03 / Agentic Commerce
A toll on every agent transaction.
As software agents begin to buy on our behalf, every purchase needs identity, authorization and settlement. The company that owns that layer holds a toll on the category, not a feature inside it.
What it is
The trust and payment rails for transactions software initiates.
Agent identity and credentials, so a merchant knows which agent this is and on whose behalf it acts. Mandates and authorization, so a person or a company can say what an agent may buy, from whom, at what limit. Settlement, from payment instruments issued to agents to wallets, escrow and micropayments between machines. Verification for the merchant side, so legitimate agents are let through and abusive ones are not. Dispute and liability frameworks for when the agent got it wrong. Machine-readable catalogs, pricing and negotiation, so agents can transact without scraping a web page built for people.
Each of these is a rail, and rails earn a toll. The economics are a take rate on volume or a fee per transaction, with the moat in the fact that both sides of the transaction integrate once and then do not want to integrate again.
The assistant that does the shopping is the model's job. The rail it has to clear to pay is a company.
Why now
The transactions have started and the rails are being chosen.
Agents already book, procure, pay bills, buy software and replenish inventory, and the volume is early enough that the rails are still being decided. The card networks and the largest platforms have each published agent-payment programs and protocols, which settles the question of whether this category exists and opens the question of who owns the layers around and between them: identity, mandates, risk, and the settlement paths the networks do not cover.
This is the layer where the wrapper trap is most visible. A shopping assistant, a checkout button or a plugin for one platform is a feature the model or the platform will ship. Identity, authorization, risk and settlement that both sides rely on is infrastructure, and infrastructure is what we fund.
Where value accrues
The mandate record, the risk data, and standing on both sides.
The mandate and authorization record
What the agent was allowed to do, and proof it stayed inside it. This is the system of record for agent spend and the evidence in every dispute. Whoever holds it decides liability.
Risk data on agent behavior
Which agents are legitimate, which merchants are safe, what abuse looks like. This accumulates only by processing volume, and it is what a competitor with the same model cannot buy.
Standing with the networks and the platforms
Being the credential a merchant trusts and the rail a platform routes through is a distribution position that does not transfer when the model improves.
Both sides integrated
A rail that only the buyer's agent uses is a feature. A rail that merchants, platforms and agents all integrate is a network, and networks are replaced rarely.
The fork
Stronger, or erased.
Sits at authorization and settlement, accumulates risk data with every transaction, and is embedded with merchants and platforms as well as agents. More capable agents mean more transactions clearing its rail.
A shopping assistant, a checkout wrapper, a single-platform plugin, or a feature the card networks' own protocol covers. More capable agents do the same job without it.
What we ask
The questions the memo has to answer.
Which side pays, and why do both sides integrate?
A one-sided rail is a feature. We want the reason the merchant and the agent both need it.
What happens when the networks' own protocols cover the use case?
The right answer names the layer the protocol does not reach and shows the company already there.
Does the take rate survive competition?
Tolls trend to zero unless the rail holds something the next rail cannot: the record, the risk data, the standing.
Who carries liability when the agent buys the wrong thing?
And does the company's record decide the dispute. If it does, it is infrastructure.
What is the risk model built on, and who else has that data?
Proprietary by construction, or licensable by anyone.
Which regulation applies?
Money transmission, know-your-customer for agents, consumer protection and payment network rules. Some of these are moats once cleared, and we want the team that has mapped them.
The bar by stage
What good looks like from first cheque to Series A.
Pre-seed
A team with payments or identity infrastructure in its history and a precise view of which rail is missing. A first integration on each side of the transaction, even at tiny volume.
Seed
Live transactions clearing the rail with both sides integrated, a risk model learning from them, and a first standing with a network or a platform. Unit economics that make sense at a take rate the market will bear.
Series A
Volume growing with the number of agents rather than with sales effort, annual recurring revenue in the low millions and growing, net revenue retention above roughly 110 to 120 percent, and a record that has already decided a dispute.