When not to use an MCP: AI agents versus integrations for Fishbowl
We sell an MCP for Fishbowl, and we also build Fishbowl integrations. The two solve different problems, and the AI projects we see go wrong usually used the first for a job that needed the second. An AI agent makes a judgment call each time it runs. An integration does the same thing with the same input every time. If a job has to run unattended, on a schedule, and match to the penny, it needs an integration.
How they differ
When you ask an agent to turn a spreadsheet into a sales order, it reads the columns, decides what each one means, and maps it to Fishbowl. Most of the time it gets it right, and the preview shows you when it doesn’t. Run the same file through it next week and it may make slightly different choices. With a person checking each result, that’s fine.
An integration has the mapping written into code. A Shopify order becomes a Fishbowl sales order the same way at 2 a.m. as at 2 p.m., and when something fails, it fails the same way and can be retried. Nobody needs to watch it.
Anti-patterns
These are the uses we’d talk a customer out of.
Replacing a recurring integration with a chat. “Every morning, ask the AI to import yesterday’s Shopify orders” works the first week. Then the volume grows, someone forgets a morning, and an order is mapped differently from the day before. If the same file or feed arrives on a schedule, build the integration.
Writing to Fishbowl with nobody checking. An agent that creates orders or adjusts inventory with no preview and no approval turns a wrong guess into a wrong transaction. ILC MCP doesn’t write anything until a person approves a preview, and it should stay that way.
Numbers that must be the same every time. A month-end report, a KPI on a dashboard, or anything your accountant reconciles shouldn’t depend on how the AI writes the query that day. Use a saved query, where the SQL is fixed and only the inputs change, or a custom report.
High-volume or time-critical work. Receiving on the dock, picking, label printing, and anything a person scans hundreds of times a day needs to respond instantly and the same way each time. That’s work for a warehouse app or an integration.
Business rules written as instructions. A note telling the AI “never sell below cost” can be misread or skipped. If a rule must never be broken, it belongs in a tool’s code, where it’s checked on every call.
A rule of thumb
| Use an AI agent when | Build an integration when |
|---|---|
| The question or job is new each time | The same file or feed arrives regularly |
| A person checks the result | It runs unattended, on a schedule |
| The input format varies | The format is fixed |
| A close answer is useful | The result must reconcile exactly |
| It happens a few times a month | It happens many times a day |
From chat to integration
Most jobs move from left to right as they repeat:
- A one-off job by conversation. A new customer’s spreadsheet, a question nobody has asked before.
- A saved query or custom tool. When the same question comes up every week, an admin saves the query, or we build a tool with the rules in code. The AI still handles the conversation, but the work it calls is fixed.
- An integration. When it runs on a schedule with nobody watching, it doesn’t need the conversation at all.
The agent is a good place to find out which jobs repeat. When one does, move it to step 2 or 3.
Related: