
Breakdown / AI at work · Service integrations
Your customer has an AI assistant. Can it use your service?
Connecting an assistant is the beginning. Permissions, missing fields and a clear record of what changed decide whether it can finish the job.
Imagine a customer asking their AI assistant: “Show me the overdue invoices, match any payments that have arrived, and prepare reminders for the rest.”
If you build financial software, can that assistant work with your product, or will the customer have to open it and repeat the job by hand?
MCP, the Model Context Protocol, is one way to make the connection. It lets an assistant discover and call tools a service provides. Those tools might retrieve an invoice, list expenses or prepare a change. But the label alone doesn't tell you which assistants can connect or what they can actually do.
Follow the customer's permissions
Payhawk, a business spending and expense-management service, documents a connector that uses each person's own login and existing permissions. Its connection guide also lists supported assistants and explicitly says sign-ins from other clients are rejected. That is a concrete limit on the idea that one MCP connection works everywhere. Payhawk connection guide
The company's playbooks show why the individual action matters. Its “check the next payment run is covered” workflow checks funding but does not pay anything. Its flight-request workflow goes through the account's normal approval process, with nothing purchased in chat. Those are documented boundaries, not financial actions we've tested. Payhawk playbooks
For a product team, the practical question is whether the user can tell the difference between asking, preparing and approving. Each stage needs an understandable result.
A successful call can leave something out
Intuit's QuickBooks Online MCP repository offers a different route: a local server for developers or partners, with an OAuth authorization setup. It provides separate switches for disabling create, update and delete tools. Turning off creation alone does not turn off the other two categories. Intuit's repository
There is an especially useful detail in its tool-registration code. The wrapper removes unsupported fields before passing a request to the underlying operation, then adds a warning identifying what was ignored. A comment describes a missing invoice shipping date as the reason for that handling. We inspected the code; we haven't reproduced the reported incident or tested a live account. Tool-registration implementation
That gives a developer a concrete test. Submit a synthetic invoice with a field the tool doesn't support. Does the assistant notice the warning, explain what was left out and check the saved record? Or does it announce that the invoice is finished?
Start with one job customers already do
Our opening invoice workflow is an example, not a claim that either connector completes every step. A sensible first implementation might retrieve outstanding invoices and prepare reminders for a person to review.
Then test the whole job in the assistant your customers use. Keep their existing permissions, make approval explicit where needed, ensure retries don't duplicate a change, and leave a record of the result. Revoke the connection and check that access really stops.
The opportunity for an existing software business is concrete: let a customer's assistant do useful work in your product. The measure of success is whether the intended job was completed, with the right details and approvals, and whether the customer can see what happened.
Sources
- Payhawk connection guidePayhawk
- Payhawk playbooksPayhawk
- Intuit's repositoryIntuit
- Tool-registration implementationIntuit