DEV Community

FusionAL
FusionAL

Posted on

A voicemail becomes a work order draft, and a person checks it in under a minute

A maintenance team gets voicemails all day. "Hi, it's the front office, the heater in room 4 is making a banging noise again, can someone look before Thursday?" Someone has to listen, type it up and guess at the missing parts.

This is a good first workflow to automate, because it is small and the failure is easy to see. Here is a hypothetical demo of how I would scope it. It is an example, not client work.

The one workflow

A voicemail comes in. It gets transcribed. A model turns the transcript into a draft work order with a few fields: location, problem, urgency, requester and callback number. The draft lands in a review queue. Nothing is created in the real system until a person approves it.

That last rule is the whole design. The model drafts. A person decides.

What the draft has to show

A person can only check something in under a minute if the draft is built for checking. Each draft should show:

  • The original transcript next to the extracted fields
  • Any field the model could not find, marked "missing" rather than filled with a guess
  • Any field the model inferred, marked as inferred, such as "urgent" from "before Thursday"

A blank field is honest. A confident wrong room number is a problem, because it looks like every other field.

Checks that run before a human sees it

Simple checks catch most of the obvious errors:

  • Does the room or asset exist in the facilities list? A read-only lookup is enough.
  • Is the callback number a valid format?
  • Does the draft closely match an open work order? If so, flag a possible duplicate instead of creating a second one.

None of these need write access. If the workflow connects to the work order system through MCP, I would give that connection read access for lookups and keep creation behind the approval step. If it does not use MCP, the same boundary applies to whatever API it uses.

Privacy still applies

Voicemails contain names and phone numbers. Even for an internal, non-regulated workflow, decide where recordings and transcripts are stored, who can open the review queue and how long they are kept. Keep credentials out of the prompt and out of logs.

The handoff

Before anyone relies on this, we review it with a handful of real, anonymized voicemails, and the team checks what the reviewer sees. Can they approve a good draft in under a minute? Can they spot a bad one without replaying the audio?

If the answer to either is no, the draft layout needs work before the model does.


I'm Jonathan Melton. I run FusionAL, focused on reviewing and integrating lower-risk internal AI workflows, one bounded workflow at a time.

Top comments (0)