The Last 20 Percent
Aadik@mostlyharmless
+5.7%
since Oct 7
The idea
My bet is that as AI makes the first 80% of building something cheap, the value shifts to whoever owns the hard last mile: deep domain expertise, proprietary data, and the customer relationships the real world already trusts. So instead of chasing the data-center buildout, I'm leaning into vertical software with entrenched workflows, companies whose know-how is baked into hardware and physical networks, owners of proprietary judgment datasets, and the distribution channels that will carry AI into everyday businesses.
What it owns
- Domain knowledge locked into hardware, physical networks and relationships over decades creates the hardest last mile to replicate and stays furthest from commoditized software. · Intuitive SurgicalSurgeon training, hospital relationships and hardware expertise accumulated over decades create an installed base and market position that new robot entrants cannot quickly displace., Axon EnterpriseIt owns relationships with nearly every US police department plus the body-camera footage dataset no competitor possesses, making it both a proprietary data owner and the distribution channel for public safety AI., Deere & CompanyIts dealer network and decades of farm data represent the only realistic way AI reaches farmers in the real economy, turning heavy equipment into one of the largest AI distribution channels., TrimbleCentimeter-level positioning technology plus construction and agriculture workflows embed domain expertise into hardware and software that contractors and farmers have built their operations around., HEICOCertified aerospace parts knowledge accumulated over decades creates expertise that a model cannot shortcut because regulators and airlines demand the real certification and approval process., CopartPhysical auction yards and a two-sided buyer network represent a marketplace no software startup can rebuild, and its damage imagery and data becomes more valuable as insurers automate claims processing with AI.30%
- Industry-specific workflows and data accumulated over decades position these companies as the natural channels through which AI reaches their niches, even as markets price them as AI casualties. · Veeva SystemsIt has spent two decades embedding itself into how pharma companies run clinical trials and sales, creating FDA workflow knowledge that AI can draft but cannot replicate without Veeva's relationships and data., Tyler TechnologiesIt sells to courts, counties and city halls, the slowest-diffusing corner of the economy, so any AI reaching those customers flows through Tyler rather than around it., SamsaraIt collects proprietary data streams from trucks, warehouses and job sites that cannot be scraped, creating real-world judgment datasets it sells back to operators as AI products., Guidewire SoftwareIt runs the core claims and policy systems insurers cannot rip out, making it the integration point through which insurers deploy AI rather than a legacy layer AI replaces., Procore TechnologiesIt is the system of record for construction projects, embedding itself into how contractors coordinate work, and the grip it holds on that workflow is the last mile AI must travel through., ToastIt owns the register, payments and customer relationship for restaurants, a sector far from technology vanguards, giving it the distribution channel through which restaurant AI reaches its customers.25%
- Datasets that cannot be scraped from the internet become more valuable as AI models improve, since every new model makes access to proprietary judgment data scarcer. · Thomson ReutersIt owns Westlaw, a proprietary legal corpus that AI models need to generate trustworthy answers in law, converting a traditional publisher into the dataset that legal AI depends on., Fair IsaacIts credit score represents embedded judgment and fifty years of outcomes data plus regulatory acceptance that no model-building exercise can quickly recreate or replace in the lending system., Verisk AnalyticsIt owns decades of insurance claims and risk data the entire industry prices against, a proprietary judgment dataset that AI makes more valuable to query rather than obsolete., S&P GlobalIts ratings and benchmarks represent judgment the financial world has agreed to trust, a proprietary asset that no model can generate and that becomes more valuable as AI increases financial data consumption.20%
- When building technology becomes cheap, the scarce asset becomes the sales channel that already reaches millions of real-world businesses, making distribution the defensible advantage. · AccentureIt sells implementation services to enterprises outside technology hubs, making it the primary channel for AI adoption in the broader economy where most companies operate., CDW CorporationIt is how mid-sized businesses in the long tail actually purchase technology through trusted relationships, and they will buy AI hardware and services the same way., CintasIts driver network visits a million businesses weekly and sells them more each visit, a non-software distribution advantage where the relationship and logistics network compound.15%
- Cash10%