Contents
01Start by counting, not by buying
Export the last three months of tickets and group them by reason. Nearly every support queue has three reasons that make up a large share of volume, and they are almost always the same shape: where is my order, how do I do this one thing, and something is broken in a specific place.
That table decides everything else. Without it you are buying a tool to answer questions you have not counted, which is how support automation projects end up measured in licences rather than in hours saved.
02The three categories
Sort the reasons into three buckets and treat each differently.
- Remove: caused by the product. Two confusing screens generating a third of the tickets is a design fix, not an automation problem.
- Automate: one correct answer that can be looked up. Order status, opening hours, how to reset something, booking and rescheduling.
- Keep with people: money disputes, anything with a legal edge, anything where the customer is already upset.
Automating the third bucket is the fastest way to turn a support problem into a reputation problem.
03What a good assistant does
It answers from your documents rather than from general knowledge, which is what keeps it from inventing a refund policy you do not have. It says when it is unsure and hands over with the full context attached. It states that it is an assistant. And it writes every conversation into the same place as everything else, so a person picking up the thread sees what was already said.
An assistant that cannot do the last two is not ready for customers, whatever the demo showed.
04The first month decides it
Review transcripts weekly for the first four weeks. You are looking for two failures: confident wrong answers, which are the dangerous ones, and unnecessary handovers, which mean the assistant is more cautious than it needs to be.
Our AI work includes four weeks of transcript review after launch for exactly this reason. A script written before launch is a hypothesis, and the only place it gets tested is in real conversations.
05What it costs
AI assistant and voice agent work starts at $9,000, scoped against your real tickets and recordings rather than a demo. If the counting exercise shows that your volume is concentrated in two product screens, we will tell you to fix the screens first. That advice is cheaper for you and it is the reason the number stays down afterwards.
Questions
Will customers accept an assistant?
Most do when it is faster than waiting and honest about what it is. What they do not accept is discovering they were misled, or being trapped with no way to reach a person.
What about answers that need to be exactly right, like policies?
Those come from your documents, quoted rather than paraphrased, and anything the assistant cannot ground in a document goes to a person. If the policy itself is ambiguous, the assistant will find that ambiguity faster than any audit.
How much volume can realistically be handled?
It depends entirely on your ticket mix, which is why counting comes first. A queue dominated by order status questions automates far further than one full of billing disputes.
Where we can help