A burst pipe at 3 AM is a useful test case for any AI agent you put in front of customers. The agent can handle almost every step around the problem. It cannot do the one step that matters: someone has to go to the house and fix it.
That makes the handoff the most important piece of the design, and it is usually the least specified.
Why this case
I run EGA Futura, a native ERP on Salesforce, and co-founded Vantegrate, an AI agent company, with Francisco Morales. In a LinkedIn post I wrote in Spanish today, I noted that AI now writes code, analyzes contracts and builds apps, while a broken pipe in the middle of the night still needs a person who shows up. For engineers, that sentence becomes a requirement: design the agent so the human who shows up gets everything they need, and nothing the agent guessed.
1. Make triage output structured, not a chat log
The agent talks to the customer, but what it hands off should be a record with fixed fields. Free text gets skimmed at 3 AM; fields get read.
{
"request_id": "REQ-2026-11-20-0412",
"received_at": "2026-11-20T03:04:00-03:00",
"category": "water_leak",
"active_leak": true,
"shutoff_valve_closed": "unknown",
"customer_answers": [
{"q": "Is water still coming out?", "a": "yes"},
{"q": "Hot or cold water?", "a": "cold"},
{"q": "Can you close the main valve?", "a": "not sure where it is"}
],
"agent_assessment": {
"urgency": "high",
"confidence": 0.7,
"reason": "active cold water leak, valve not closed"
},
"proposed": {
"work_type": "emergency_plumbing",
"technician_id": null,
"window_start": null
}
}
Two details matter. "unknown" is a valid answer and must stay distinct from "no". And the agent's assessment carries a reason and a confidence, so the human can disagree with something specific.
2. Split permissions into read, propose and write
I described this pattern for sales data in Read, propose, write. It fits field work even better:
| Action | Agent | Human on call |
|---|---|---|
| Read schedules, skills, absences, van stock | yes | yes |
| Create the request and store answers | yes | yes |
| Propose technician, time window, part | yes (draft) | yes |
| Confirm a night dispatch | no | yes |
| Tell the customer someone is on the way | only after confirmation | yes |
| Diagnose the cause | no | technician on site |
The night dispatch has a cost and a risk. A named person should accept it.
3. Write the escalation rules down as code, not prompt text
Prompts drift. Escalation rules belong in a function the agent calls, with tests.
def needs_human_now(req) -> bool:
if req["active_leak"] and req["shutoff_valve_closed"] != "yes":
return True
if req["agent_assessment"]["confidence"] < 0.6:
return True
if req["category"] not in KNOWN_CATEGORIES:
return True
return False
If this returns true, the agent pages the on-call person and stops proposing.
4. Make the handoff idempotent
Customers send the same message three times when they are scared. Key every request on channel plus sender plus a short time window, and update the existing record instead of creating new ones. Otherwise the on-call person wakes up to three dispatches for one leak.
5. Close the loop after the visit
The technician knows things the agent never will: what was behind the wall, which part failed. Let the agent draft the job summary from the technician's notes and photos, and require the technician to confirm it. That record is what improves the next triage.
What to measure
Count visits that needed a second trip because a part or a question was missing at first contact. If that number does not go down, the agent is not helping, no matter how fluent the chat looks.
The longer version of the idea, in Spanish, is on EGA Futura Discover. If you build this on Salesforce, the native ERP for Salesforce keeps billing for the visit in the same org, and Vantegrate covers the agent side in AI agents for field service (Spanish).
The person who rings the doorbell at 3 AM should arrive knowing what they will find. That is the whole job of the agent.
Top comments (0)