We didn't build our first AI assistant for a client.
We built it for ourselves.
It seemed like the right place to start. We wanted to understand what it actually feels like to hand off a workflow to an AI — the friction, the learning curve, the moments where it earns trust and the moments where it doesn't. So we picked the workflow we knew best: managing a support inbox.
Here's what we built, the decisions behind it, and what we'd do differently.
The problem we were solving
Our inbox was functional but inefficient. Every morning started with the same ritual: scan all new tickets, figure out which ones needed immediate attention, draft responses, log activity. The work wasn't hard — it was just repetitive and time-consuming in a way that added up.
We wanted an assistant that could handle the cognitive overhead of triage, so the human could focus on the actual response.
What we built
The assistant sits alongside the inbox. It does four things:
Triage by urgency and sentiment — it reads every incoming ticket and flags the ones that are high priority, either because of the issue type (billing, outage) or because the customer's tone signals frustration.
Drafts context-aware replies — for each ticket, it pulls the conversation history and suggests a response. Not a generic template, but a draft that accounts for what's already been said.
Answers quick questions inline — the support agent can ask the assistant directly without switching to a separate tool.
Summarizes weekly activity — every week it generates a short report: tickets handled, average response time, recurring issue categories.
The decisions we debated
How much should it do automatically? We considered letting the assistant draft and send responses without human review. We decided against it — not because the drafts weren't good, but because the review step keeps a human in the loop on every customer interaction.
How do we handle wrong answers? The assistant will occasionally misread a ticket or suggest an off-tone response. We built in a simple pattern: every suggestion is labelled as a suggestion, and the agent always reviews before sending.
What does the interface actually look like? The assistant needed to feel like a natural part of the inbox, not a separate tool demanding attention. It lives in a side panel that appears when you open a ticket.
What we learned
The first two weeks were rough in the most instructive way. The assistant was technically working from day one, but it took us two weeks of real use to understand where it was strong and where it needed refinement.
The biggest lesson: an AI assistant is not finished when it's deployed. It's finished when the team has internalized how to work with it — and that takes time, real usage, and an honest feedback loop.
We now show this assistant on every discovery call. Not because it's perfect, but because it's real. Everything we learned building it is now part of how we build for clients.