.4Both the US and Korea hold three numbers at once

The same thing happens in the United States. A FEDS Note published by the Federal Reserve Board on 3 April 2026 sets three surveys of US AI adoption side by side and explains, in its own words, why they disagree.

Source: Federal Reserve Board, FEDS Note, "Monitoring AI adoption in the U.S. economy" (3 April 2026). The note covers adoption rates only and does not address productivity effects.

The Fed's explanation reads: "The different sampling distributions combined with adoption heterogeneity across firm size classes likely drives a considerable share of the gap between the BTOS and SBU estimates." To that it adds differences in question wording, how much respondents know about their own firm, and social desirability bias — the pressure on an executive to report AI adoption as a mark of efficiency.

The survey producing the highest of the three numbers, the SBU, is the US source feeding the seven-country main sample of this very paper. Which means the Fed's own note has flagged the possibility of social desirability bias in that survey. Since the note deals only with adoption rates, we do not carry it across into productivity figures here.

Korea has the same structure. The 2024 Survey of Business Activities puts firm-level AI use at 9.6%; the Bank of Korea's household survey puts it at 51.8% for individual workers; Appendix C of this paper puts it at 89% among executive respondents. The three look incompatible, but each counts a different thing. The Bank of Korea issue note itself reads the gap between 9.6% for firms and 51.8% for workers as a sign that AI is spreading at the level of the individual worker rather than through changes to firm-level workflows. For reference, another research team analysing the same Survey of Business Activities, narrowed to incorporated companies with 50 or more regular employees and at least KRW 300 million in capital, arrives at 9.2%. Same source, different population filter.

6.5The biggest block in the seven-country average answered in January

There is something to note on the denominator side too. In the aggregate data the German column carries identical values, to the decimal, in the winter and summer waves. Germany did not field this question in the summer wave, and the January 2026 data was reused as is. That German column is the single largest block in the seven-country average at 26.7% by response count, ahead of the UK at 23.3%. The paper states this in a table note; we did not uncover it so much as pass along what the paper wrote down first. Still, anyone reading the "seven-country average" as a present-tense number should know that the heaviest weight in it belongs to a country whose answers are six months old. The winter wave had four participating countries — the UK, the US, Australia and Germany — and seven only from the summer wave onwards.

Read this far and it is easy to conclude that surveys are the problem. But go looking for an alternative to surveys and, as of 2026, the answer is close to "there isn't one". No central bank and no statistical agency holds an instrument that measures the effect of AI directly. What separates these institutions is not whether they use a survey but whether they kept a second thing to check the survey figure against. Three institutions that did so follow.

7.1The Bank of Korea asked one more question

In May and June 2025 the Bank of Korea's Economic Research Institute surveyed 5,512 employed people aged 15 to 64 across the country. It drew the sampling frame from the microdata of the Regional Employment Survey, stratified by occupation subgroup, age band and sex, and applied post-stratification weights. The result is BOK Issue Note 2026-12, "Does AI adoption raise productivity? An analysis of the first three years". One of its three authors, Samil Oh, is also a co-author of the NBER paper above. The same person asked executives on one side and workers on the other.

What the worker survey found is this. Workers using generative AI spend on average 3.8% less time on the same task — about 1.5 hours a week against a 40-hour week. Assume every saved hour flows back into production and the potential productivity effect converts to roughly 1.0 percentage points, but a footnote in the issue note states explicitly that this is to be read as an upper bound. The authors wrote down the strength of the assumption before anyone else could.

What comes next is the important part. The Bank of Korea asked separately how much time had been saved and how much more work had been done, then compared the two. At the level of the individual worker, the correlation between time saved and increase in work completed was 0.00. In a regression controlling for individual characteristics and fixed effects the coefficient was 0.011 with a standard error of 0.166, nowhere near significance. The researchers named the phenomenon the "AI productivity disconnect". That conclusion was available only because a second question existed to check whether saved time turned into output. The same problem viewed through different data is covered elsewhere on this blog (a recount of 246 METR developer trials).

7.2The European Central Bank checked its survey measure against accounting data

The European Central Bank also counts AI through surveys. It adds special questions on AI use to the SAFE survey of roughly 5,000 euro area firms, and picks up the worker side through its Consumer Expectations Survey. The difference is that it did not take the survey-built measure at face value. Occasional Paper 395 constructs a proxy for firm labour productivity out of SAFE responses and then benchmarks that proxy against ORBIS company accounts and the euro area labour productivity index. This is not a study that measured the effect of AI; it is a study of how well a productivity measure built from a survey matches accounting data. As an instance of a second source placed beside a survey figure, though, it is exactly on point.

The ECB blog also records a range. Estimates of the additional annual productivity growth AI might deliver over a ten-year period scatter between 0.1% and 3.4% depending on who is estimating, and the ECB's own estimate is around 0.35 percentage points a year for the euro area. A spread of more than thirty-fold is currently in circulation as the answer to one question.

7.3The DMP has a validation system too, and it does not reach the AI questions

The third case is the DMP itself. Section 6 of Bank of England Staff Working Paper 1096, "The Decision Maker Panel: a user's guide", checks the panel's forecast accuracy by a published method. Panel firms are shown five scenarios for sales, prices, employment and investment and attach a probability to each, which yields a mean and a standard deviation. Figure 25 sets those expectations against what actually happened a year later: "firms make reasonably accurate predictions in the cross-section, with a strong correlation between expected growth rates and subsequent outturns a year later." Accuracy holds at the aggregate level too, though forecasts "have, perhaps unsurprisingly, been less accurate in periods when the economy has faced large unanticipated shocks". Sales growth outside ±100% is treated as an outlier, as are prices, employment, costs and wages outside ±50%.

The AI special questions sit outside all of that machinery. They are not five scenarios with probabilities but one of five boxes, so there is no standard deviation. There is no realised observation a year later to compare it with, so no Figure 25 or 26 style validation is possible. And because the value comes out of a box, it cannot stray beyond ±5%, so no outlier rule will ever bite. None of this is a knock against the DMP. Quite the opposite: it is one of the rare business surveys with a published ex-post validation system for its core variables. The point is that the system structurally cannot reach the special questions, while the sentence in the report presents both kinds the same way, as what firms answered.

Where the system cannot reach, the paper found a way in from the side. The authors took the most obvious objection — that firms expecting a lot are simply optimistic firms — and tested it first. Figure A23 reports "no systematic relationship between expected AI productivity impacts and firms' sales growth forecast errors. Expectations of larger productivity impacts from AI are not associated with firms' general forecasting optimism." Figure A18 finds the same on the employment side. That test was possible only because each firm's past forecast errors are stacked up in the panel. A validation system that cannot reach the AI question reached it sideways, through the habits of the firms answering it.

7.4Collecting the instruments that exist today

Gather up what the major central banks and statistical agencies use to measure AI in 2026 and the table below is what you get. Not one of them holds a measuring instrument.

Sources: each institution's own 2026 publications. The right-hand column is where this article is looking.

Central banks lean on surveys for a reason other than laziness: surveys are the only instrument that exists. Which means the test of an institution cannot be whether it used one. The test is whether it put one more thing beside the survey figure to check it against, and in that slot sit the Bank of Korea's second question, the ECB's accounting data, and the DMP's ex-post validation system. By this test, the best thing the Bank of England did was to keep the survey figure out of its central projection and hold it as an upside risk.

What we did in this article is what DataClinic does on its first pass through a client's metrics. Instead of trusting the number written in the report, open the source aggregate table, look at the distribution, rebuild the average, and shake it to see what it is sensitive to. Doing that showed that "0.9 percentage points" is a value four transformations downstream of a five-box multiple choice, that eight firms in ten ticked the empty box, and that the average hangs on one number nobody picked. This time we learned one more thing. Following the same question through three editions of the report exposed the place where the unit name changed. The cheapest check in a metrics review, and the one that pays off most often, is holding the same metric up against its own earlier editions. Tracing that kind of history can be automated.

8.1Three data quality lessons

First, the unit breaks before the lineage does. Put a three-year cumulative figure next to an annual contribution and a gap of 4.5 to 9 times looks like 1.8 times. And "0.9% per year" against "0.9 percentage points per year" will never show up in a value comparison, because the figures are identical. A unit name is a label rather than a value, and labels generally do not travel with the number.

Second, a value placed into a categorical response is configuration, not data. Putting 7.5 into "above 5%" was a choice, not a record, and moving that number only between 6 and 10 drops the survey average by 13% or raises it by 22%. Configuration ought to travel with the value. As things stand it lives in the body of the paper and does not follow the sentence into the report.

Third, averaging erases the distribution, attribution erases the pipeline, and rounding erases the steps. The mean of a distribution where eight in ten sit at zero becomes "firms estimate"; a value four steps downstream becomes "firms answered"; and 0.873 and 0.85 become the same 0.9 at one decimal place. All three show up in every client metrics review we run.

8.2Five questions to ask before publishing an AI impact number

Your own company's "impact of AI adoption" number is also, in all likelihood, something somebody wrote on a survey form somewhere. Checking five things before it reaches a slide will head off most of the failures this article has walked through.

• Is this a recorded value or a value assigned to a chosen box? If it was assigned, does the assignment rule travel with the metric?
• Does the unit hold all the way through? Are cumulative and annual, level and growth rate, mixed on one slide? Have you compared it letter by letter against how the same metric's unit was written last quarter?
• What shape is the distribution behind the average? What percentage of respondents sit at zero?
• Who answered? Does how much AI the respondent uses predict their answer? Does seniority predict it?
• Is there a second thing to check against? Another question, accounting data, or a later outturn.

All five transfer outside central banking unchanged. In this survey, how much AI the respondent used spread expectations by more than threefold, and seniority moved the adoption rate by 4 to 7 percentage points. There is no reason an internal survey would behave differently.

8.3Where Pebblous stands on this

Earlier pieces on this blog mostly interrogated how something was measured: how a summarisation rule gets chosen, what a comparison design can hide, where a denominator changes. This one looked at something slightly different — the record of the steps a value came through. From a five-box multiple choice through four transformations to a sentence in a central bank report, we followed what gets added and what falls away at each step.

Sources for the figures in the text, grouped by kind. Items 1 to 7 are reports and official material published by the institutions themselves. Items 8 to 11 are the paper the figures come from and the methodological literature. Items 12 to 14 are 2026 material from other central banks and policy bodies. From item 15 onwards are the outlets that covered this report and Pebblous blog pieces that take up the same question elsewhere.

Institutional reports and official material

• 1.Bank of England, Monetary Policy Report — November 2026, 5 November 2026. bankofengland.co.uk Box C and Box D in section 1, and Table A1.C plus the upside-risk quotation in section 2, come from here.
• 2.Bank of England, Monetary Policy Report — July 2026, July 2026. The "0.9% per year" row in the three-edition table in section 2 is this edition's sentence.
• 3.Bank of England, Monetary Policy Report — February 2026, February 2026. The "0.6% per year" row in the three-edition table in section 2 is this edition's sentence.
• 4.Bank of England Staff Working Paper No. 1096, The Decision Maker Panel: a user's guide, November 2024. The forecast-accuracy checks and outlier rules in section 7.3 are in chapter 6 of this document.
• 5.Bank of Korea, BOK Issue Note 2026-12, "Does AI adoption raise productivity? An analysis of the first three years", Seo Dong-hyun, Oh Sam-il and Yoon Jong-won, June 2026. The sample design, the 3.8%, the correlation of 0.00 and the regression coefficient in section 7.1 come from the 17-page full text.
• 6.Bank of Korea, BOK Issue Note 2025-22, "The rapid diffusion of AI and its productivity effects: evidence from a household survey", 18 August 2025. The original source for the 51.8% worker-level figure in section 6.4.
• 7.Office for National Statistics, "Measuring labour productivity: our new approach", ONS Blog, 17 September 2026. Background to the change in labour input measurement in section 1. The 2009–2024 figures in the body follow Box D of the report rather than this blog post.

Papers and methodological literature

• 8.Yotzov, V. et al., Firm Data on AI, NBER Working Paper 34836, written February 2026, revised October 2026. The aggregate data file is the version published on the project site (updated 6 October 2026). The five steps and Table 3 in section 3, the distribution and the rebuild in section 4, Figure A24 in section 5, Appendix C and the three survey-design factors in section 6, and Figures A18 and A23 in section 7.3 all come from this paper.
• 9.Bick, A., Blandin, A., Deming, D. J., Fuchs-Schündeln, N. & Jessen, J., Mind the Gap: AI Adoption in Europe and the U.S., NBER Working Paper 34995, 2026. Cited alongside in the report as evidence on adoption comparisons. This article does not use its adoption figures.
• 10.von Hippel, P., Scarpino, S. & Holas, I., "Robust estimation of inequality from binned incomes", Sociological Methodology. The discussion of open end bins in section 4.3 draws on this line of work.
• 11.Jargowsky, P. A. & Wheeler, C. A., "Estimating Income Statistics from Grouped Data: Mean-constrained Integration over Brackets", Sociological Methodology, 2018. The mean-matching prescription in section 4.3 is this paper's method.

Other central banks and policy bodies

• 12.Federal Reserve Board, "Monitoring AI adoption in the U.S. economy", FEDS Notes, 3 April 2026. The 18%, 41% and 78% figures and the passage on social desirability bias in section 6.4 are here.
• 13.European Central Bank, Adoption and investment in AI across the euro area, Occasional Paper No. 395, 2026. The ORBIS accounting benchmark in section 7.2 is this paper's work.
• 14.Dias da Silva, A., Lebastard, L. & Sondermann, D., "AI adoption and the productivity promise: what workers report", ECB Blog, 26 August 2026. The 0.1% to 3.4% estimate range and the euro area figure of around 0.35 percentage points a year in section 7.2 come from here.

Press coverage and adjacent Pebblous pieces

• 15.ppc.land, "AI accounts for nearly a fifth of UK growth in 2026, Bank of England finds", 5 November 2026. One of the two pieces of coverage checked in section 1; it kept cumulative and annual figures apart.
• 16.The Register, Carly Page, "Bank of England pours cold water on AI productivity boom", 6 November 2026. The second piece of coverage on the same box.
• 17.A Recount of METR's Experiment Finds No Ledger for Saved Time — on whether saved time turns into output. Takes the same question as the "AI productivity disconnect" in section 7.1 through developer trial data.
• 18.Torsten Slok's Zero AI Payoff, Absent or Just Invisible? — the same question approached from macro indicators.
• 19.In Bessemer's Healthcare AI Scorecard, What Runs Without a Human Sits Outside the Exam Room — another case of an executive survey becoming a metric.
• 20.AI Risk Disclosure Is Up at UK Listed Companies, and the Incident Line Is Blank — on corporate disclosure in the same country.
• 21.Ramp's Card Data: AI Use Climbs, the Bill Lags Behind — a case where the AI adoption figure turns on what the denominator is.