SaaS & Customer Support// topic
Drafted replies and agent assist
In short
The stage where a model writes and a human still sends — how a reply is drafted in your team's voice, cited back to something real, and how the agent's edits become the best evaluation data you will ever get.
5 pages
diagnostics
- Agents are deleting the same sentence out of every draftA repeated identical deletion is the cheapest diagnostic signal a drafting system produces. It says something is still emitting that sentence — and the prompt is the wrong place to stop it.diagnostic11 min
- An unrendered placeholder shipped in a live reply"Hi {{first_name}}" reaching a customer is not one bug. It is 5, and the first question — did the model write those braces, or did your template engine fail to remove them — decides which.diagnostic10 min
- The draft promised a refund window that does not existA model asserting a commercial term nobody wrote is not a hallucination in general. It is one of 3 specific defects, and the first search you run tells you which.diagnostic9 min
definitions
- Draft acceptance rate: three outcomes it has to separateOne acceptance percentage hides the only signal worth having. Sent unchanged, sent after edits and discarded are three different verdicts, and the middle one carries almost all the information.definition7 min
- Internal note: the half of a ticket the customer never readsEvery helpdesk has 2 compose boxes behind one text field. The internal note holds the reasoning, the doubt and the account context — and it is discoverable, which changes what belongs in it.definition6 min
Other topics in SaaS & Customer Support
See allTicket and shared-inbox triageThe first thing any support AI should do and the lowest-risk place to start: reading every inbound ticket, email and chat, deciding what it is about, how urgent it is and who owns it — before a customer ever sees a word the model wrote.9 pagesBuild, buy, or extend what you already pay forYour helpdesk vendor already sells an AI add-on, usually priced per resolution — an honest read on when that is the right answer, when a custom layer earns its keep, and how the two coexist.2 pages
Working on something in this space?
Tell us where you are in a sentence or two. We'll tell you honestly whether we're the right team, and what a sensible first slice of the work looks like.
Start the conversation