An AI customer service agent is a business agent who works your support inbox: triaging what comes in, drafting replies from your own documentation and past tickets, resolving the questions that have known answers, and escalating the ones that do not. The goal is not to hide from customers behind automation. It is to answer faster without answering worse.
This is the support chapter of our complete guide to business agents.
What an AI customer service agent actually does
Triage. Every ticket read, tagged, and routed within minutes of arriving: billing to billing, bugs to engineering, the angry one to the top of your list. Nothing sits unread over a weekend.
Replies, sent. For the questions your help docs already answer, the agent answers end to end and cites where the answer came from. Password resets, shipping status, plan questions: the repetitive half of most inboxes has a documented answer, and that half is the agent's to resolve, not just to draft. You name in the brief which categories they own outright and which wait for your eyes.
Escalation with context. When a ticket needs a human, the handoff includes what the customer already said, what was already tried, and what the agent recommends. Your team starts from the middle of the story instead of the beginning.
Patterns, not just tickets. Ten tickets about the same confusing screen is a product finding. An agent who reads every ticket is positioned to notice, and to write it up.
Speed is what customers notice first. Accuracy is what they remember.
What a customer service agent does not do
They do not calm someone down on a phone call, make exception decisions about refunds you have not authorized, or paper over a product problem with charm. The judgment calls and the relationships stay with your team. An honest support agent makes your humans faster, and makes sure the easy tickets never reach them at all.
One boundary worth stating because this category overstates it constantly: a business agent is not a wall between customers and people. Every escalation path should end at a human, visibly.
How a support agent works in Dock
In Dock, your support agent is a named agent with their own memory, which matters more in support than almost anywhere else. They remember the customer who reported this same bug in June, the workaround that fixed it, and the phrasing that made it worse. They work inside your real help desk, and their work is visible: ticket summaries, the replies they sent, and weekly pattern reports land in a shared workspace your team can open and check.
They also run on a schedule. Inbox sweeps happen every morning whether or not anyone remembered to ask, and a weekly report tells you what customers are actually writing in about.
Triage that reads the customer's history
A ticket on its own can mislead. "This isn't working, I want my money back" looks like a refund request. Read next to the customer's history, it might be the third complaint this month from someone on your biggest plan, which makes it a customer about to leave. A quick refund reply would close the ticket and lose the account.
So a good triage pass looks past the message itself:
- The customer's record. Past tickets, what they bought or which plan they are on, and how the last conversation ended.
- A label, a priority, and a reason. Each routing decision comes with a sentence or two on why, naming what the agent looked at. You can read the reason in seconds and disagree with it.
- A clear home for the risky ones. Anything that reads as churn risk goes to whoever handles your most important customers, with the history attached, instead of into the refund pile.
Once a week, pick a handful of routed tickets and compare where the agent sent them with how they were actually resolved. Where the two differ, tighten the brief. That short review is how triage judgment gets sharper over time instead of drifting.
Replies that sound like your business
Generic AI replies sound like a different company. They apologize too much, hedge, and promise to "look into this right away." Customers notice, even if they could not say why. A support agent drafts in your voice when you give them three things:
- The customer's earlier conversations, so the reply answers this person and not an average customer.
- A short written voice guide. Pick a position on a few plain questions: formal or casual, warm or matter-of-fact, how often you apologize. Add a few example sentences for each and a list of phrases you never use.
- Your best past replies, so the agent starts from answers that already worked.
The step most teams skip is the edit. When you rewrite a draft, the rewrite is the lesson. If the agent opens a late-order reply with "I'm so sorry for any inconvenience" and you change it to "That's too long. It's due Thursday, and here is the tracking link," the agent keeps that pair in memory. The next late-order draft starts from your version, and so does the one after it.
Spot-check a few sent replies each week against the voice guide. Drift shows up early when someone is looking.
The work around the inbox
Answering tickets is the visible part of support. Two quieter jobs decide whether it keeps getting better, and both suit an agent who reads everything.
A case brief that lives as long as the case. Some problems span a week, three people, and a dozen messages. Instead of rebuilding the story from forwarded emails at each handoff, the agent keeps one running brief per open case in your shared workspace and updates it as things happen: what the customer reported, what you have promised them, what is still open, and who owns it now. Anyone who picks the case up starts from the brief. When someone asks what you told this customer last month, the answer is one search away.
Help doc gaps, written up as fixes. When the same question keeps arriving and your docs do not answer it, or customers reply that an answer did not help, that is a gap in your documentation. The agent writes it up as a proposed edit: the question, the answer your team actually gives, and the page it belongs on. You make the change or tell the agent to, and the next customer who asks gets the better answer.
Start strict, then loosen on evidence
The line between "handle it" and "flag it" is the whole job, and you draw it in stages.
In the first weeks, keep it tight. Set the agent's permissions so sending a reply always asks you first, and let them flag more than they need to. You read the drafts, fix what is off, and send. That costs you some time early, and it buys you a record to judge from.
Then let the record decide, one category of ticket at a time:
- If you approved most drafts in a category without changes, move it to the list the agent owns outright.
- If you keep rewriting drafts in a category, leave it in review and fix the docs or the brief behind it.
- If an owned category starts producing complaints or reopened tickets, move it back. Trust you gave can be taken back.
Some categories never move. Angry customers, important accounts, real defects, and anything with money at stake go to a person every time, with the context attached. Everything the agent does lands in the shared workspace, so you can always check what was sent and why.
Hiring one: a short checklist
- Start with triage week one. Let the agent sort and route while your team answers; you will see where their judgment is trustworthy. Week two, hand them the known answers.
- Feed them your real documentation. The agent answers from what you give them. Thin help docs produce thin answers; fixing the docs is part of the hire.
- Set the escalation line explicitly. Refunds above a number, legal threats, churn risk: name the categories that always go to a person.
- Spot-check replies weekly, not never. Tone drifts, products change, and the review is how the agent learns.
- Measure first response time and reopen rate. Faster is only a win if the answers hold.
What it costs
Dock pricing is flat per-person monthly: Starter at $20 per month, Standard at $100, Scale at $200, each with a monthly allowance of business credits. Compare that to the cost of the customer who left because nobody answered for three days.
Go deeper
The broader picture is in Business agents: the complete guide.