Ten minutes before a customer call, the same scramble starts. Open the CRM. Check the last invoice. Scan support for open tickets. Dig for the last note someone left. Then paste scraps into chat and hope the assistant stitches them into something useful.
That prep is a real job. It should not depend on which tabs are open, or on how the model decides to hunt today.
What breaks
A freeform prompt like “brief me on Northwind” sounds fine until it isn’t. The assistant invents which systems to check. It mixes stale notes with live tickets. Two runs on the same customer return different shapes. You spend the first minutes of the call reconstructing what you already knew.
The failure is not “AI can’t research.” The failure is that the pack was never a contract — only a request.
What to build
One tool. One input: customer name or id. Fixed output every time:
- Last invoice status (paid / overdue / amount / date)
- Open tickets (id, status, owner)
- Last note (who wrote it, when, one-line summary)
Wire each field to the source you already trust — sheet, CRM, helpdesk, Drive folder. Lock the args you do not want the model to invent. Keep the surface small: this tool prepares the call; it does not send email or close tickets.
In chat you type something like: “Call pack for Northwind.” The assistant calls the tool and reads the pack back. Same three sections, same order, every time.
How to verify it worked
Run the pack twice on the same customer five minutes apart. The structure should match. Numbers should match the source of truth, not a paraphrase. If a field is empty, the tool should say empty — not invent a status.
Then run it once more right before a real call. If you still open three tabs, the tool is missing a field you actually need. Add that field. Do not grow a second “research everything” tool.
Where Smia fits
Design that small tool, host it on a server URL, and call it from ChatGPT, Claude, or Cursor. The assistant still does the talking. The pack stays a fixed job you can re-run without renegotiating the path.