Anthropic filed its S-1 on October 26, and most of the commentary went where commentary always goes: the financials and the governance. George Serafeim, at Harvard Business School, read further back — twenty pages of customer stories — and found a different pattern. The biggest gains didn't come from doing existing work faster. They came from work nobody was doing at all.
AIG is the case to read. Its Lexington middle-market property business went from reviewing some broker submissions to reviewing all of them. Quoted submissions rose 30%, time-to-quote fell 55%, and submissions bound rose 40%. Accenture took security test coverage from about 10% to over 80%. PayPal started spotting merchant growth signals it had been missing. One caveat, and it matters: these are Anthropic's figures in its own filing, which says its operating metrics have not been independently verified. Treat them as direction, not audit.
What it means for carriers
First, "hours saved" is the wrong metric — and Serafeim says why in one line: you can't save time on work you never did. Hours saved only measures the work you were already doing. It says nothing about the work that never got touched. The metric that matters is coverage: what share of the work that should be done actually gets done. For underwriting, that is the share of submissions that got a real review. The submissions that were never reviewed were never quoted, never bound, never priced. The 40% is not a speed gain. It is revenue that was sitting in the inbox, untouched.
Second, attention is the scarce resource, and every carrier rations it. The underwriter's day is the bottleneck, not the model. The machine's job is to stretch that attention: read everything, hand the underwriter a summary with the evidence attached, and let the human spend their minutes on the decision instead of the data gathering. AIG Assist reads every broker submission and hands underwriters a risk summary with a full audit trail. That last part is the one that makes it defensible. The machine does the reading, the human makes the call, and the record shows both.
Third, this is the middle-market story, and it is worth naming plainly. Lexington is middle-market property — submissions arriving in every format, from every broker, triaged by whoever has time. The carrier that reviews all of them is not faster than the carrier that reviews some. It is a different business. It quotes risks its competitors never saw.
Serafeim maps the payoffs on a 2x2 and says most AI business cases are written for the bottom left; he suspects the big prize sits in the top right. The image was cut off in the version I saw, so I won't pretend to know the axes. But the pattern in the stories is plain enough: the prize is the work that was never staffed, never queued, never done.
From the operator's chair
One: measure coverage, not hours. For every queue that matters — submissions, referrals, claims triage — report the share that got a human-quality review and the share that got none. The unreviewed share is your prize. If you cannot measure it, you do not have a queue. You have a pile.
Two: price the outcome, not the effort. If the work that matters is a submission reviewed, a quote issued, a policy bound, then that is the unit to buy and sell. Hours saved is a vendor's metric. Submissions bound is an operator's.
Three: keep the human on the decision and the record on the file. The machine reads everything. The underwriter decides. The audit trail shows what was read, what was recommended, and who signed. That is what makes coverage defensible — to the regulator, the auditor, and the broker whose submission you declined.
Speed is a feature. Coverage is the business.
— Krishna