How AI Becomes Economic Value—and Who Captures It

In the previous note, I focused on the low-hanging fruit: assets that already have distribution, trust, workflow ownership, or transaction rails, with AI as an amplifier. This note looks further out, at opportunities whose infrastructure and business models are still developing. Three directions stand out to me.

A2A Across Organizations: From Matching to Execution

The infrastructure for agents is taking shape. MCP standardizes how agents connect to tools and data, while A2A standardizes how agents communicate with other agents. A2A today is deployed in controlled enterprise environments within organizations. As identity, delegated authority, confidentiality, liability, and settlement develop, cross-organization A2A will be the next step.

Many transactions still begin with static documents such as a job description or an RFP. Each document captures part of the relevant information at one point in time. Agents could maintain current representations of each side's needs, capabilities, and authority, fill information gaps, and negotiate routine terms within human-defined limits.

Over time, they could also coordinate qualification, quoting, scheduling, contracting, and payment.

Agents can evaluate a broader set of potential counterparties at much lower coordination cost. Transactions that once required repeated emails, phone calls, and manual comparisons can move forward faster.

Existing marketplaces start from a strong position because they already have customer relationships, supplier data, transaction history, and working payment and fulfillment rails. AI can interpret requests, identify suitable counterparties, increase conversion, and complete more transactions. The strongest opportunities combine fragmented participants and complex requirements with enough transaction volume to observe outcomes and improve future matching.

From Transactions to Risk Pricing and Insurance

Completed transactions produce verified outcome data: whether work was completed, a supplier delivered on time, equipment remained operational, a borrower repaid, or a claim occurred and at what severity. More completed transactions create more outcome data, improving future matching, qualification, and pricing.

Insurance underwriting becomes possible when operating data can be linked consistently to verified financial losses. Aggregated outcomes establish base rates. Capturing the economics requires data rights, financial-product distribution, intervention, and payment-collection rails.

Stripe Capital provides one example: payment volume and history inform financing eligibility, while repayment runs through the same payment flow. Samsara collects fleet data to improve driver behavior and insurance discounts. Coalition has built a tighter loop across cyber-risk assessment, insurance pricing, continuous monitoring, intervention, and claims.

Specialized robots insurance could be an attractive long-term opportunity. A multi-vendor platform that links machine behavior and human intervention to failures, downtime, repair costs, and financial losses could become the system of record for robot performance, pricing, and risk.

The commercial path could develop in stages:

Observability, fleet control, and recurring service revenue should support the initial investment case. Financing, guarantees, and insurance provide longer-term upside once the platform has established a credible loss curve.

AI4Sci and Recursive Self-Improvement

AI4Sci deserves a separate article. Companies such as Lila Sciences, Isomorphic Labs, Periodic Labs, and Ricursive Intelligence are worth watching as different models for using AI to automate R&D. Isomorphic Labs focuses on AI-first drug design; Lila Sciences and Periodic Labs connect scientific reasoning with autonomous physical experiments; Ricursive uses AI to design and verify the chips that will power future AI systems.

These also point to a recursive improvement loop: automated experiments generate proprietary data that improves scientific models, which then design better experiments. Ricursive applies similar logic to computing, using AI to design better chips that can support more capable AI systems.

A scientific discovery generates revenue only after it has been validated, scaled, and brought to market. The key questions are who owns the resulting intellectual property, who funds commercialization, and how the economics are shared among the AI platform, laboratory, research partner, and manufacturer. Drugs require clinical development and regulatory approval; materials require process scale-up and customer qualification; chips require verification, tape-out, and fabrication.

The long-term value of these companies will depend on both their scientific capabilities and their control over intellectual property, capital, and the path to commercialization.