The work nobody announced as an AI initiative — underwriting queues, VAT regulation scanning, SKU reconciliation — is where AI is paying back fastest. Not because it’s technically easy, but because it was always the highest-friction, lowest-status work in the building. The glamorous use cases get the press releases. The unglamorous ones pay back in weeks and then quietly disappear from the org chart.
Hiscox - Insurance - Underwriting drops from three days to three minutes
Hiscox built a generative AI underwriting model on Google Cloud’s Gemini, running through its internal AI lab Hailo, and went live with it in December 2024 across its London Market sabotage and terrorism book. Lead open-market quotes that previously took up to three days now take three minutes — the model runs submission to quote automatically, generates a broker email with pricing pre-filled, and routes only the complex or borderline cases to human underwriters. Three-day turnaround used to be a selling point. Now it’s a liability. If you run a commercial book and your quote-to-bind is still measured in days, you’re losing quotes to whoever answers first.
Amazon FinTech - Financial Services - VAT regulatory reading cut from 26 minutes to 2
Amazon’s internal tax team built World Wide Watch, a generative AI scanner that tracks VAT regulation updates across geographies and tax types. Time to gain insights dropped 92% — from 26 minutes to 2 minutes per update, with 8,000-word source documents compressed to 250-word summaries. Tax experts approved 80% of machine-drafted titles, summaries, and impact predictions without revision. The question isn’t whether AI can read regulation anymore. It’s whether your team’s review time is what’s slowing the business down.
K&L Wines - Retail - Fourteen hours of daily SKU matching, gone
K&L Wines was burning 14 hours of human labor every day matching incoming supplier inventory lists to internal database records. Wine names don’t reconcile cleanly — vintages shift, regions get abbreviated, producers get reordered. NineTwoThree built a custom entity-resolution model using feature engineering on grape variety, vintage, region, and producer, plus vector embeddings to measure semantic distance between product names. Reported accuracy hit 99%, and the system paid for itself in weeks through labor cost savings. The interesting wrinkle is that an LLM alone would have hallucinated its way through this — SKU matching needs confidence thresholds, not eloquence. The boring reconciliation work hidden inside every receiving dock is cheap enough to automate now, and it pays back in a quarter. Most specialty retailers have a version of this problem. Not many have looked at it as an AI project yet.
Throwback: Walmart, 1991 — Retail Link and the Supplier Data Commons
In 1991, retail was a guessing game played in the dark. Buyers placed orders based on quarterly reports and gut feel; suppliers shipped into a void and discovered demand weeks later through reorder patterns. Kmart and Sears treated point-of-sale data as proprietary exhaust — something to be hoarded, occasionally summarized, and sold back to vendors as research. Walmart’s contrarian bet was the opposite: that the data was worth more given away than kept. If suppliers could see store-level inventory in real time, they would solve Walmart’s stockout and overstock problems for free, using their own forecasting talent and their own balance sheets.
The infrastructure had been quietly assembled for years. In 1987 Walmart spent $24 million on a private satellite network — the largest of its kind in the US — connecting every store to Bentonville. The 1988 P&G collaboration was the proof of concept: shared daily Pampers data produced a $50 million profitability swing within eight months. Retail Link formally launched in 1991–1992 as an EDI mandate on the top 2,000 suppliers, then migrated to an internet extranet in 1996–1997 that opened the system to 20,000+ vendors with 104 weeks of item-level history queryable on demand. By 1989, before most of this had even scaled, Walmart’s distribution costs were 1.7% of cost of sales against Kmart’s 3.5% and Sears’ 5% — a structural 2x advantage that compounded every quarter into the everyday-low-price flywheel competitors kept trying and failing to copy.
The transformation was invisible while it happened. Walton did not announce a “supply chain revolution” or convene a steering committee on data strategy. He extended a network that already existed to people who already had problems to solve, and let them solve them. The moat was not the satellites or the extranet — those were commodities by 1998. The moat was the decade of supplier behavior trained against Walmart’s data, the forecasting muscles built inside Procter & Gamble and Kraft and Unilever that only worked when pointed at Walmart’s shelves. Hiscox and the insurance carriers now wiring AI into underwriting are running the same play in a different century: the advantage will not be the model, which everyone will eventually have, but the years of claims and broker behavior compounded around it while everyone else was still debating the business case.
The work didn’t disappear this week — the waiting did. Walmart figured out in 1991 that shared data changes who holds the advantage; what’s changed is that the advantage now compounds in seconds, not seasons.