AI is booming and productivity isn't. A new deep dive from Social Capital — I have read only the public teaser, not the paywalled full report, and everything attributed to it below should be read that way — offers a simpler explanation: the bottleneck is not the model. It is the work itself.
Start with the cost side, because it kills the most popular excuse. Ramp's AI Index, cited in the teaser, tracks what more than 70,000 US businesses spend on AI: a median $12.50 per employee per month. Against an average employee cost of about $8,500 a month, that pays for itself at 0.15% more productivity — roughly three productive minutes a week. The token bill is a rounding error. What costs real money is everything around it: change management and retraining.
The value side is where the paradox lives. AI mostly speeds up production. But most of a corporate employee's week is not production. It is coordination: meetings, alignment, sign-offs, waiting on legal. At P&G, the teaser reports, one person using AI produced proposals as good as a two-person team without it. MIT economists found the sharper form: automation pays when AI steps are grouped so a person checks the work once, not after every step. AI can hand back those three minutes; it cannot decide what happens to them. If they flow into another meeting, the return is zero.
That may also explain why gains are so uneven: AI multiplies output rather than adding to it, and multipliers reward the already-good. A 2026 experiment with 515 startups makes the point brutally. Every firm got the same tools, training, and support; a random half also saw how other firms reorganized work around AI. For the typical startup, revenue barely moved in either group. Almost all gains came from the top 10%, and the reorganized group reached 1.9x the control group's revenue overall. Same tools. Different work design.
What it means for carriers
First, this is the Pilot Trap wearing an economist's clothes. Carriers have spent three years bolting AI tools onto journeys designed for paper, phones, and handoffs. A copilot on a broken submission intake does not fix the intake; it produces faster versions of the same queue. The firms pulling ahead rebuilt the journey, then pointed the model at it. Today's headlines agree from the other direction: Allianz and Hartford executives put data standards first, and an AM Best survey ranked data readiness the single largest barrier to AI deployment. The foundation is the rebuild, not the purchase.
Second, this reframes the BPO conversation entirely. The renewal debate is priced as if AI were a cheaper pair of hands for the same process. It isn't. The return comes from deleting steps, not staffing them differently. Treat AI as a headcount negotiation and you keep missing the number; treat it as a journey redesign and you keep beating it.
Third, the rebuild has a governance shape, and it is the one this series keeps returning to: agents pre-qualify, a licensed person confirms, the ledger keeps the tape. It means moving the human to the one check that matters, once, with a record — the MIT finding as an operating principle. Rebuild intake around that discipline and you get both speed and defense. Automate the old sign-off chain and you get neither.
From the operator's chair
One: delete before you automate. Take your submission intake, or your claims status queue, and run the five steps in order: question every requirement, delete what you can, simplify what is left, speed it up, and only then automate. Most carriers do it backwards — they automate a process they already suspected was unnecessary, then wonder why the ROI deck doesn't close.
Two: measure the coordination tax, not the token bill. Stop reporting what AI costs per employee. Start reporting how many handoffs, approvals, and waiting days sit between a submission arriving and a decision being made. That number is where your productivity went, and it is the only number a rebuild can move.
Three: rebuild one journey end to end before you buy the next tool. Intake is the obvious candidate: it is where the broker's time, the underwriter's time, and quote-cost economics collide. Rebuild it — delete the steps, group the AI work under a single human check, keep the tape — and let that one journey be the proof. The top 10% did not get there with ten pilots. They got there with one rebuild done properly.
AI lowers the cost of doing things. It does nothing about whether they were worth doing. The carriers that learn that distinction will see their productivity numbers finally move. The rest will keep paying $12.50 a month and wondering where the return went.
— Krishna