AI Customer Service Agent: What It Handles, What It Costs, and How to Pick One

An AI customer service agent is software that reads or hears a customer's request, finds the answer in your own systems, and either resolves it or hands it to a person with the full context attached. For a small business it is worth buying when the same dozen questions arrive every day — order status, appointment changes, hours, billing basics — and not worth buying to replace judgment, apologies, or anything that needs a signature.

This post is the small-business version of a topic mostly written for call centers: what the agent should handle, what it costs under each pricing model, where it goes wrong, and how to choose one. It covers chat, email and text first. Phone calls get their own treatment in our guide to AI voice agents for business, and larger support teams should start with our piece on conversational AI for enterprise customer support.

One distinction before the details: an agent is not a chatbot with a nicer name. A chatbot answers; an agent looks things up and takes an action. Scott McAuley's breakdown of AI agents vs chatbots is a useful primer if the two still sound alike.

An AI customer service agent is a software system built on a large language model that can understand a request in plain language, retrieve the relevant facts, and complete a task inside the limits you set. Three parts make it an agent rather than a script.

Knowledge. The agent answers from your material — help articles, policies, price lists, past tickets — not from the open internet. If the answer is not in an approved source, a well-built agent says so and hands off.

Tools. The agent can do things: look up an order, move an appointment, open a ticket, issue a return label, update a contact record. Each tool is a permission you grant on purpose, and each one should have a limit.

Rules for handing off. The agent knows which topics, phrases and customer states end its turn. An angry customer, a refund over your limit, or the words "cancel my account" should route to a person every time, with the transcript and the customer record attached.

The analyst firm Gartner predicts that by 2029 agentic AI will autonomously resolve 80% of common customer service issues without human intervention, with a 30% reduction in operational costs. Read the word "common" carefully. That is a forecast about routine requests, and a forecast is not a result. The uncommon 20% is where your reputation is made, and it still belongs to your staff.

The honest answer is narrower than the sales pages and wider than most owners expect. The test for any request type is simple: could a new hire handle it correctly on day three with a written procedure and read access to your systems? If yes, an agent probably can too.

These are the request types where we see agents earn their keep in small businesses:

• Status questions. Where is my order, is my appointment confirmed, did my payment go through, when does my warranty end.
• Scheduling changes. Booking, rescheduling and cancelling against a live calendar, with the confirmation sent by text or email.
• Policy and how-to answers. Hours, service area, return windows, what to bring, how to reset a password, how to read an invoice.
• Intake and routing. Collecting the details a person needs — account, product, photos, urgency — and opening a complete ticket for the right team.
• Simple account tasks. Updating an address, resending a receipt, issuing a return label, taking a payment through a secure link.

This is the part worth being blunt about. Some conversations cost more to automate badly than to staff well:

• Complaints and anything emotional. The agent can acknowledge, collect the facts and escalate fast. It should not negotiate.
• Money beyond a set limit. Refunds, credits and fee waivers above the threshold you choose.
• Regulated advice. Medical, legal, financial and insurance coverage questions. An agent can schedule the conversation; it should not have it.
• Exceptions to policy. If the right answer is "normally no, but for this customer yes," that is a human decision.

Our post on the calls you are most afraid to automate walks through how those conversations actually go when the handoff is designed first. The decision rule: automate the request types you can write a procedure for, and design the handoff before you design the answers.

An AI customer service agent is sold under three pricing models, and they are not directly comparable until you convert each one to cost per resolved conversation. Prices below are the vendors' published list prices as of November 2026; treat them as examples of each model, not as a ranking.

Per resolution or outcome. You pay only when the agent finishes the job. Intercom's Fin agent is the best-known example: its pricing page lists $0.99 per outcome, charges one outcome per conversation, and sets a 50-outcome monthly minimum when Fin runs on a helpdesk you already own. Conversations passed to your team without an outcome are not charged. This model fits a business that wants spend to track results.

Per conversation or per action. You pay for activity, resolved or not. Salesforce lists Agentforce at $2 per conversation, or through Flex Credits at $500 per 100,000 credits, where a standard action uses 20 credits, which works out to $0.10 per action. This model fits a business already running on that platform's data.

Per seat or per minute. A flat platform fee plus usage. Voice agents are almost always priced this way because phone time has a real per-minute cost. Our own Talos Connect AI receptionist is $149 per month including 500 minutes, and our breakdown of AI receptionist pricing covers what moves that number.

Take a business with 400 customer conversations a month, where an agent resolves half on its own. This is arithmetic on the list prices above, not a quote. At $0.99 per outcome, 200 resolved conversations is $198 a month, plus the helpdesk seats the agent runs on. At $2 per conversation, all 400 conversations is $800 a month whether or not they were resolved. Same business, same month, four times the bill.

Neither number includes the costs that do not appear on a pricing page: writing and cleaning the knowledge base, connecting your systems, testing, and the hour or two a week someone spends reading transcripts. On the small deployments we have run, that labor has been larger than the software fee in the first quarter. The decision rule: ask every vendor for the total at your real volume and your expected resolution rate, in writing.

Most failures are not the model inventing nonsense. They are business decisions that nobody made on purpose.

In Moffatt v. Air Canada, decided in February 2024, British Columbia's Civil Resolution Tribunal held the airline liable for negligent misrepresentation after its website chatbot told a customer he could apply for a bereavement fare after travelling. The airline's own policy page said otherwise. The tribunal rejected the argument that the chatbot was a separate entity responsible for its own actions and awarded the customer about $650 in damages plus interest and fees. The amount was small. The principle is not: an answer from your agent is an answer from your business.

The practical fix is to limit the agent to approved sources, keep those sources current, and block it from stating policy it cannot cite. When a price or policy changes, updating the agent's source is part of the change, not a follow-up task.

Tell people at the start of the conversation. Beyond being good manners, in some places it is the law: California's bot disclosure statute makes it unlawful to use a bot to mislead a person in California about its artificial identity in order to incentivize a sale, and it protects businesses that disclose clearly and conspicuously. If you sell nationally, you sell in California. This is general information, not legal advice; ask your attorney how disclosure rules apply to your business.

The failure customers remember is asking for a person three times and getting the same menu. Every channel needs a visible way out: a plain "talk to a person" option, a handoff that carries the transcript, and an honest message about response time when nobody is available. An agent that answers at 2 a.m. and books a callback for 8 a.m. is a good experience. An agent that pretends nobody will ever be available is not.

Choosing well is mostly preparation. Run these steps in order and the vendor decision gets easy:

1. Count your requests. Pull two to four weeks of emails, chats, texts and call notes and sort them into types. In our experience a handful of request types make up most of the volume.
2. Pick three request types to start. Choose the ones with the highest volume and the clearest procedure. Write down what a correct resolution looks like for each.
3. List the systems each one touches. Calendar, order system, CRM, billing. An agent that cannot read the system that holds the answer can only apologize politely.
4. Write the handoff rules. Topics, dollar limits, trigger phrases, and who receives the handoff during and after business hours.
5. Test with your real questions. Feed each vendor 50 actual customer messages, including the messy ones, and grade the answers yourself.
6. Price it at your volume. Convert each quote to cost per resolved conversation using the method above.
7. Agree on who reads the transcripts. Someone owns weekly review for the first 90 days. If nobody owns it, do not launch.

Three questions separate serious vendors from the rest: what counts as a billable resolution, where your conversation data is stored and whether it trains anyone's model, and what happens to your knowledge base and transcripts if you leave. Our guide to choosing an AI agent platform goes deeper on the evaluation, and healthcare practices should confirm a Business Associate Agreement is available before any patient message touches the system.

There are three ways to get an agent into production, and the right one depends on how unusual your systems are.

Buy a platform and configure it yourself. Fastest and cheapest to start. It works when your help content is in good shape and your systems are mainstream. The hidden price is that you are the one who writes the instructions, tests the flows and notices when an answer is wrong.

Add the agent your current tools already offer. If you run a helpdesk or CRM with a built-in agent, start there. The data is already connected, which is most of the work.

Have it built around your process. This is the path when the answer lives in several systems, when the agent has to write back into them, or when phone, text and email need to behave as one conversation. It is what our workflow automation and CRM automation work covers; most single-system installs go live in two to three weeks, and we follow the phased approach described in our process.

For the phone side specifically, the answered conversation also has to show up where the office works the next morning. Talos Connect, our phone and AI receptionist platform, has built-in connections our team can set up for GoHighLevel, HubSpot, Salesforce, Pipedrive, Zoho CRM, Close, monday.com and Airtable: each caller becomes a contact, and the call follows with its outcome and AI summary. Those connections are one-way, and anything beyond that list is handled as a custom build.

The decision rule: if the agent only needs to read, buy. If it needs to act across more than one system, plan for a build.

Vendors lead with deflection rate, the share of conversations that never reached a person. It is the wrong number, because a customer who gave up counts as deflected. Track these instead:

• Resolution rate. Conversations where the customer got what they needed and did not come back about the same issue within a few days.
• Handoff quality. When the agent escalated, did the person receiving it have the transcript, the customer record and the reason? Ask your staff; they know.
• Repeat contacts. A rise in second contacts on the same issue means the agent is closing conversations it did not solve.
• Time to first response. The plain win. If this is the pain today, our page on improving customer response time with AI covers the fix on its own.
• Wrong answers found in review. Read a sample every week. The target is a short list that gets fixed, not zero forever.

Give it 90 days before judging. The first month is tuning, the second is where the numbers settle, and the third tells you whether to add request types or stop. If after-hours contact is the main driver, AI after-hours call handling is usually the cleanest first deployment because the alternative is voicemail.

An AI customer service agent is software that understands a customer's request in plain language, looks up the answer in your approved content and systems, and completes a defined task such as rescheduling an appointment or opening a ticket. When a request falls outside its rules, it hands the conversation to a person along with the transcript and customer record, so nobody has to start over.

A chatbot follows a script and answers questions. An AI customer service agent can also take action: check an order, change a booking, update a record, or route a ticket, within permissions you set. The practical difference shows up in the handoff. A good agent knows when to stop and passes the full context to your staff instead of sending the customer in circles.

Cost depends on the pricing model. Published list prices in late 2026 include about $0.99 per resolved outcome on one well-known platform and $2 per conversation on another, while voice agents are usually a monthly fee plus minutes. At 400 conversations a month with half resolved by the agent, that arithmetic gives roughly $200 to $800 a month in software, and setup plus weekly review is often the larger early cost.

Rarely, and that is the wrong goal for a small business. An agent takes the repetitive requests so your people have time for complaints, exceptions and sales conversations that need judgment. The small teams we work with use it to cover nights, weekends and busy periods and to stop losing customers to slow replies, not to cut the people customers already trust.

Plan as if the answer is yes. In a 2024 Canadian tribunal decision, Air Canada was held liable for inaccurate information its website chatbot gave a customer. Limit the agent to approved sources, keep policies current, disclose that customers are talking to software, and review transcripts weekly. For specific legal questions about your business, ask your attorney.

Want to know which of your customer requests are worth automating? Book a free consultation — bring two weeks of your real emails, chats and call notes, and we will sort them into request types, tell you which three to start with and what each pricing model would cost at your volume, and tell you honestly if the agent built into a tool you already pay for is enough. If phones are the bigger leak, our voice AI agents page shows how the same approach works on calls.