Take a rambling quarterly operations report with buried findings and numbers that quietly disagree with each other, and use AI to restructure it into a clean report plus a one-page executive summary — without letting the AI smooth over the inconsistencies it was supposed to surface.
Arjun Mehta is an operations analyst at Cascade Logistics Pvt Ltd, a fictional mid-size 3PL running warehouses in three regions. His manager forwarded him the Q1 operations report drafted by a departing colleague, with one line: "Board pack due Thursday. Make this readable. One page on top." The report below is real-world messy: findings buried in paragraph six, numbers repeated in three places with three values, and a structure that follows the writer's memory rather than the reader's needs. The trap of this lab: a language model will happily "clean up" contradictory numbers by picking one and moving on. Your job is to make it flag them instead — and to catch whatever it still misses.
Paste-ready. This is the full text Arjun received.
CASCADE LOGISTICS - Q1 OPERATIONS REVIEW (DRAFT v3 - internal)
It has been a busy quarter for the operations team with many challenges
and also many wins across all our regions and service lines. The team
has worked hard despite attrition and the festival season impact.
Total shipments handled in Q1 were 12,450 across all three regions,
which represents growth over last quarter. West region led as usual.
The new WMS rollout at the Bhiwandi warehouse went live in February,
slightly behind the January target, causing some disruption to
dispatch operations for around two weeks.
On-time delivery performance for the quarter came in at 94 percent
which is above our contractual SLA of 92 percent with most clients.
Damage claims were 31 in the quarter totalling Rs 2.1 lakh, most
relating to the fragile-goods lane for our electronics client.
Region-wise: West handled 6,200 shipments, North handled 3,900, and
South handled 2,850. North faced driver shortages in January due to
which some volumes were diverted to rail. The rail pilot was a
success and management may consider expanding it.
Warehouse utilization stood at 78 percent on average. The Bhiwandi
disruption mentioned above caused a temporary dip. Staff attrition
for the quarter was 14 percent annualized, concentrated in the
loader workforce, and 22 new hires were onboarded.
Cost per shipment for the quarter was Rs 218, calculated on a total
operations cost of Rs 26.4 lakh. This is an improvement on last
quarter's Rs 231. Fuel costs remain the main pressure and the
finance team's hedging proposal is still pending a decision.
It should also be noted that on-time delivery was impacted by the
WMS disruption and came in at 89 percent for February specifically,
though the quarter overall as noted above met the SLA. One client,
Meridian Electronics, has raised a formal concern about the damage
claims and a review meeting is scheduled for April 12. Actually the
overall on-time figure including the February dip works out to 91
percent for the quarter per the dispatch system export.
In summary the team delivered a solid quarter and with the WMS now
stable, Q2 is expected to be stronger. Key asks for management:
approve the rail expansion study, close the fuel hedging decision,
and sanction 2 additional QC staff for the fragile-goods lane.
Read it yourself first — ten minutes, pen in hand. Before any AI touches this, read the report once and write down: the three most important findings, every number that appears more than once, and anything that made you say "wait, what?". This is not busywork. You cannot audit an AI's restructuring of a document you have never read — and the whole failure mode of this lab is delegating the reading, not just the writing.
Run the arithmetic. Three checks, calculator in hand: do the region shipment figures sum to the stated total? Does the stated cost per shipment equal total cost divided by total shipments? Is the on-time percentage internally consistent across the three places it appears? Write down what you find — exact figures, not impressions. Do not fix anything yet.
Ask the AI to audit before it rewrites. Paste the full report with: "Do not rewrite this yet. First, audit it: list every numeric claim, every place the same metric appears with different values, and every internal contradiction. Present as a table: claim, location, conflicts with, severity." Compare its audit against your step-2 notes. If it found issues you missed, verify them against the text. If you found issues it missed, tell it — and note that a model asked to audit finds far more than a model asked to rewrite.
Resolve or flag every discrepancy — your call, not the AI's. For each conflict, decide: is one value verifiably correct from context, or is the truth unknowable from this document? Rule for the rewrite: a resolvable conflict gets the correct value plus a footnote; an unresolvable one gets flagged in the text as "requires reconciliation before circulation — source figures conflict". Never let the report silently present either version of a disputed number. Draft your resolution table in your doc.
Restructure — outline first. Prompt: "Restructure this report for a management audience. Propose an outline only: sections in order of reader importance, one line on what each contains. Findings and asks come first, narrative context last." Review the outline. A good one leads with performance vs SLA and the three management asks — not with "it has been a busy quarter". Adjust it, then approve.
Generate the rewritten report section by section. Feed the approved outline back: "Write the report following this outline exactly. Use only facts from the original. Apply my resolution table for the disputed figures — flag unresolved conflicts exactly as instructed. Neutral, direct business prose. No filler phrases. Bold-lead bullet lists for findings." Generate, then check each section against the original for invented facts — pay special attention to any sentence that sounds plausible but appears nowhere in the source.
Write the one-page executive summary. Prompt: "From the rewritten report, produce a one-page executive summary: 3-sentence overview, 5 key findings as bold-lead bullets with figures, 3 decisions requested with what happens if delayed, and a risks line. It must say explicitly that two metrics required reconciliation." The summary must survive being the only page a board member reads — check that a reader of just this page learns about the disputed figures.
The inversion check, then ship. Read the exec summary against the original report one final time asking a single question per sentence: "Does the source actually say this — with this number, this direction, this confidence?" Summaries invert conclusions more often than they invent facts: "may consider expanding" becoming "will expand", a concern becoming a resolved issue. Fix anything that drifted, save both documents, and note the total time — most people finish under an hour against a half-day manual rewrite.
An AI asked to "clean up" a document treats contradictions as noise and resolves them silently — always run a separate audit pass before the rewrite pass.
Restructuring is safe to delegate; reconciliation is not. Audit first with your own arithmetic, make the AI list conflicts before it rewrites, and ship summaries where every disputed number is flagged — because the exec summary is the only page most decision-makers will ever read.