Everyone Is Building MCP Connectors. That’s Not the Hard Part.

Let’s take a typical accounts payable process. A vendor emails an invoice and forgets the W-9. Someone forwards a past-due notice with no context. A supplier sends instructions to change its bank details. A bill sits in one system while the conversation about it happens in an inbox.

These present challenges because the inputs arrive as a mess, and sorting them out means working across several systems at once. One invoice email might need to be checked against the bill in BILL and the payment record in QuickBooks. No single system holds the whole picture, so today the person doing the work is the integration layer: reading the context, checking the systems, and escalating anything risky.

That translation work is most of the job, and it is the part software has never been able to touch. It is what we are building TenKey to do: a vendor-neutral layer that sits at the AP inbox, reads the mess, checks it against the systems a team already runs, and turns it into clean, governed action, with a person in the loop for anything risky.

The first wave of AI in finance was mostly talk: summarize a contract, explain a variance. The answers were good, and they were still just answers. The next wave has to operate inside the systems where the work happens, without breaking anything.

That is why the Model Context Protocol, or MCP, has started showing up in finance conversations that have nothing to do with engineering. MCP is an open standard, introduced by Anthropic in late 2024, for connecting AI systems to outside tools and data. OpenAI, Microsoft, and others support it too.

The major finance vendors are racing to make their systems reachable this way. Ramp initially shipped a connector that lets AI assistants like Claude and ChatGPT analyze spend, review bills, and complete expenses, with each user’s permissions carried over. This week Ramp went further and launched Stack, an AI operating system for accounting firms that runs the monthly close end to end, every step reviewable and auditable. Intuit has shipped an early-preview QuickBooks server for developers. NetSuite has launched an AI Connector Service on MCP, and Microsoft is adding a version to Dynamics 365.

Each of these agents starts from inside one system. Ramp reasons over Ramp, Brex over Brex. The incumbents are reaching outward now, and Stack is the clearest example, but they still begin from the center of their own platform. The work that creates the most drag does not sit in any one system. It spans several at once, and it usually starts upstream of all of them, in the inbox.

MCP has a limit, too. It is plumbing, and plumbing does not make a workflow safe. Finance teams still need the controls that make automation trustworthy: role-based access, approval gates, audit logs, source evidence, and rules about what it can do. Every serious vendor is adding these, so controls are becoming the price of entry rather than the thing that sets a product apart. What separates one agent from another is judgment: how to read a vendor’s intent, when an action is safe, what evidence it needs, and what has to go to a person. That judgment does not come from the connector.

The practitioner market is converging on the same point.

Gaapsavvy recently hosted a finance AI session with Adam Dix from Anthropic’s finance team. More than 1,300 people attended, most of them in-house finance and accounting practitioners, and their questions were revealing. They cared less about whether a model could write a memo and more about where AI-built workflows should live, how finance data should be governed, and how a non-deterministic tool can be audited or recreated later. The real question in accounting has moved past whether the model can do the task. It is whether the workflow can be trusted and verified.

The demand mapped to real finance work: reconciliations, accruals, variance analysis, close, AP, AR, cash, and audit support. Practitioners want AI that operates inside the systems and controls where the work already happens. The best of it will not come from a model vendor or a single ERP working from the outside; it has to absorb the judgment of the people who actually close the books and answer the auditors.

MCP can help with the integrations and give agents access to systems. It cannot decide where evidence belongs, how approvals should work, what an auditor needs to recreate, or when a human has to stay in the loop. The protocol gives agents reach. Practitioner knowledge gives them judgment. Governance makes the result usable. Accounting AI needs all three.

For fractional CFO and outsourced accounting firms, the reason to care is leverage. These firms rarely lose margin because their partners cannot think strategically. They lose it because staff drown in client inboxes, spreadsheets, bill systems, and vendor threads. Every new client brings another finance stack, another set of vendor behaviors, and another place for small errors to become client-visible.

One super-agent spanning every client is dangerous and sloppy. The real opportunity is a repeatable operating model: one workspace per client, client-scoped access to email and accounting systems, read-only to start, human approval for every external reply and every write, and a visible record of what the agent found, drafted, escalated, or cleared.

If a firm can take on more client AP volume without hiring in proportion to client count, the unit economics change. The metric that matters is clients per staff member with no drop in service quality.

For a controller or finance lead at a company of 50 to 500 people, the calculus is different. The goal is narrower: keep accounts payable from falling apart while the team stays small.

These companies often already own modern tools: QuickBooks, Ramp or Brex, maybe Tipalti or BILL. Vendors do not live tidily inside any of them. They email, follow up, ask whether they were paid, send the same document twice, and escalate when something is late.

The stack is modern. The front door is still email.

A connected agent belongs there first, as a controlled AP clerk that watches the inbox, checks the systems, and keeps people focused on judgment. The test is simple: can it tell you what AP-relevant work arrived today, what it did with each item, what is blocked, and what needs approval?

The next stage in accounting will be agentic workflows that move across systems while preserving control. The ledger stays in the system of record, and the agent becomes the translation and coordination layer around it.

This is why the connector race is a tailwind rather than a threat for an independent layer. The more systems become agent-accessible, the easier they are to coordinate across. The incumbents will reach across systems too, and some are already saying so, but each does it from the center of its own platform, with its own product as the destination.

Ramp Stack is the sharpest version of this, and the contrast is the whole point. Stack starts at the close. It works downstream, on structured records that have already made it into the system: reconciliations, journal entries, recurring schedules. That is the back half of the process. The front half is the inbox, where the invoice shows up with no PO and the past-due notice lands with no context. None of that is a clean record yet. Someone has to read it, check it across systems, and turn it into something a close engine can use. A close is only as good as the inputs that reach it, and the inputs start in the inbox.

MCP matters, but it does not finish the job. The future of accounting AI is a governed agent layer that turns unstructured finance work into controlled system action. For fractional firms, that means delivering more service with the same team. For lean finance teams, it means using AI to take back control of the AP inbox so the accounting team can get back to doing accounting.

That gap is where TenKey sits: upstream of the close, and upstream of any single platform. Stack and the platform agents reason over the records inside their own walls. TenKey reasons over the mess before it becomes a record, checking it across BILL, QuickBooks, Brex, and Tipalti as equals.