The Pace Picked Up. Did the Process?

This week brought two kinds of news. The pace of AI capability picked up again, and the physical and commercial limits around it became easier to see. For operators, both point to the same question: is the work itself ready for faster tools?

The Week in AI

1. An outside group says Anthropic's progress has roughly doubled in speed.
METR, an independent evaluation nonprofit, reviewed Anthropic's internal data from mid-August to late October and estimated that the company's models are improving on well-specified tasks about twice as fast as before. METR attributes most of the current speedup to AI helping Anthropic's own staff work faster, and it cautions that the gain may be smaller on messy, real-world work. For operators, the useful line is that distinction: tools improve fastest on work that is clearly defined, so how well a process is written down now affects how much a company gets from each new release.
https://www.google.com/url?q=https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleration-assessment/&source=gmail&ust=1794221222096000&sa=E
https://www.google.com/url?q=https://www.wcax.com/2026/11/05/anthropic-is-rapidly-accelerating-ai-development-amid-slowdown-calls-outside-research-group-says/&source=gmail&ust=1794221222096000&sa=E (Associated Press)

2. OpenAI changed how it deals with data center towns after the midterms.
Two days after voters backed limits on data centers, OpenAI announced what it calls a National Community Compact, The New York Times reported. It says it will pay the full cost of its electricity and the grid upgrades it requires, publish power and water use for each site, and stop asking public officials to sign nondisclosure agreements, with narrow exceptions. The pledges cover only sites OpenAI owns, not capacity built by partners. My read: the cost and timeline of AI capacity are becoming public, negotiated items, and that belongs in any multi-year plan that assumes cheap, plentiful compute.
https://www.google.com/url?q=https://gvwire.com/2026/11/05/after-midterm-rebuke-openai-pledges-to-give-data-center-towns-more-say/&source=gmail&ust=1794221222096000&sa=E (The New York Times, republished)

3. OpenAI released a restricted model built for security research.
On November 5, OpenAI published the system card for GPT-6 Cyber, a version of its model for authorized security work such as finding software vulnerabilities. OpenAI rates it "Critical in Cybersecurity" under its own risk framework and limits access to vetted organizations with identity checks. The operating point is practical: if vulnerabilities get found faster, patches arrive faster, and the teams that test and apply updates will feel it first.
https://www.google.com/url?q=https://deploymentsafety.openai.com/gpt-6-cyber&source=gmail&ust=1794221222096000&sa=E

4. Airbnb set a date for its AI travel assistant.
In its third-quarter shareholder letter, Airbnb said the assistant will reach consumers in the first half of 2027, TechCrunch reported. Beyond trip planning, it is meant to handle work after booking, including reservation changes, payment issues, refunds and rebooking. That after-sale work is where service operations are tested, because it depends on clear policies and on knowing who can approve an exception.
https://www.google.com/url?q=https://techcrunch.com/2026/11/05/airbnb-says-its-launching-an-ai-travel-agent-in-the-first-half-of-2027/&source=gmail&ust=1794221222096000&sa=E

5. A major publisher tied part of its revenue drop to AI search.
USA Today Co. reported third-quarter revenue down 7.4 percent to $519.2 million, according to Poynter. CEO Mike Reed pointed to changing search behavior, as readers see AI-generated summaries and click through less often. Any business that depends on search for demand should read this as a planning signal. A channel can shift faster than an annual forecast assumes.
https://www.google.com/url?q=https://www.poynter.org/business-work/2026/usa-today-co-ai-use-advertising-video/&source=gmail&ust=1794221222097000&sa=E

The Analyst Read

Harvard Business Review's November and December issue carries a clear argument for operators. Companies that add AI to single tasks see limited gains. The payoff comes from mapping a workflow from its first step to its last and then reworking the steps around whatever bottleneck caps performance.

The article draws on research covering 515 companies. Firms that were shown examples of how other organizations had redesigned work around AI identified 44 percent more use cases and completed 12 percent more tasks than firms that were not. They were also 18 percent more likely to win paying customers and generated nearly twice as much revenue. The article's advice, in short, is to "redesign surrounding processes rather than simply automating individual tasks."

What holds up: the workflow is the right unit of improvement. That matches what operators see. A faster step in the middle of a slow process mostly creates a longer queue at the next step.

What it underplays, in my view, is ownership. A workflow map shows the steps. It rarely shows who owns the handoff between them, and that is often where work waits. Many bottlenecks turn out to be decisions waiting for someone with the authority to make them.

Maps also go stale. At MUJI North America, I built the company's MUJINET intranet to replace scattered documents. My view since: a shared reference only stays useful when someone owns keeping it current, and a workflow map is no different. Without an owner and a review date, it describes last quarter's operation.

What a leader should do with it: choose one workflow that crosses at least two teams. Map it end to end with the people who do the work. Mark every handoff, and write the name of the person who owns each one. Where no one owns it, fix that before buying anything.

The Operating Note

This week's items share a thread. Capability is improving fastest on well-defined work, while capacity and demand are both less predictable than most plans assume. Neither trend rewards a company that is waiting for the next tool.

My suggestion for this week is modest. Take the workflow you mapped, or one you know well, and measure how long work waits between steps compared with how long it is actively worked on. Waiting time is usually the larger number, and it mostly comes down to unclear ownership at the handoffs. Those are operating decisions, and they pay off with or without new technology.

The question I would put to a leadership team: in your most important cross-team process, who owns the handoff where work waits the longest?

At Bound, we help leadership teams redesign operating models, decision rights and handoffs so new ways of working actually hold.