Three Days Of Paperwork, Done In Two Hours

The Den is a membership social club for families, and like most small operators it loses enormous stretches of time to paperwork that only happens a few times a year. Grant applications, liquor-license renewals and compliance filings each arrive as a blank form plus a thick packet of rules, and someone on staff has to reconcile the two by hand every time. The club moved that work into ChatGPT Work, OpenAI's version of ChatGPT that can read a team's own stored documents rather than starting from a blank chat.

The mechanism is less about writing and more about assembly. The club's prior filings, mission statements and financials become the raw material; the funder's requirements become the template; and the model drafts the submission by matching one against the other. A grant application that used to take three days now takes about two hours, and liquor-license materials that took four days take three.

Across those recurring tasks the club reports freeing up 10 to 15 hours a week, which is the real story for a team opening a second site. The hours do not disappear into a report nobody reads. They go into the work of actually opening the new location.

EvenUp builds software for personal injury law firms, where the bottleneck is a document called a demand package: the letter and exhibits a firm sends an insurer to justify a settlement. Building one means reading hundreds of pages of medical records, bills and incident reports, pulling out the facts that matter, and arranging them into an argument. Done by hand, that is slow, expensive work that a junior lawyer or paralegal does line by line.

EvenUp runs that extraction and assembly on Claude. The case file goes in; the system identifies the injuries, ties each to the supporting record, totals the damages and produces a structured draft a lawyer then reviews. What the model replaces is not the lawyer's judgment but the hours of sorting and transcribing that come before judgment can happen.

The firm reports cutting document drafting from 15 hours to 15 minutes. That number only holds because a human still signs off on the draft, and because the underlying records are complete enough for the model to cite. The saving is real, but it is a saving on the reading and assembly, not on the decision about what a case is worth.

Anthropic released Claude Sonnet 5.5, a straightforward upgrade to its mid-tier model rather than a new flagship. It runs more than 30% faster than the previous Sonnet and costs up to 30% less for most work, while still landing as a clear quality improvement over the version it replaces.

Sonnet is the model most businesses actually run in production, because it sits between the cheap, fast tier and the expensive frontier tier. For the kind of high-volume document work in this issue's case studies, speed and price per task are the whole economics. A third off both means the same monthly agent workload either costs noticeably less or can do noticeably more before the bill changes.

The practical consequence is that anything already built on Sonnet gets cheaper and quicker the moment the model is swapped in, with no change to the surrounding system. For teams weighing whether a high-volume automation pencils out, the math moved in their favor this week.

A federal judge dismissed the antitrust lawsuits brought by Chegg, the education company, and Penske Media against Google over its AI-generated search answers. The core complaint was that Google's AI Overviews summarize a publisher's content directly on the results page, so users get the answer without clicking through, and the traffic that publishers depend on dries up.

The court accepted that AI search has real consequences for sites that rely on Google referrals, but ruled that those consequences are not an antitrust violation. In plain terms, losing traffic because a search engine answers the question itself is a business problem, not an illegal one, and the courts will not treat it as grounds to force a change.

For any business whose customers arrive through Google, that settles a question worth knowing. There is no legal lever here to pull. Plans that assume steady search referrals now have to account for a results page that increasingly answers on its own, because the fix will not come from a courtroom.