Published on petlpay.com/blog | Category: Agentic Economy, Agentic Payments |
-
The question is not whether AI agents will do real work on real projects. They already are. At least to some extent, in collaboration with humans, as human-assistants, but increasingly, as fully autonomous contributors to work, each with their own wallets.
The question is: who’s built the payment infrastructure for them?
The answer, right now, is nobody specifically has won. And that gap, between the emergence of AI agents as genuine project contributors and the financial infrastructure that can pay them compliantly and automatically, is one of the most significant unaddressed and interesting future-based problems in fintech.
This post is about what we believe that infrastructure needs to look like, why existing payment tools cannot provide it, and how MCP (Model Context Protocol) is the architectural bridge between AI agents and the payment rails that will serve the agentic economy, which in turn, will serve nearly all sectors, including professional services as well as building and construction.
What AI agents are actually doing on projects today
Let’s start with the concrete reality, not the science fiction.
AI agents in 2026 are being deployed on specific, scoped project tasks: writing and reviewing code, producing content at scale, running data analysis, managing customer interactions, generating design assets, and executing research briefs. In each case, there is a client, a deliverable, a defined scope, and a payment obligation.
For the most part, these agents are tools, invoked by a human, paid for via a SaaS subscription, with no direct payment relationship. But increasingly, agents are being structured as independent contributors, engaged for a specific project, paid on completion, with output-verified compensation.
The structural characteristics of this are identical to how human subcontractors and freelancers are paid:
- Project-based engagement, not a salary
- Output-verified compensation, payment triggered by delivery, not by time
- Compliance considerations (the tax treatment of AI agent earnings is an open regulatory question, but it is arriving)
- Multi-party coordination, an agent may subcontract to other agents, creating nested payment obligations
No existing payment infrastructure was built for this. Payroll assumes a human with a tax ID. Stripe assumes a merchant with a bank account. Wise assumes a person sending money to another person.
The payment rail for agentic work needs to be quite different.
Why the project is the right financial unit
Before getting into MCP, it is worth grounding this in the broader thesis.
The shift from employment to project-based work, which is happening simultaneously across human and AI workforces, requires a corresponding shift in financial infrastructure. The salary was the right financial unit for the employment economy. The project is the right financial unit for the economy that is emerging.
A project has: a defined scope, multiple contributors (human and AI), milestone-based payment triggers, compliance obligations per contributor and jurisdiction, and a reconciliation requirement at close.
No current payment tool manages all of this natively. Accounting software records it after the fact. Payment tools move money point-to-point. Compliance tools handle one jurisdiction at a time.
What is missing is a payment rail that makes the project the native unit, where invoices, approvals, disbursements, compliance, and reconciliation are all anchored to the project rather than to individual bilateral payment relationships.
That is what Petl Pay is built to be. And MCP is how AI agents plug into it.
What is MCP and why does it matter for payments
Model Context Protocol (MCP) is an open standard, developed by Anthropic, that allows AI agents to call external tools and services as part of their reasoning and execution flow. Think of it as an API layer specifically designed for AI agent interactions, structured so that an LLM can discover, invoke, and receive results from external services without requiring a human to manage each interaction.
MCP servers expose capabilities, called tools, that agents can invoke. A tool might be "create an invoice," "check payment status," or "approve a disbursement." The agent calls the tool, passes the required parameters, and receives a structured response. The payment action happens in the real world. The agent continues its reasoning.
This is the architectural primitive that connects AI agents to payment infrastructure.
Petl's MCP server exposes the core payment primitives of the platform as callable tools, for example, they could be:
- create_invoice - generates a compliant invoice for a specified project, contributor, and amount
- submit_for_approval - routes the invoice to the designated approver
- check_payment_status - returns the current status of a payment or invoice
- trigger_settlement - initiates payment on an approved invoice across the appropriate rail
- get_project_balance - returns the current financial state of a project
An AI agent working on a project can call any of these tools as part of its execution flow. No human needs to manage the payment transaction. The agent completes the work, creates the invoice, submits it for approval, and, once approved, settlement is triggered automatically.
How this works in practice: three scenarios
Scenario 1: AI agent as project contributor
A software agency engages an AI coding agent to complete a defined module of a larger project. The engagement is scoped: X hours of code generation and review, deliverable specified, and payment on completion.
When the module is complete, the agent calls Petl's create_invoice tool, passing the project ID, deliverable description, and agreed amount. Petl generates a compliant invoice. The agency project lead receives an approval request, one tap. Payment is triggered. Settlement happens on the agreed rail. Xero is updated.
The human contributor’s involvement is essentially one approval tap. Everything else should be automated.
Scenario 2: AI agent orchestrating a multi-agent project
A more complex scenario: an orchestrator AI agent is managing a project with multiple subagent contributors, one for research, one for writing, and one for design. Each subagent completes its component. The orchestrator verifies outputs and triggers payment to each contributor.
Using Petl's MCP tools, the orchestrator calls create_invoice for each subagent contribution, submit_for_approval for the consolidated project completion, and trigger_settlement once approved. Multi-party disbursement, to multiple agent contributors, happens in a single coordinated flow.
This is not a future scenario. The infrastructure for it exists. The regulatory framework for paying AI agents is still catching up, but the payment primitives are ready.
Scenario 3: Human contractor using an AI agent to manage their invoicing
A subcontractor on a construction project uses an AI assistant, via WhatsApp or Claude, to manage their project admin. They describe the work done to their chat-based assistant, the assistant calls Petl's create_invoice tool, generating a CIS-compliant invoice with the correct deduction calculation. It submits for approval to the GC's dashboard. When approved, the assistant confirms payment status via check_payment_status.
The subcontractor described their work. Their agent handled everything else.
This is the WhatsApp flow that Petl is built around, and MCP is the architectural layer that makes the AI assistant a first-class participant in the payment workflow rather than just a messaging interface.
The compliance question for agentic payments
One question that comes up immediately: how do you handle tax and compliance for AI agent payments?
The honest answer: the regulatory framework is still forming. HMRC has not yet issued clear guidance on the tax treatment of AI agent earnings. The US IRS is similarly behind the curve. This will be resolved, probably within the next 2-3 years, and the resolution will likely treat AI agent earnings similarly to how software licensing or service fees are treated today.
What Petl can do now: structure agentic payments in a way that creates a clean, compliant audit trail that will satisfy whatever regulatory framework emerges. Every payment is documented: who initiated it, what work it was for, which project it belongs to, what amounts were involved, and what rail it settled on. That audit trail is the compliance foundation.
The parallel to CIS in construction is instructive. CIS compliance was mandatory and complex. The infrastructure that encoded it natively, rather than leaving it as a manual process, created the moat. The same dynamic will play out with agentic payment compliance. The platforms that build the compliance layer now will own it when regulation arrives.
Claude + Petl: what this looks like for users today
The most immediate version of this is Claude as the interface for Petl's payment workflows.
A user in Claude.ai, or any Claude-powered application, can interact with Petl's MCP tools directly:
- "Create an invoice for £2,400 for the Bermondsey site groundwork, 3 days labour, CIS-registered subcontractor"
- "What's the payment status on project 247?"
- "Approve all pending invoices under £500 for the Canary Wharf project"
Claude interprets the natural language instruction, calls the appropriate Petl MCP tool, and returns the result. The payment action happens in Petl's infrastructure. The user never leaves their chat interface.
This is the chat-native payment experience that Petl is built around, extended to any LLM interface, not just WhatsApp or Petl's own dashboard.
For agencies already using Claude for project work, this means payment management becomes part of the same workflow as the work itself. No context switching. No separate payment tool to log into. One interface for everything.
The long arc: Petl as the payment rail for human and AI contributors
Zoom out, and the thesis is actually pretty straightforward.
The workforce is becoming hybrid, human contributors and AI agents, working together on project-based engagements, each needing to be paid compliantly and automatically.
Rafiki Works, Petl's sister company, is already a fractional talent marketplace connecting human specialists, engineers, AI automation experts, and GTM talent with businesses that need them. As AI agents become genuine project contributors, the same marketplace infrastructure applies: engagement scoped, work delivered, payment triggered.
One payment rail, for both. Petl as the infrastructure layer for a workforce that is increasingly human and AI in the same project.
That is not a prediction; the components are being built and connected today.
FAQ
What is an MCP server and how does it work?
Model Context Protocol (MCP) is an open standard developed by Anthropic that allows AI agents to call external tools and services. An MCP server exposes capabilities, like creating invoices or triggering payments, that AI agents can invoke as part of their reasoning and execution flow. It is the architectural bridge between AI agents and real-world actions.
Can AI agents actually be paid through Petl today?
Petl's MCP server is live and exposes core payment primitives as callable tools. AI agents that support MCP, including Claude, can invoke these tools to create invoices, submit for approval, and check payment status. Settlement with AI agent contributors is structurally possible; regulatory clarity on the tax treatment of agentic earnings is still forming.
How does Petl's MCP server integrate with Claude?
Petl's MCP server can be connected to any Claude-powered application or to Claude.ai directly. Once connected, users can instruct Claude in natural language to perform payment actions, creating invoices, approving payments, checking project balances, and Claude calls the appropriate Petl tools.
What is the tax treatment of AI agent earnings?
This is an open regulatory question. Current guidance from HMRC and other tax authorities does not yet specifically address AI agent earnings. Petl structures agentic payments to create a clean audit trail that will support compliance with whatever framework emerges. The parallel to CIS compliance in construction, where Petl encodes the rules natively, applies.
How is Petl different from Stripe for agentic payments?
Stripe is a payment processing tool; it moves money from A to B. Petl is a project payment infrastructure and handles the full workflow from work logged to payment settled, including invoice generation, multi-party disbursement, compliance per jurisdiction, and reconciliation. For agentic payments specifically, Petl's MCP integration means AI agents can trigger the full payment workflow, not just a point-to-point transfer.
What rails does Petl use for settlement?
Open banking (UK and EU), local rails (South Africa, LatAm), and USDC stablecoin for any contributor globally. Rail selection is automatic based on jurisdiction and preference.
-
Petl Pay is the payment rail for the agentic economy. MCP-native, compliance-first, built for human and AI contributors. Sign up at petlpay.com.
