Start here · The thesis

One human and one agent

Why B2B payments fail at the data layer, why orchestration wins the decade, and what would prove me wrong.

By 2030, the B2B AP function becomes one human and one agent. The agent runs settlement, reconciles, and handles the transactions that follow the playbook. The human handles exceptions, approves edge cases, and makes the calls a regulator still wants a person to make. B2B payments fail at the data layer, not the money-movement layer. Convert unstructured supplier data into a deterministic schema and autonomous settlement falls out. The decade ahead belongs to platforms that route across bank rails, not platforms that pick one. The losers stop at workflow and never reach the rail. Orchestration wins. Single-rail middleware does not.
The thesis in 100 words

The argument

01
Where do B2B payments actually fail?
Watch a mid-market AR team close a single payment. The wire lands Monday. The 835 follows an hour later. The paper stub shows up two weeks after that. Three formats, one dollar, zero clean handoffs to the ERP.
That department is not a relic. It is the median. Most companies still use paper checks for B2B payments. Manual invoice processing costs $12.88 to $19.83 per invoice and takes 10 to 30 minutes of human attention. The accounts payable department operates with a workflow that a time traveler from 2005 would recognize immediately.
The failure is not the money movement. B2B payments fail at the data layer. The data lands in incompatible formats and the cost of resolving it scales linearly with humans. Structured data is the prerequisite for autonomous financial execution, and the operators who get the data layer right capture the network effects that follow.
The full essay: Canonicalization Beats Workflow →
02
Why does canonicalization beat workflow?
The standard fix is to buy software that helps humans read the formats faster. Better OCR. Better dashboards. Better exception queues. Cost per line drops. Headcount does not. That is the workflow tax. Teams pay it because incompatible inputs still have to land in one canonical ERP schema, and nobody has built the schema.
Translation is the actual work. Workflow software is what you buy when no one has solved translation. The vendor category that wins the spend is the one that treats the schema as the product, not the workflow. The pattern holds across healthcare claim adjudication, EDI clearinghouse modernization, and B2B remittance reconciliation. Different verticals. Different buyer titles. Same insight underneath.
Convert unstructured supplier data into a deterministic schema and autonomous settlement falls out. The human-in-the-loop disappears for standard transactions. Network liquidity expands. Basis points scale without headcount.
The full essay: Canonicalization Beats Workflow →
03
Who told you to do this?
When the Bardi family of Florence extended credit to wool buyers in England in the 1300s, the seal on their letter of credit was not decorative. It was the authorization. It answered the only question the receiving banker needed answered. Who told you to do this? That question has anchored every payment authorization framework built since.
Every framework in the lineage assumes the answer ends with a person. The Electronic Fund Transfer Act of 1978. The ACH framework NACHA built in 1980. UCC Article 4A. Now AI systems make consequential decisions upstream of payment execution. A human sets parameters on Monday. The agent acts inside those parameters at 2 a.m. on Thursday. The transaction still reaches a settlement layer that was built to recognize credentials, not delegated machine authority.
The industry needs an agent attestation standard. Not a new law. A new authorization primitive that sits beside identity verification in the settlement stack. The system does not need to prove the agent is human. It needs to prove the agent is acting with valid authority. If the execution layer scales before the authorization layer catches up, the market will learn the hard way, one disputed transfer at a time.
The full essay: Authorization Without Identity →
04
Why does service-as-software replace SaaS, and why payments first?
For twenty-five years, SaaS had a limitation we papered over. It sold you the tool, not the work. When you bought an AP automation platform, you got a nicer interface, OCR that could read most documents most of the time, and dashboards that told you, with great precision, exactly how behind you were. What you did not get was the invoices processed.
Service as software inverts the bargain. Traditional SaaS sells access to a tool, and the customer does the work. Service as software sells the completed outcome, and the AI does the work. SaaS sold you a gym membership. Service as software does your pushups for you. When the work itself is performed by AI, the seat is empty, and per-seat pricing does not just look outdated. It looks absurd.
The flip starts in payments because payments has every characteristic that makes AI agency possible and profitable. It is the rare industry where the data is structured, the outcomes are measurable, the rules are explicit, and the cost of doing it wrong is both high and quantifiable. If you were designing a domain for AI agents to conquer first, you would design something that looks exactly like B2B payments.
Chapter one, free: The Last Invoice →
05
What would prove it?
Put together, the argument lands on one prediction. By 2030, the B2B AP function becomes one human and one agent. The agent runs settlement, initiates payment, reconciles, and handles the transactions that follow the playbook. The human handles exceptions, approves edge cases, and makes the calls a regulator still wants a person to make.
The competitive claim rides with it. The losers stop at workflow and never reach the rail. Orchestration wins. Single-rail middleware does not.
Every call behind this argument is dated, falsifiable, and graded in the open. Each one comes from something already published, with the date it shipped. Nothing gets marked Right or Wrong without a public source showing resolution. The 2030 thesis runs the longest. The scoreboard exists so the misses stay on the record next to the hits.
The scoreboard: every call, dated and graded →

What would change my mind

The AP team of 2030. The core call fails if the standard mid-market AP team in 2030 is still a department instead of one person and one agent, with the agent running settlement end to end. the call →
The middleware survivors. The orchestration claim fails if single-rail AP middleware holds its share while multi-rail orchestration platforms fail to take the category. the call →
The pricing model. The service-as-software call fails if back-office buyers keep paying for seats instead of completed outcomes, or if payments does not flip before the rest of the stack. the call →
The authorization gap. The infrastructure call weakens if agent-initiated volume scales and no card network or regulator ships agent-native authorization credentials after the liability gap surfaces in a real dispute. the call →

9 calls on the public scoreboard · 9 open · 0 graded. Track them →

The vocabulary

Agentic payments: B2B payments executed by software that acts, not just assists. The agent decides whether to pay an invoice and executes the payment, operating inside spending limits, approval chains, and escalation rules set by humans.
Orchestration layer: The layer that routes payments across bank rails rather than picking one. Multi-bank orchestration replaces manual reconciliation with a structured data schema that enforces B2B payment logic across regulated environments.
Canonicalization: Collapsing many incompatible input formats onto one canonical schema so that translation becomes infrastructure instead of headcount. The schema is the product. The workflow is the consequence.
Service-as-software: The successor model to SaaS. Traditional SaaS sells access to a tool and the customer does the work. Service as software sells the completed outcome and the AI does the work.
Authorization without identity: The gap that opens when AI agents can initiate payments while the settlement layer still expects a human authorizer. The fix is not proving the agent is human. It is proving the agent is acting with valid authority.
The workflow tax: The recurring cost of buying software that helps humans read incompatible formats faster instead of solving translation. Teams pay it because incompatible inputs still have to land in one canonical schema and nobody has built the schema.

The long-form version

Book · 2024