Office agents are cheaper and faster than juniors, not better

What we checked

• Do Chinese LLM agents reach about 64 percent of junior-human office-deliverable quality on OmegaUse-OfficeVal at a small fraction of Chinese outsourcing cost? confirmed. Baidu's Agent Frontier Team scored GLM-5.2 at 17.91 against a best-of-annotators human baseline of 27.79 on 100 long-horizon office tasks, a 64 percent ratio. LLM inference cost was $0.2152 to $1.7572 per task versus a $6.8560 human price proxy, and four of five models finished far faster than the 2.324-hour human mean. [1][2]
• Is India's IT-services industry shifting from billable hours toward outcome-based contracts under AI pricing pressure? partly confirmed. A Reuters dispatch carried by CNA quotes TCS chief executive K. Krithivasan saying about 80 percent of contracts in finance, HR, and other business services are now tied to outcome measures, and Persistent Systems chief executive Sandeep Kalra saying clients want the same work for 25 to 30 percent less. LiveMint's TCS earnings report independently describes outcome-based and output-commitment AI deals with a 10 to 15 percent productivity-savings rule of thumb, so the shift is real but not shown to cover the whole $315 billion sector. [6][7][8]
• Do US workers now delegate whole tasks to AI and rarely review the output? partly confirmed. An Epoch AI/Ipsos KnowledgePanel survey of employed US adults fielded 10–19 July 2026 finds that 20 percent say AI now handles at least one task previously given to a coworker or contractor, and that 66 percent of AI outputs are used unchanged or with only minor edits versus 5 percent heavily reworked. The same survey says AI rarely does most or all of a task: full or nearly full reallocation tops out at 10 percent for software and systems design and stays below 7 percent on other tasks. [3][4][5]
• Does enterprise AI token spend follow a power law inside companies, with a small cadre driving most cost? partly confirmed. Rippling reports that roughly 10–15 percent of employees drove about 60 percent of AI spend, with one engineer at $50,000 a month, on a trajectory toward 40 percent of the R&D headcount budget before controls cut that forecast to about 10 to 15 percent. That is extreme concentration in one firm, not a fitted power-law across companies. [9][10]
• Does OmegaUse-OfficeVal support Nielsen's 90-percent-by-end-2026 timetable and a measured $1-versus-$100 US labor comparison? refuted. The paper is a single snapshot of five Chinese models under one programmatic scaffold; it contains no time series and no US wage measurement. The $6.86 human cost is a China outsourcing price proxy, and Nielsen's loaded US $100 figure and 90/100 percent forecasts are commentary, not results. METR's roughly seven-month (and recently faster) doubling applies to software-task time horizons, a different metric and task distribution. [1][11][12][13]
• On this office-deliverable benchmark, do coding actions beat GUI-only computer use? confirmed. With Kimi K2.6, a coding-only CoAct variant scored 14.41 on OmegaUse-OfficeVal versus 0.78 for computer-use-only; hybrid sat in between at 13.77. The ranking reversed on OSWorld office-related tasks, where hybrid led at 77.58, so file-level code wins here because scoring judges the artifact, not the GUI trajectory. [1][2]

What we found

• OmegaUse-OfficeVal is a fully released Baidu benchmark of 100 practitioner-derived office-suite tasks (documents, spreadsheets, slides, PDFs) that take 2.32 hours of junior human labor on average. Scoring uses 219 usability checks and 2,009 task-completion rubric items compiled into code verifiers: an unopenable or uneditable file scores zero, and collateral damage is penalized. The human baseline of 27.79 is itself far from perfect—humans zero out 29 percent of tasks and score above 50 on only 21 percent—so the 64 percent GLM-to-human ratio is a ratio of two incomplete sums, not 64 percent of a finished job. [1][2]
• Cheap and fast are the clear wins; quality is not. Qwen3.7-Plus costs $0.2152 per task and DeepSeek-V4-Pro averages 0.184 hours, versus $6.8560 and 2.324 hours for humans. Weighting by price or human time changes the winner to Qwen3.7-Plus, so “best model” depends on whether you count tasks or dollars. Two of five models still zero out about half the set. Main-board evaluation used a programmatic scaffold with no GUI computer-use, LibreOffice in Docker, and a 14,400-second timeout, so the numbers describe file-level agents, not Microsoft 365 click-through. [1][2]
• US workers are reallocating some tasks, not whole jobs, and they under-edit. Epoch AI and Ipsos surveyed 1,103 employed adults after quality control (1,747 adults screened to 1,106 employed; Ipsos reports ±3.0 points at 95 percent for the full adult sample). AI use among people who do each task runs from 25 percent (record keeping) to 57 percent (designing systems and software). Time savings are reported on 53 percent of tasks where AI does most or all of the work versus 37 percent when it only assists, and about one in six AI-assisted tasks take more time. Editing is not a quality score: 66 percent of outputs ship with at most minor edits. [3][4][5]
• Indian IT pricing is moving off the timesheet in named segments, with clients demanding the same work for less and some work moving in-house. CNA's Reuters dispatch puts the industry at $315 billion and records TCS's 80 percent outcome-based share in business services plus Persistent's 25–30 percent price ask. LiveMint quotes TCS leadership on outcome and fixed-capacity AI deals and a 10–15 percent productivity pass-through; The Economic Times quotes Kalra on 25–30 percent effort deflation. That is sector evidence of unbundling hours, not proof that every contract is now an SLA. [6][7][8]
• Rippling's internal X-ray is the clearest public view of concentrated token spend: 80 percent month-over-month growth, a 40 percent-of-R&D-headcount trajectory, 10–15 percent of employees driving about 60 percent of spend, and a $50,000 monthly individual. After caps, an AI gateway, and spend dashboards, the company says the R&D token share fell to about 10–15 percent while July token volume (about 600 billion) nearly matched the peak even as July cost was 37 percent of April's. TechCrunch independently reports the same shock number and the routing explanation. Defaults to the newest frontier model, not measured productivity, drove the blowout. [9][10]

What it means

Treat office agents as cheap junior contractors that fail silently. Ship review, handoff, and recovery surfaces before generation chrome: status while the agent works, diffs against the last good file, reversible retries, and a path to a human. Score models on value-weighted deliverable quality and repair burden, not prompt latency, and prefer file-level coding actions for long document workflows while keeping GUI only where layout must be seen. Price and instrument completed outcomes—openable files, rubric pass rates, rework in review—because clients in IT services are already buying that, and because two-thirds of workers will not edit unless checking is cheaper than trusting.

Do not staff or forecast from the 64 percent headline, the $1-versus-$100 contrast, or a 90-percent-by-December curve. Humans on this set average well under one-third of rubric points; Western frontier models have not been reported here; METR's doubling time is a software-horizon statistic, not an office-quality speedometer. Inside the enterprise, assume spend is skewed, put token meters next to output metrics, and route off expensive defaults. Power users are worth studying for tomorrow's workflows, but they are not the median employee and they will burn the budget if unconstrained.

Still open

• No public OmegaUse-OfficeVal numbers were found for US or other non-Chinese frontier models under the same scaffold.
• The benchmark is one snapshot; nobody has published a repeated time series that would support a 2026–2027 quality forecast.
• The roundup's roughly $100 US loaded cost for these tasks is not measured in the paper.
• It is unknown whether the 66 percent light-edit rate reflects adequate quality, low-stakes tasks, or review UIs that are too costly to use.