What Happens When You Let an AI Agent Near Your Accounts Payable Process
Whenever I talk with finance leaders about AI agents, one question shows up within the first few minutes. What if the agent decides to approve a $5 million invoice?
It is a fair question, and it is the kind that keeps a CFO up at night. The short answer is that it cannot, and the reason has less to do with the AI than with how the system around it is built.
An agent is a user with a job description
When an agent connects to Direct Commerce through our MCP server, we treat it like any other user. It has a role. Its access is scoped to specific actions, such as reading purchase orders, reading invoices, or writing invoices, and nothing beyond that.
We also keep strict parity between what you can do in the application and what an agent can do through MCP. Anything available in one is available in the other, and the same roles and permissions apply to both. There is no back door for AI.
Every change to a workflow needs a human
The bigger fear is usually about workflow changes. If agents can help improve how invoices move through your process, what stops one from quietly rewriting the rules?
Every change to a workflow is guarded by a person's approval. Someone at your business reviews the change and signs off before it takes effect.
That approval can also be how you give AI a little room to help. You might say you are comfortable with AI auto-approving invoices under $25,000 that are five days from their due date. That is a safety net you chose, with limits you set. The AI gets just enough judgment to move a routine invoice over the finish line, and you stay in charge of where the controls sit.
These are the same APIs we have always used to connect to your ERP. What changed is that the workflow on top of them can adapt. And when it changes, we check that the change cannot hurt you, for example by blocking anything your ERP could not process.
What this looks like for a supplier
Here is a concrete example. A supplier connects their own Claude or ChatGPT to our MCP server. They ask which invoices are approved and waiting on payment, across every buyer they sell to. No more logging into separate portals to piece it together.
Then they go a step further. They ask the agent to review open chargebacks and decide which are worth disputing. The agent finds the ones that qualify, asks a clarifying question where it needs one, and files the dispute. If it needs information it does not have, it can ping the supplier in Slack and finish the job once it hears back.
Setup takes about five to ten minutes. You authenticate, describe what you want in plain language, and schedule it.
Before, finding a disputable chargeback meant logging in, reviewing items by hand, and working out what evidence you needed. With an agent, the system points straight at the evidence gap. Your time goes to the judgment calls, such as what proof actually supports this claim, instead of clicking through screens.
AI in the right places
I do not believe the future of AP is AI everywhere. I believe it is AI in the right places, surrounded by deterministic building blocks that keep it inside the lines. That takes time to build and to earn trust, and it will not all happen on day one.
Where we are heading is event-driven. Instead of an agent checking on a schedule, it reacts the moment something happens, like an invoice getting approved, and decides whether to accelerate payment. We are building toward that, and the foundation is already in place.
If you want to see how agents work inside Direct Commerce, we would be glad to walk you through it.