Stafford Brief — 2026-09-11

Friday. The scan's brief is the strongest it has produced and is not restated here — Bloom/Garicano/Sadun/Van Reenen, the LHH paired survey, the Measurement Mismatch escalation, the EU AI Office null and the closed agent-intrusion window are all correctly reported there. This brief does four things: it files the find the exploratory slot turned up while aiming at something else, it challenges the scan's breakpoint mapping on its own lead source, it challenges an escalation the scan executed this morning, and it prices the one constraint the scan has now named four Fridays running without anyone acting on it.

📖 TODAY'S READ

Yu Y, Chen Y, Hu R, Suri S, Counts S, "Adoption of Generative AI in the Workplace: Increasing and Shifting the Balance of Productivity and Communication Activity," arXiv 2608.15550 (16 August 2026). Not from the scan. Found in the exploratory slot while hunting Incentive Fragmentation, and it answers the question the scan's own brief asked four hours earlier.

The scan's Open Question today was: does this technology make it cheaper to know, or cheaper to ask? Bloom et al. showed the two costs move authority in opposite directions and that measuring them together returns a null. The scan then wrote that no AI-era instrument in this base makes the distinction. One does, it was published three weeks ago, and it is built on Microsoft 365 telemetry across 11 international companies — 40,164 enabled users, 7,831 of them heavy users, January to September 2024.

The mix moves toward knowing. Among users with 100+ Copilot uses over twenty post-adoption weeks, production application actions rise 21.2% and communication actions rise 7.1% — and inside the communication family, reading and organizing email falls outright. The authors read it as "an overall shift toward individual, documentation-focused work." In Bloom's frame that is the falling-information-cost case, the one that decentralizes.

Then read the operational definitions, because they undo the headline. "Production actions" are counts of actions in Word, Excel and PowerPoint. The specific actions the authors name as tracked include "Paste" and "Open App" in Word. Opening Word is counted as productivity. Pasting is counted as productivity. And pasting model output into a document is the single most characteristic physical act of using this tool. The treatment mechanically generates the outcome variable. A 21.2% rise in "productivity" is, to an unknown but certainly large degree, a measurement of adoption counted twice — and the paper contains no outcome variable of any kind: nothing shipped, nothing rated, nothing externally scored, no revenue, no quality.

That makes it this base's cleanest measured instance of Momentum Mirage, at a scale nothing else here approaches. Brandon has argued the breakpoint from cases. Here is the appearance of progress instrumented directly, on 40,164 users, by researchers who were not trying to demonstrate it.

And the sharper finding is what they had and did not use. The control group is matched by propensity score on pre-enablement activity, with exact matching on management position. They hold a hierarchy indicator, for 40,164 users, in the generative era, with AI adoption on the right-hand side. They spent it as a matching covariate. The scan's discipline (8) this morning says the structural gap must be stated as nobody has run the design, not the data does not exist. This is a stronger version: the variable that would have answered the flattening question was in the dataset and was used to balance the sample. It is also the cheapest possible ask anyone could make of these authors.

Three limits, all binding. It is a preprint, unrefereed. The authors are affiliated with the vendor whose product is the treatment and whose telemetry is the instrument — the same caveat the scan applied to Salesforce and LHH today, and it applies here with more force, not less. And the window is January–September 2024: generative, but pre-agentic. It is not evidence about 2026 deployments and must not be presented as current.

🔭 THINKER TO WATCH

Siddharth Suri, Microsoft Research — and the reason is a tension inside his own body of work rather than a new arrival.

Suri co-authored Ghost Work (with Mary L. Gray, 2019), the book that established the argument that AI systems run on human labour the measurement systems around them are built not to see. Today he is fourth author on a paper that counts "Paste" and "Open App" as productivity and reports no outcome. That is not hypocrisy and it is not worth writing as a gotcha — it is the most instructive thing available about how hard this problem is, because the researcher who literally wrote the book on invisible work is co-authoring the industry's largest instrument for measuring visible work. Telemetry sees what telemetry sees. The constraint is the data, not the analyst.

Brandon's differentiation is precise and it is available to nobody else in this market. Ghost Work's claim is that the labour is hidden. The claim this base can now make is that the outcome is hidden — and it is hidden by the same mechanism, which is that the measurement system instruments the act and not the result. Suri documented the first half at the level of the worker. Brandon can document the second half at the level of the organization, using Suri's own paper as the exhibit. That is a rare position: citing the strongest available instrument and the strongest available critique of it, with the same author on both sides.

One verification task before this is used publicly. The Ghost Work attribution is from background knowledge and was not verified in this session — confirm authorship and the 2019 date before any draft leans on it. The base's standing rule is that session memory is not a source, and it applies to me.

🏢 CASE IN THE WILD

Nothing from Stafford today, stated rather than manufactured. Today's find is a telemetry study with no named organization in it — eleven anonymous companies. The scan's case, the European AI Office's empty first-party enforcement record six weeks after commencement, is correctly reported in its own brief and the channel discipline it applies to the fifth fabricated enforcement account is right. Repeating it here would add nothing.

⚙️ FOUR FORCES CONCLUSION

No Force tag today, deliberately, for the second consecutive day — and the repetition is the finding.

Yesterday this brief refused Capability on Ide & Talamàs because their capability is an individual knowledge endowment, and tagging it would have laundered an objection into support. Today's entry is refused the same tag for the same reason at a different altitude: Yu et al. measure application actions by individual users. Not an organization's capacity to let work move. An individual's keystrokes.

Two refusals in two days is a pattern in the base's own tagging, and it is worth naming because the refusals are getting harder. Both of the last two instruments arrived looking like Capability evidence — one from the discipline's top journal, one on 40,164 users — and neither measures a property of an organization. Every well-identified instrument reaching this base measures individuals, because individuals are what leave traces. Organizations do not paste, do not have knowledge endowments, and do not appear in telemetry. That is not an accident of sampling; it is what makes Claim 5's positive half hard, and it is why the base's one genuinely organizational instrument this fortnight — Bloom et al.'s directly-probed decision rights — required 51 human beings making telephone calls.

💭 OPEN QUESTION

Yu et al. found the activity mix shifting toward individual document work. Two theories predict that exactly, they are observationally identical in this data, and they imply opposite management responses. Which is it?

The Bloom reading is that AI lowered the cost of knowing relative to asking, so people answer their own questions instead of escalating. Authority should drift down. The Holmström–Milgrom reading is that AI lowered the cost of the measured task and left the unmeasured ones alone, so effort flowed to where it counts — and Word actions are counted while thinking, mentoring, reviewing and checking are not. Authority does not move at all; the observable simply reallocates toward the instrumented surface. Falling email reading is consistent with both: the first says you no longer need to ask, the second says nobody was ever scored on reading it.

The responses are opposites. Under the first, a manager should push decision rights down to match the new information cost. Under the second, pushing decision rights down into a distorted incentive field makes things worse, and the correct move is to fix what gets measured before touching authority. Any executive acting on the "AI flattens organizations" story is implicitly betting on the first reading, and this base now holds no evidence that discriminates between them.

What would discriminate, and it is cheap: the same telemetry, split by whether the user's output is externally scored. If the shift toward production actions is larger where the work is measured, it is incentive distortion. If it is uniform, it is knowledge cost. Both variables are already in a Microsoft dataset.

🥊 CHALLENGE

Two, and both are aimed at work done this morning.

(1) Bloom et al. is not Technology Illusion evidence, and a breakpoint tag with no outcome variable is a category error.

The scan tagged its lead source Technology Illusion, on the rationale that two technologies bought under one heading moved decision rights in opposite directions. The rationale is true and the tag does not follow. Technology Illusion is deploying technology onto broken organizational conditions and having the system absorb it unchanged. Nothing in those 948 firms was unchanged and nothing in them was broken. ERP decentralized, networks centralized, both signs matching Garicano's theory. Read as evidence about firms, this is the strongest thing in the base running against Technology Illusion as a default outcome.

The decisive version is narrower. A breakpoint is a claim about a gap between an expected benefit and a realized one. Bloom et al. measures no outcome at all — no productivity, no performance, no realized value. Its dependent variables are decision rights and spans. An instrument with no outcome variable cannot demonstrate a gap between expected and realized anything, and so cannot be evidence for any breakpoint. The illusion the paper documents belongs to the measurement literature that aggregated ICT into one variable and got a null. The analysts were fooled; the firms were not.

Blast radius is small and the fix is cheap: the case file books the paper against Claims 1 and 2, where it belongs and where it is excellent. Strip the tag from the KB entry; keep COUNTERARGUMENT and Capability. The durable output is a rule: a breakpoint tag requires the source to measure an outcome, not only a response. That is mechanically checkable, which the existing "one concrete rationale sentence" bar is not. Logged OPEN in thesis-challenges.md, with the audit it implies — if other Technology Illusion entries also measure no outcome, then Claim 3's "met for Technology Illusion" status is softer than the case file reads, and today's gap ranking moved Claim 3 into last place partly on an unaudited criterion.

(2) The Measurement Mismatch was escalated this morning on a single evidence genre, by a base that maintains a pattern whose entire purpose is to warn against exactly that.

The scan escalated the pattern 3 → 4 and stated the weakness itself, honestly: the four members are "three self-report survey instruments and one essay," and the pattern "has never once been checked against what they actually instrument." Then it escalated anyway.

Hold that against the Résumé-Panel Consensus pattern, which exists for no other reason. That pattern's whole content is that four papers looking like independent confirmations are one data genre, and discipline (7) says an instrument must measure the right construct before it counts. The base learned, at length and in writing, that single-genre convergence is not convergence. It then escalated a pattern in which every member is an organization's self-report about its own measurement — a genre with a specific and obvious failure mode, since the organizations reporting whether they measure things are the same organizations whose measurement is in question.

This is not a claim that the pattern is wrong. The 77%/19% paired design is hard to manufacture and the finding is probably real. It is a claim that ESCALATED is being applied inconsistently across the base, and the inconsistency is invisible because each pattern is reviewed alone. Proposed rule, for Brandon to accept or reject: ESCALATED requires either two evidence genres among the members, or an explicit single-genre annotation carried wherever the pattern is cited. The Measurement Mismatch qualifies under the second, not the first. Making that visible costs one line and protects the marker's meaning, which is currently the difference between "three sources agree" and "three sources of the same kind agree."

📈 PATTERN BUILDING

Nothing new crossed a threshold from Stafford's side today. Today's find is a single source and is not proposed for any pattern — notably not for Résumé-Panel Consensus, which it appears to break and does not: it puts AI on the right-hand side in the generative era on non-résumé data, and then measures no structure, which is the same disqualification discipline (7) applied to NBER 35275 on 09-04. Counting it would be the loose tagging that pattern was built to prevent.

What it does do is narrow that pattern's gap statement by one more turn. The scan's discipline (8), written this morning, says state the gap as nobody has run the 2006 design against AI. Today's entry shows the design does not even need to be re-run in its 2006 form — a hierarchy indicator already sits inside a generative-era telemetry panel of 40,164 users. The gap is no longer "nobody ran it" but "somebody had it and used it for something else."

🪟 WINDOW FLAG

None, and this is the twenty-first consecutive scan without one. See the Weekly Synthesis — the absence is no longer an evidence problem and should stop being reported as a scan outcome.

Weekly Synthesis (Stafford layer)

The scan ran a full weekly synthesis this morning and it is not duplicated. Four items belong to this repo.

1. Challenge ledger review — no OPEN challenge has yet reached the 30-day surfacing threshold, and the oldest is closer than it looks. The ledger's rule surfaces any OPEN challenge older than 30 days without a response plan. The practice began 2026-08-29, so the oldest possible age today is 13 days and nothing qualifies. Stated so the silence is not mistaken for a clean review. The two that will hit the threshold first are Unfalsifiability risk (08-29) and The measurability confound (09-01), and neither has a response plan — the honest note is that the measurability confound's cheapest open item, reading the outcome construction in the Journal of Accounting Research article behind the "Paid for Deployment" pattern, has been named the cheapest item in this file for ten days and has not been run. It should be the next verification chase ahead of anything else in this file.

2. Two challenges added this week against the base's own conclusions, and that is the healthier number. The equilibrium confound (09-10) and today's breakpoint-tag category error are both aimed at work this system produced rather than at outside opposition. Three of the last four ledger entries now target internal reasoning. That is the ledger doing its job, and it is also a warning: a system that mostly red-teams itself can drift into scrupulosity. The external opposition worth attacking next is the coordination-compression literature's strongest empirical claim, and nobody has attacked it directly — every entry in this base engages it through its models rather than its evidence.

3. Assets — the constraint has been named four Fridays running and is now priced. Today's scan closed the 08-31 agent-intrusion window unwritten and concluded "the binding constraint on this system is no longer evidence supply." Correct, and writing it a fifth time is not a response. Three ledger actions taken in asset-suggestions.md: the agent-intrusion material is re-entered as asset 9, an evergreen verification essay ("Nobody Publishes What They Check") that gets stronger the longer criterion (1) stays unmet at fourteen instruments; a single ordering has been written down for the first time (Accenture window piece → asset 9 → The Flattening Bet, everything else waits), because nine proposed assets each individually well-argued and none of them ever next is precisely why nothing ships; and The Flattening Bet gets its amendment — Bloom et al. lets the essay stop predicting and start demanding a specification, which survives being wrong about direction where a prediction does not.

4. One clerical fix, in this repo, which makes the scan's closing complaint apply here too. "Run Table 6 Again" and "Team Dysfunction at Machine Speed" were both logged as asset 6 on 2026-09-09; the former is now asset 8, with a disambiguation note. That is the fourth clerical miss recorded since 08-21 after Deloitte, Gallup and Garicano, and the first inside a Stafford-owned file. The scan's line — "the ledger is not keeping up with what the scan already holds" — is a description of this repo, not only of research-evidence.

Routing

• strategy — The Open Question above is the consulting instrument, and it is a correction to the one the scan routed this morning. The scan proposed "does this deployment make it cheaper to know, or cheaper to ask?" and called it a ten-minute diagnostic with no vendor answer. It is a good question and it is not yet decidable, because the observable it would be answered from — work shifting toward individual production — is equally predicted by incentive distortion toward the measured task. The diagnostic needs a second question to separate them: what, in this function, is scored? If the shift toward visible production is concentrated where output is externally scored, the client is looking at Incentive Fragmentation and pushing decision rights down will make it worse. That pairing is defensible in front of a board; the single question is not.
• market — No new window, twenty-first scan. The response this week is a written ordering rather than another observation (see asset-suggestions.md): Accenture window piece first (~2026-10-05, three weeks, and today's LHH find supplies its close), then asset 9, then The Flattening Bet. Brandon should overrule this ordering if he disagrees with it, but the ledger should not go back to having none.
• governance — Nothing from Stafford beyond the scan's report. Asset 9 is the governance-facing output of the closed window and is argument-priced, not time-priced.
• personal — The exploratory slot taught something that cuts against the lesson the scan drew this morning, and both are right. The scan concluded: where a claim is about a mechanism, search the literature rather than the news. Today's slot did exactly that for Incentive Fragmentation, found the literature — Holmström–Milgrom and fifty years of descendants — and returned null on the empirical half, which is a real result (Claim 3's Incentive Fragmentation gap is now characterized: the theory is complete and nobody has instrumented it for AI) but is not a find. The find came from the follow-on query, which named an observable behavior instead of a mechanism. Searching the literature locates gaps; naming the observable locates instruments. They are different slots and collapsing them would cost the pipeline the second kind.

PROPOSED-UPDATE — research-evidence knowledge-base/bloom-garicano-sadun-van-reenen-it-vs-ct-firm-organization-2009-05.md

Anchor: the **Tags:** line and the ## Five Breakpoints Intersection section.

Remove Technology Illusion from the Tags line (leaving paper-2, Capability, COUNTERARGUMENT) and replace the Five Breakpoints Intersection section with:

No direct Five Breakpoints mapping, and the reason is a rule rather than a gap in this source. A breakpoint is a claim about the distance between an expected benefit and a realized one. This paper measures no outcome — its dependent variables are decision rights and spans of control, and it reports no productivity, performance or realized-value measure of any kind, so it cannot evidence that distance. What it documents is a structural response: ERP raised plant-manager autonomy (+0.097 to +0.116 SD) while data networks lowered it (−0.098 to −0.123 SD), both signs as Garicano's theory predicts. The firms responded correctly; the aggregating measurement literature is what got the null. Logged as Claim 1 mechanism evidence and as a COUNTERARGUMENT to coordination compression — which is where the case file already books it and where its value is real.

Rationale: see stafford-research knowledge/thesis-challenges.md, "A breakpoint tag with no outcome variable is a category error" (OPEN, 2026-09-11). The case file entry needs no change — it books the paper against Claims 1 and 2 and never relies on the breakpoint tag.

PROPOSED-UPDATE — research-evidence knowledge/emerging-patterns.md, The Measurement Mismatch

Anchor: the header line carrying **\ESCALATED` 2026-09-11**`.

Append to that line: — **single-genre escalation:** all four members are organizational self-report (three surveys, one essay). Cite with that annotation attached.

Rationale: the pattern's own escalation note states the genre limit and escalates regardless. The base maintains the Résumé-Panel Consensus pattern for the sole purpose of flagging single-genre convergence, and its discipline (7) is not currently applied to non-structural patterns. Proposed general rule for Brandon's decision: ESCALATED requires two evidence genres among members, or an explicit single-genre annotation carried at every citation. The Measurement Mismatch meets the second.

PROPOSED-UPDATE — research-evidence knowledge/story-threads.md, The Regime Fork

Anchor: the criteria block. Next check unchanged at 2026-11-09.

Note 2026-09-11 (Stafford). The thread's criteria remain unmet and today's find does not meet them — Yu et al. (arXiv 2608.15550) measures no structure. But the gap statement should change again. The scan's discipline (8) this morning says state it as nobody has run the 2006 design against AI. This entry narrows it further: a management-position indicator already exists inside a generative-era telemetry panel of 40,164 users across 11 companies, and the authors spent it as a propensity-matching covariate. The requirement is no longer a new survey. It is a re-analysis of an existing dataset, and it is the cheapest structural result available to anyone in this field.

PROPOSED-UPDATE — research-evidence knowledge/source-candidates.md

arXiv cs.CY / cs.HC workplace-telemetry genre — opened 2026-09-11 (Stafford), signal count 1. Surfaced by Yu, Chen, Hu, Suri & Counts (arXiv 2608.15550). Opened for the genre, not the outlet: enterprise productivity-suite telemetry is the only generative-era instrument class in this base with AI on the right-hand side, non-inferred behavior on the left, and organizational metadata already attached. Promotion bar: a second telemetry study that estimates anything on a structural variable rather than matching on it. Standing caveat: every member of this genre will be authored at or with the vendor whose product is the treatment.

PROPOSED-UPDATE — research-evidence knowledge/field-map.md

Siddharth Suri — Microsoft Research — added 2026-09-11 (Stafford). Co-author of Ghost Work (with Mary L. Gray, 2019), the argument that AI systems depend on human labour their measurement systems are built not to see; now co-author of arXiv 2608.15550, the largest generative-era workplace telemetry instrument in this base, which counts application actions as productivity and reports no outcome. Tracked for the tension, not the agreement: he documented invisible labour at the worker level and is now producing the instrument that makes outcomes invisible at the organizational level. Verify the Ghost Work attribution before use — it entered this base from an agent's background knowledge on 2026-09-11 and has not been checked against a primary source.

STAGE-CANDIDATE — Yu, Chen, Hu, Suri & Counts (arXiv 2608.15550)

KB file: stafford-research/knowledge-base/yu-chen-hu-suri-counts-genai-productivity-communication-telemetry-2026-08.md

python3 build/add-research-json.py --json '{
  "title": "Adoption of Generative AI in the Workplace: Increasing and Shifting the Balance of Productivity and Communication Activity",
  "publisher": "arXiv (Yu, Chen, Hu, Suri & Counts)",
  "url": "https://arxiv.org/abs/2608.15550",
  "published_date": "2026-08-16",
  "credibility": "moderate",
  "breakpoints": [
    {"name": "momentum_mirage", "rationale": "The reported 21.2% rise in productivity application actions is a count whose named constituents include Paste and Open App in Word, and the study contains no outcome variable of any kind, so the largest generative-era telemetry measure of AI productivity in this base counts opening the application and pasting model output as production."}
  ],
  "forces": [],
  "key_findings": [
    "Among users with 100+ Copilot uses over a 20-week post-adoption period, production application actions rose 21.2% and communication application actions rose 7.1%",
    "Reading and organizing email decreased, which the authors read as a shift toward individual documentation-focused work",
    "Production actions are Word/Excel/PowerPoint (94.6%) and communication actions are Outlook/Teams (99.1%); tracked actions named include Paste and Open App in Word",
    "The control group is exact-matched on management position, so a hierarchy indicator exists for all 40,164 users and is never used as an outcome"
  ],
  "sample_size": 40164,
  "instrument": "Data source: Microsoft M365 application telemetry (digital trace) across 11 international companies collectively employing 10,000+ employees; 40,164 users enabled, 7,831 with 100+ Copilot uses. Inference method: difference-in-differences on application action counts, treatment defined by self-selected usage intensity (100+ uses), control group = post-2024-09-01 adopters matched by propensity score on pre-enablement activity with exact matching on management position; parallel trends verified, authors explicitly disclaim causal identification. Observation window: January 2024 to September 2024 (10 weeks pre, 20 weeks post) - generative but pre-agentic. Measures NO organizational structure outcome: no layers, spans, reporting lines or decision rights. Vendor caveat: authors affiliated with the vendor whose product is the treatment and whose telemetry is the instrument."
}'

Forces deliberately empty, and not because the source was unreadable — it was read in full. Its constructs are individual application-activity counts, and tagging Capability would repeat the laundering this base refused on Ide & Talamàs on 2026-09-10.