An AI agent for customer service reads your incoming messages, answers the ones it can from your own help content and account records, and hands the rest to a person with the history attached. In a business of 5 to 50 people it should own the questions that repeat. It should not own the ones that cost money or patience.
That split is the decision, and you can make it in an hour with your own inbox. The sorting exercise is below, along with what the vendors charge once you have the counts.
What can an AI agent do for customer service?
A support agent sorts, answers, looks things up and hands off. Sorting runs on every message that arrives. Answering is limited to the kinds of message you approved it to answer, which should be a short list at the start. Looking something up means it needs read access to wherever that answer lives. The handoff is the part buyers skip, and it is the part customers notice.
| Job | What the agent does | Who finishes it |
|---|---|---|
| Sort and route | Reads every new message, tags it, sends it to the right place | Agent |
| Answer the repeats | Replies from your published help content | Agent |
| Look up a status | Checks an order, booking or account and replies with what it found | Agent |
| Draft the hard ones | Writes a reply and holds it | A person approves and sends |
| Chase a missing detail | Asks for the order number, address or photo so the queue is not stuck | Agent |
| Escalate | Passes the thread over with what it already knows | A person |
Only two of those rows send a reply with nobody looking, which is where a sensible first version lands. The roles this fits into across a business are in eight practical AI agent use cases, and the four shapes you can buy one in are in AI agents for business.
Sort your last 100 customer messages
Open the last 100 messages your team answered and put each one in a bucket. Do not sample from memory. Memory keeps the dramatic ones and loses the people who asked where their order is. The counts are what decide whether an agent pays off, and they take about an hour to get.
| Bucket | What it looks like | What the agent gets |
|---|---|---|
| 1. Same answer every time | Hours, policy, how to return something, how to reset a password | Answers and sends |
| 2. Same answer, needs a lookup | Where is my order, is this in stock, when is my appointment | Answers and sends, once connected |
| 3. Same question, different answer | Pricing, a custom request, can you do this for us | Drafts, a person sends |
| 4. Money, anger or risk | Refunds, cancellations, complaints, anything legal or safety related | A person, always |
Then do the arithmetic. Add buckets 1 and 2. That number is what an agent can take off your team this month. Multiply it by the per-ticket price in the next section and compare it with the hours it gives back.
As a planning rule we use, not a benchmark anyone has validated: if buckets 1 and 2 together are under about a third of your volume, fix your help content first and buy the agent later. An agent with nothing written down to answer from is an expensive way to say "let me check on that."
The other thing the exercise tells you is which bucket 1 answers should not be answers at all. If ten people asked the same question about your return window this month, the honest fix is a clearer return policy on the page where people buy. Answering it ten more times next month looks like success on a dashboard, and it is a leak you are now paying by the message to plug.
What never reaches the agent on its own
Four kinds of message stay behind a person no matter how good the agent gets: anything that moves money, anything that ends the relationship, anything where the customer is already upset, and anything a lawyer would want to read first. Write that list down before launch, in your own words, with your own examples.
- Refunds, credits, discounts, and any change to what somebody owes you.
- Cancellations, downgrades, and "I want to speak to the owner."
- Second contacts about the same unresolved problem. The agent already had its turn.
- Injury, safety, discrimination, privacy, and anything involving a minor.
- Anything a customer could screenshot and post.
That last one is not a legal category. It is the practical one. A reply that is technically correct and tone-deaf costs more than a slow one. How to decide this per action, rather than per inbox, is in human in the loop for AI agents, and the matching question of what the agent should be able to read and change is in what access should an AI agent have.
How much does an AI agent for customer service cost?
Both of the big help desk vendors now charge you per ticket the agent closes. Seats still cost what seats cost, and the agent is billed on top, by the resolution. That is good news at low volume and it gets expensive quietly, so run your own numbers from the bucket counts rather than a vendor's example.
Prices below come from each vendor's own public pricing pages, read on October 17, 2026.
Zendesk says its AI agents are included in every Support and Suite plan, billed on what it calls automated resolutions, which it defines as a customer request the AI agent resolved without escalating to a human. Paid yearly, seats run $19 per agent per month on Support Team, $55 on Suite Team and $115 on Suite Professional, and month to month those are $25, $69 and $149. Support Team and Suite Team include 5 automated resolutions per agent per month, and Suite Professional includes 10. Past that you pay $1.50 per committed resolution, or $2.00 if you take them as you need them, on every plan. Those figures are on Zendesk's pricing page.
Run a small business through that. Three people on Suite Team paid yearly is $165 a month, with 15 resolutions included. If the agent closes 120 tickets, the other 105 cost $157.50, so the month lands near $322.
Intercom prices its agent, Fin, at $0.99 per outcome with a $49 monthly minimum that includes 50 outcomes, no seat cost, and no setup or connection fee. It states that you pay nothing when Fin cannot answer, when it fails the action, or when the customer asks for a human at any point. It also works with other help desks, including Freshdesk, Salesforce and Zendesk. See Fin's pricing page and Intercom's, where help desk seats start at $29 per seat per month on the Essential plan. The same 120 closed tickets costs $118.80.
Read the "you pay nothing if it cannot answer" clause as the useful signal it is. It tells you the vendor is confident enough to put its fee on the outcome, and it tells you how they count. A ticket the agent gave up on is free, and it is also still in your queue.
Our own work is priced differently because it is not a support product. The build starts at $1,000 one time, including three coaching sessions for your team, and hosting and maintenance start at $200 a month, which does not change with ticket count. That is on our pricing page. Market ranges for having an agent built, by us or anyone else, are collected in how much AI agent development costs.
What the agent needs from you before it can answer
It needs written answers it can quote, read access to whatever system holds the status people ask about, and a rule for what it does when it is not sure. None of those are technical. Miss the last one and you get a confident guess where you wanted a handoff.
The written answers are the part that stings. If your refund window, your lead time and your service area only live in the head of whoever has been there longest, an agent cannot answer from them. Neither can your next hire. Writing them down is work you owed yourself anyway, and in our experience it moves the quality of the answers more than any choice you make about the agent itself.
The access part is narrower than people expect. For most of bucket 2, read only is enough. The agent needs to see the order, not change it. Start there and widen it on evidence.
What we do with our own inbound messages
We run our own agents on our own messages, so this part is first-hand rather than a recommendation we have not tested ourselves. Three rules came out of it, and each one was written after something went sideways.
An agent replying to someone outside the company says only what is already published on our site. Not the reasoning behind a price, and not who a client is. When somebody keeps pushing past that line, the agent stops answering and tells a person. We wrote it because a friendly question is a remarkably effective way to get an eager assistant to explain more than it should.
An agent drafts our customer email and a person sends it, including the easy ones, because the easy ones are where carelessness collects.
The failure we did not see coming was never a wrong answer we could spot. It was a confident one. An agent that says "done" convincingly is harder to catch than one that breaks, which is why we read a sample of what ours sent rather than trusting the summary. The rest of what broke, and what it cost us, is in what went wrong when we gave AI agents real work.
How do you know it is working?
Track four numbers for the first 60 days, and read a sample by hand every week. The numbers tell you whether the agent is carrying load. The sample tells you whether it is carrying it well, which is the harder question and the one that costs you if you skip it.
- Closed with nobody involved, as a share of all messages. This is the number the vendor bills you on.
- Handed over, and how long the customer waited for the person.
- Reopens. Customers who came back about the same thing are resolutions you already paid for.
- Drafts you had to fix before they went out. If this stays high, bucket 3 is not ready to move.
Then pull 10 replies the agent sent each week for the first month and read them as a customer would. You are looking for the fluent, friendly, wrong one, which no dashboard is going to flag for you.
Frequently asked questions
Can an AI agent answer our phone calls?
Phone is a different build. It has to hear, speak and decide while the customer waits, and it fails in ways email never does. Treat voice as its own project with its own budget rather than an add-on to a messaging agent.
Will customers know they are dealing with an AI agent?
Tell them. Say it in the first reply, keep the route to a person short and obvious, and never let the agent imply it is a colleague. Being caught out is worse than disclosing, and in support the moment somebody feels handled is the moment the ticket gets expensive.
What happens when the agent gets it wrong?
Decide in advance who finds out and how fast. A wrong reply that a person reviews the same day is a small problem. The same reply discovered in a review next quarter is the problem. That is what the weekly sample and the reopen count are for.
Do we need a help center before we start?
You need written answers. A full help center is one way to hold them. A plain document with your 25 most common questions and the current answer to each is another, and it is enough to start on bucket 1.
Is this the same as the chatbot on our website?
No. A chatbot waits for someone to open a window and type. An agent holds a standing job across your inbox, forms and chat, whether or not anyone is in the room. Where that line falls is in where an AI assistant for business stops.
Where to start
Sort the last 100. Count buckets 1 and 2, multiply by a per-ticket price, and you will know inside an hour whether this is worth a project right now or a help page rewrite instead.
If the counts say yes and you would rather not own the build, the maintenance and the handoff rules yourself, that is what we do. It starts with 30 minutes on your workflows.
