Customer service teams answer the same questions every day, under time pressure, in front of customers. AI can take the searching and the first draft off their hands. It should not take the promise.
Why service teams override AI
In Camunda's September 2026 survey of 5,000 employees, 44% said they had manually overridden AI outputs because the underlying process was not set up correctly, and 65% said they could use AI much more effectively if it was implemented differently.
In service, the pattern is familiar. A chatbot answers from an out-of-date article. A drafted reply promises a date the warehouse never confirmed. The agent fixes it, the fix is not recorded, and the dashboard still says the AI is working.
The answer is not less AI. It is a clear line between what AI may send and what it may only draft.
Three workflows that keep the human
Support tickets: triage and draft replies
AI reads each ticket, classifies it, finds similar solved tickets and the right help article, and drafts a reply with its sources. Hard cases go to a specialist with a summary. An agent approves every reply before it is sent. Edits feed back into better drafts and flag help articles that need updating.
Customer status updates
Coordinators spend hours writing arrival, delay and completion messages, and customers still call to ask. AI sends routine updates from live job or order status, using only templates the team has approved, on confirmed status events. Delays, damage and anything unusual go to the coordinator with a draft.
Field service: breakdown triage
When a customer reports a breakdown, someone has to pull asset history, parts and technician availability from several systems. AI assembles that into one summary with a likely cause and a drafted customer notice. The service manager picks the technician and approves the message. Overnight, AI flags equipment with repeat visits, so a person can open an investigation before the next breakdown.
The approval rule
Write down, before you build, which messages may go out without review. Our default:
- May send alone: routine updates from approved templates, on confirmed events, within a frequency cap
- Drafts for approval: every reply to a customer question, every delay notice, every apology
- Never: offer refunds, credits or compensation; commit to a date operations has not confirmed; reply to a complaint
How to tell whether it is working
Measure first response time, reopen rate and escalations before the change. After it, add two numbers most teams forget: how long agents take to approve a draft, and how often they edit it. Fast approvals with no edits can mean excellent drafts or nobody reading. A weekly sample review tells you which.
Get the use cases
The short guide below covers these three workflows. Newsletter subscribers get the full playbook, with the before-and-after workflow, control points and baseline sheet for each.