Notes on CX

The three buckets

How I decide what AI should touch in a support queue, what it should start, and what it should never see. A strategy I learned at Uber long before the agents arrived.

AJ Musselman  ·  Published 2026  ·  4 min read

"How much of this can AI do?" It's the question everyone is asking about support right now, and there's no useful answer for a whole queue. There's an answer for each type of ticket, and the answers are very different.

My approach didn't start with AI. It started at Uber, on the first in-house social media response team, while I was carrying Incident Response casework at the same time. The job every day was triage: what can be resolved right there on social, what needs to move to a DM, and what has to go to a phone call or an email with a specialist. Three lanes, sorted by how much judgment and how much risk a conversation carries. When AI agents came along, I was already sorting tickets this way. The lanes just got new names.

So I sort every ticket type into one of three buckets.

Bucket 1
AI, end to end
Cancellations, where-is-my-order, general FAQs. The answer lives in a system and the customer just needs it fast.
AIHuman
Bucket 2
AI starts, a human finishes
Replacement orders and anything else where AI can gather the facts but a person needs to take the action.
AIHuman
Bucket 3
Humans only
Adverse events and anything brand-critical. AI never sees these, and the routing makes sure of it.
AIHuman

AI, end to end

These are the tickets where the customer already knows what they want and the answer lives in a system. Cancel my subscription. Where is my order. What's your return window. There's no judgment call. There's a lookup and an action, and a well-built agent does both faster and more consistently than a person at 11pm on a Sunday.

This bucket is where the automation numbers come from, and it's larger than most teams think. The catch is that "end to end" has to mean end to end. The agent needs the tools to actually cancel the subscription, not just explain how. Customers can tell when an AI can only answer questions about cancelling.

It's also where the business case is. These are the most common tickets in any DTC queue, and they arrive at every hour. Handling them with AI gives customers an answer at any time of day on the things they ask about most, and it means the human team doesn't have to grow in step with ticket volume.

AI starts, a human finishes

This is the bucket where the design matters most. AI does the collecting and a person does the deciding. A customer writes in about an order that arrived damaged. The agent confirms the order, asks for a photo, checks the timing, and writes up a clean summary. Then a person looks at it, decides, and sends the replacement. The person's time goes to the part that needs judgment instead of the intake.

Done well, this bucket cuts handling time more than the first one does, because the human part of these tickets goes from ten minutes to one.

Humans only

Someone had a reaction to a product. Someone is describing a safety issue. Someone is upset in a way that could turn into a public problem for the brand. These tickets are a small share of volume and a large share of risk, and AI should never be the first thing that touches them.

That isn't because the model would say something terrible. Usually it wouldn't. But "usually" is the wrong standard for an adverse event, and a friendly automated acknowledgment that arrives before a human gets involved is hard to take back. This bucket needs hard routing that pulls the ticket out of the AI flow before any reply is generated, and a human-owned SLA that is faster than the AI's.

How I roll it out

Bucket one goes live first. On day one, buckets two and three skip AI entirely and route straight to people. That gets the easy, high-volume wins right away without putting the conversations that matter most at risk, and it gives the team a few weeks to QA the agent on the simplest tickets before it's trusted with anything else.

Once the QA holds, bucket two ramps in: the agent starts gathering information on replacement-type tickets and hands them off. Bucket three never ramps.

I've done this twice with Siena AI, for a functional beverage brand and a wellness supplement brand while I was consulting at Create CX. Both reached 40 to 65% ticket automation within three months of implementation, and both got there by sorting first and automating second. Before that, I applied the same thinking to Gorgias automation at Juice Beauty. The automation rate depends on the mix of tickets a brand has, which you mostly can't change. The sorting is the part you control.

Views are my own and don't represent any current or former employer or client.

More notes on CX