An AI agent must hand a conversation to a human when risk, emotion, or a customer right outruns FAQ. A clean handoff carries history. A bad one asks people to repeat themselves. In Indonesia, bot acceptance is high only while the errand finishes.
Acceptance is not permission to drop humans
At the WhatsApp Business Summit on 11 August 2026, Kompas wrote that 79 percent of surveyed adult internet users found Meta AI responses useful. ANTARA cited 83 percent saying AI agents help them get what they need faster. Meta Business Agent is built to pass to a human team. That is a feature, not a footnote.
Seventy-one percent of people in Indonesia message businesses every day, up from 60 percent in 2024 and 66 percent in 2025, per the same Meta figures. That volume makes bots tempting. The same volume makes a messy handoff feel industrial.
Triggers you do not negotiate
Fraud or a frozen account. A complaint headed to OJK or another regulator. An erasure or access request under UU PDP. Legal threats. Children and health data. A customer who types “agent” or “human” twice. A bot repeating the same answer. A transaction value above a threshold you wrote down.
OJK’s IASC, as reported by Kompas on 8 September 2026, blocked Rp 771.2 billion and returned Rp 206 billion from 19 banks. “Why did my transfer vanish” is not a place for a model to improvise. Hand off with the log intact.
How to transfer without breaking SLA
One thread, one case, one summary an agent can read in 15 seconds: intent, identity already checked, bot steps, reason for escalation. Do not open a new channel. Do not re-ask for an OTP already passed. Measure time to the first human reply after the handoff flag, not only the bot’s first response.
Marketing-Interactive on 21 September 2026 wrote that Telkom is integrating Llama into WhatsApp, with Telkomsel’s Veronika on that channel. Lesson: a telco bot lives next to billing. If handoff drops context, the customer blames the brand, not the model.
AI contact center, omnichannel service and compliance notes for Indonesian service leaders.
Every Monday, five minutes.
Labour and vendors
Minister of Manpower Regulation 7/2026, in force 30 April 2026, limits outsourcing to supporting work. Core customer service is not on that list. A bot does not turn agents into an “AI vendor.” Humans who take handoff remain employees or another compliant scheme, with older arrangements due to adjust by 30 April 2028.
What must not be a trigger
Night hours alone. Accent. A long question that is still an FAQ. Template message cost. Handing off every time the bot is “unsure” will kill deflection. Write a confidence threshold, sample it, and publish the rubric to supervisors.
Since 1 July 2025 WhatsApp templates are billed per message. A handoff that sends a new template has a cost line. Do not withhold a human just to save one utility message.
Frequently asked questions
When must AI hand off to a human?
Fraud, regulators, data rights, repeated anger, or a looping bot. Also when the customer asks for a human twice.
Do customers reject bots?
Meta’s WABS 2026 figures show high acceptance when things finish fast. Rejection shows up when the bot spins.
What must travel with the escalation?
History, identity already checked, and the reason. Not “please tell your story again.”
Does handoff hurt deflection?
Honest deflection will fall. That is better than fake containment that reopens.
Does UU PDP change bot rules?
Yes. Data-subject requests and sensitive data need a recorded human process.
How do we measure a healthy handoff?
Time to first human reply, reopen rate inside 24 hours, and a filled escalation reason.
Book a demo and see your own contact flows running on ConverseAIght
ConverseAIght
