The biggest open slot is the gateway that lets websites welcome machine customers
Asked (summary):
The current product is weak in a crowded space with serval.com. What else could this team build — what other products — especially given long-horizon agents and today’s strong LLMs?
Asked (follow-up):
“give me 5 more.”
Five new horizontal infrastructure theses — all hypotheses, none validated — each targeting initial contracts above $200K with million-dollar ceilings. They are grounded in 27 curated, source-linked evidence rows spanning Action’s own capability, browser-agent runtime and identity infrastructure, enterprise spreadsheet estates, artifact verification and emerging AI warranties. Sources mix vendor and secondary claims; treat the economics as directional and the warranty market as nascent.
Bet #1 — Machine customer gateway: how a transaction flows
One governed path from a sponsoring human or company to a signed receipt. Hover or tap a stage for what it does.
Browser-agent runtime is already converging on verified identity, persistent sessions, audit logging and human-in-the-loop intervention at the scale of hundreds of concurrent sessions — the raw material of a gateway protocol. cloudrps.com
Microsoft now frames cloud browsers as enterprise infrastructure for fleets of AI agents across thousands of websites, with credentials, observability and replay — evidence the substrate bet (#2) is being contested now. techcommunity.microsoft.com
Spreadsheet estates are enormous — up to 75 million files at a mid-sized financial firm and 8–10 million spreadsheets at a global bank — which is why Spreadsheet Foundry (#3) must sell parallel validation, not one-shot conversion. spreadsheetweb.com
AI insurance is real but tiny: current policy limits run $1–5 million, with continuous production monitoring as an underwriting condition — the warranty exchange (#5) rides an experimental market. ifamagazine.com
Poor data quality reportedly costs organizations an average of $12.9 million a year — the directional wedge for Proof-Carrying Work (#4). kanerika.com
The five bets
Hypotheses, ranked. Price points are proposals, not observed market prices.
#1 — strongest recommendation
Machine customer gateway
What
The Cloudflare/Stripe layer through which sites and enterprise apps transact safely with external AI agents: published machine-readable actions and policies, verified agent identity and sponsor, scoped sessions, signed intents, scoped human confirmation, metering and cryptographic receipts — with a fallback browser where no API exists.
Buyer & price
CTO / CDO / head of ecommerce or platform. $300K–$1M annual minimum plus transaction fees; $2M–$20M for high-volume networks.
Why now
Agent adoption creates a new internet participant — the machine customer — and sites either block bots or expose bespoke APIs.
Moat
Trust protocol, site integrations, signed intent/receipt history, two-sided network.
Initial wedge
B2B SaaS and service portals: retrieve invoices, change subscriptions, schedule service, submit requests.
Fatal flaw
Standards adoption and platform incumbents. If it stays a browser proxy, it is not a large company.
#2
Digital worker workstation cloud
What
A managed operating environment for enterprise digital workers: each agent gets a persistent workstation, corporate identity, scoped credentials, inbox, phone/SMS, files, apps, budget, supervisor, approval policy and replayable audit history. IT hires, pauses or terminates a digital worker like an employee.
Buyer & price
CIO / CTO / head of AI platform. $250K–$750K platform minimum plus $5K–$25K per active worker per year; $1M–$10M at hundreds of workers.
Why now
Enterprises already run fleets of browser and computer-use agents on substrates built for sessions, not workers.
Moat
Owning full persistent workstations, identities, communications, budgets and durable multi-day task state — not browser sessions alone.
Initial wedge
Companies already running 20+ agents across customer operations, finance or IT.
Fatal flaw
Hyperscalers, Browserbase and VDI/RPA incumbents; costly support. Positioned as execution infrastructure, not a security dashboard.
#3
Spreadsheet foundry
What
Discovers the critical spreadsheet estate, observes workbook inputs, formulas, macros and approvals, infers the data model, generates a governed application with tests and lineage, runs workbook and app in parallel, reconciles outputs and gradually retires the sheet.
Buyer & price
CFO / CIO / head of enterprise applications. $500K annual minimum for discovery plus 10–20 conversions; $2M–$10M for large estates.
Why now
Estates are vast and spreadsheet dependence persists even after EPM purchases; frontier agents still fail most realistic spreadsheet workflows, so rigorous parallel validation is the product.
Moat
Workbook behavior graph, generated tests and a growing conversion corpus.
Initial wedge
Finance, pricing, capacity and planning sheets whose failure would delay close or change a decision.
Fatal flaw
Spreadsheets encode undocumented human judgment; the services burden may dominate.
#4
Proof-carrying work
What
A verification and provenance layer: every consequential number, claim, chart and action ships with a machine-checkable proof bundle — source snapshot, query, lineage, assumptions, reviewer, freshness. Plugs into docs, sheets, BI and agent frameworks; independently re-runs calculations and blocks approval when proof is missing. A type system and CI for business work.
AI multiplies artifact volume while verification stays manual and expensive.
Moat
Executable provenance and independent recomputation — not citations generated by the same model.
Initial wedge
Board and investment papers, financial models, pricing approvals — where one wrong number is costly.
Fatal flaw
Customers treat it as a feature of Microsoft/Google/BI suites; it needs an open proof standard and cross-tool coverage.
#5 — ranked last on regulatory and capital difficulty
AI warranty exchange
What
Technical underwriting and warranty administration that lets AI vendors guarantee measurable production outcomes: connect to traces and business results, define a narrow warranty, continuously evaluate, price risk, require controls, and pay service credits or insured claims on breach — with an insurance/reinsurance partner carrying balance-sheet risk.
Buyer & price
AI vendors closing large deals and enterprises buying consequential AI. $250K+ platform minimum plus a share of covered value or premium; million-dollar accounts when warranties cover tens of millions of deployed AI value.
Why now
AI performance guarantees and insurance-backed commitments are emerging, but the market is early and much evidence is secondary — label it experimental.
Moat
Action’s own execution data as an underwriting advantage.
Initial wedge
Warranty browser/computer-use agents on bounded workflows with observable outcomes and financial caps.
Fatal flaw
Regulation, correlated losses, moral hazard, thin actuarial history, carrier dependence. No self-insuring before substantial data and capital.
Qualitative ranking
Author’s scores out of 10 — judgment, not measurement. Darker is stronger.
Evidence rows
All 27 source-linked rows behind the theses. Experimentation rows are incidental carryover from earlier research and back no recommendation here.
Evidence
Source
What it shows
Metric
Method: 27 curated source-linked rows across six evidence types (Action capability 2, browser runtime 5, spreadsheet estate 5, verification 5, AI warranty 5, enterprise experimentation 5 — the last omitted from recommendations as carryover). Rows gathered from public web pages across multiple hosts; no single URL backs more than half. Metrics are vendor or secondary claims quoted as published and are directional; prices on the five bets are hypotheses, not observed contracts. Long findings and metrics are truncated for space. The ranking grid reflects qualitative judgment, not data. 90-day validation plan for the gateway omitted for space: five signed-intent actions, two portal adapters, two agent clients, three $100K+ design partners, 95% completion on 1,000 transactions, two conversions to $300K+ annual.