Two field papers landed in August 2026, both built on firm telemetry, and read together they describe an office from opposite ends of the ladder. Chatterji, Holtz and colleagues (OpenAI, Columbia and Wharton, arXiv:2608.12236, 12 August) linked 17,446,551 ChatGPT Enterprise messages across 1,764 organisations to job titles. Their sentence: “Among adopters, early-career workers and trainees send roughly eight to nine more weekly messages than the average active user within the same firm, while managers, directors, and executives send fewer messages.” Juniors are 7% of weekly active users and the heaviest senders per user; managers and directors are the largest active block at 24%, and below average.

The other paper reads the prompts themselves. Hallman, Kowaleski and Schmidt (UT Austin) with Puvvada (KPMG LLP, arXiv:2608.27364, 27 August) scored 713,564 prompts from 3,925 back-office staff: mean 40.38 conversations per active employee, median 14.00, and a top decile carrying “51% of conversations.” Sophistication climbs the ladder — “each measure rises with seniority, with staff ranking below managers and managers below employees above the manager level” (staff −0.056, manager +0.023, above-manager +0.036 on the clarity z-score). Keep it honest: the senior seat runs the fewest chat conversations but uses Microsoft Copilot the most — “Senior employees use the tools more frequently, primarily because of greater Copilot use.” It hasn’t left the assistant; it moved to the one inside Outlook and Teams.

Then one untested line, which is where our corridor starts. The KPMG authors float, but do not test, that senior people prompt better because they “face greater accountability for the resulting work product.” The junior sends the most; the partner signs. The review load has moved up, onto the seat the chat telemetry shows prompting least often.

For a 14-person studio this is not “AI is coming for juniors” and not “buy a licence.” It is a question about who reviews whom. A squad wins on its bench — Spain didn’t take the last World Cup on one striker — and a practice ships on its review bench, not on its fastest prompter. The tool spreads fastest where that bench is deepest: OpenAI’s adopters are far larger firms, yet “conditional on adoption, measured use per employee is lower.” Big firms have the reviewers. We have two.

The training caveat is blunt. KPMG’s within-employee result finds a course lifts sophistication the month it happens (+0.051 clarity in the fixed-effects spec, p<0.01) “but not in the months that follow.” The one intervention every office buys does not stick inside the person who took it. Aim that at the measurement, not the person: a message count is activity, not quality, and both teams call their scores “proxies for sophisticated use, not… direct measures of output quality.” Note who paid — Holtz and Tambe were “paid contractors for OpenAI”; Puvvada is “a full time employee of KPMG LLP.” As PAZ has argued before, vendor-owned telemetry reports only what the vendor chose to log, and only the vendor can re-run it.

Atelier: What our desk does with this is one page beside the BEP — not a ban, not a course. It names which document classes the assistant may draft (Erläuterungsbericht, minutes, SIA-112 Phase-32 Bauprojekt text, research notes — never the quantities, never the fire concept), who reviews each class, and the pairing rule: junior prompter, senior reviewer, always. No vendor dashboard has a column for the Teilphase-32 signature; that liability lives in the SIA contract, not the console. The “top decile” in a 14-person office is one or two people — name the role, not the person, file the prompt log with the document, and pin the pairing rule to the coordination model as your Monday move.

←TODAY: the one-page rule and the prompt-log-beside-the-document habit are already runnable on a live project. →3012: by Zurich-3012 the signature that carries liability is a reviewer’s diff against a logged trail, not a cold re-read. Fulcrum: the junior’s volume and the partner’s accountability only reconcile when the log turns the review into a diff instead of a re-read.

Hack: Measure your own concentration and breadth without the vendor’s dashboard. Export the admin-console CSV, then compute the two KPMG numbers — the share your top decile carries, and how evenly use spreads across the eight categories (normalised Shannon entropy: 0 = one use only, 1 = evenly spread).

import math
c = sorted(seat_counts, reverse=True)                    # conversations per seat
top_decile = sum(c[:math.ceil(len(c)/10)]) / sum(c)      # concentration, 0..1
p = [n/sum(cats) for n in cats if n]                     # cats = 8 use-case totals
use_div = -sum(x*math.log(x) for x in p) / math.log(8)   # breadth, 0..1

PAZ Takeaway: The capability that fits this problem is an AI implementation, not a licence — a lightweight local logging wrapper (Python) that files each prompt/response pair and the model name as a sidecar beside the document it produced, with author and reviewer fields. The partner’s review then becomes a diff against the log instead of a cold re-read. PAZ builds that kind of bespoke, office-owned tooling alongside a practice; it is the openBIM instinct — keep every party’s work legible to the next — applied to the assistant.

Don’t buy the deeper bench you can’t afford. Build the page that lets your two reviewers cover it: write the one-pager, wire the log, and make the junior’s speed the office’s asset instead of its exposure.