AI agents for customer service: a buyer's guide

What an AI agent for customer service actually is

An AI customer service agent is software that can understand a customer's question, decide what to do about it, and act, either resolving the issue itself or handing it to a person with the context attached. That is the part worth being precise about. A scripted chatbot follows a decision tree you wrote; an agent reasons over your knowledge base and your systems and can take an action, like looking up an order or issuing a refund, within limits you set. The current wave of tools sits on large language models, which is what makes them flexible and also what makes guardrails necessary: a model will confidently make something up if you let it.

The useful way to shop is not tool by tool. It is category first. The four categories below solve different problems, and picking the wrong category is a more expensive mistake than picking the wrong tool inside the right one.

The four categories, and the real tools in each

Every product on the market fits roughly one of these. The vendors named are real and taken from what actually ranks and sells today; we implement across all four categories and resell none of them, so read this as fit, not endorsement.

1. Help-desk-native agents

• What it is: An AI agent built into the support platform you already run, sharing its tickets, macros, and knowledge base out of the box.
• Examples: Zendesk AI agents, Intercom Fin, Salesforce Agentforce, Freshworks Freddy.
• Fits when: you are already committed to that help desk and want the fastest path from knowledge base to deflection with the least integration work.
• Watch for: you inherit the platform's limits and its pricing model, and switching help desks later gets harder, not easier.

2. Standalone resolution agents

• What it is: A best-of-breed agent that sits on top of whatever help desk, CRM, and knowledge sources you have, rather than being tied to one vendor's suite.
• Examples: Ada, Sierra, Decagon, Kore.ai.
• Fits when: your stack is mixed, you want to keep your help desk but upgrade the automation, or you care about controlling the model and the guardrails yourself.
• Watch for: you are now integrating two systems and paying two vendors, and the quality depends heavily on how well it is wired into your data.

3. Contact-center and voice agents

• What it is: Agents built for phone-first support: they handle spoken calls, route to human agents, and assist those agents live during a call.
• Examples: NICE (Cognigy), Genesys, Uniphore, Omilia, SoundHound.
• Fits when: voice is a real channel for you, not an afterthought, and you run a contact center with call volumes that justify the heavier setup.
• Watch for: voice is harder than chat, the projects are longer, and a bad voice deflection annoys customers more than a bad chatbot does.

4. Build-your-own agent platforms

• What it is: Tooling to assemble your own agent, from low-code builders to raw model APIs, when no off-the-shelf product matches your workflows.
• Examples: Botpress and Microsoft Copilot Studio at the low-code end; the OpenAI and Anthropic APIs at the build-it-yourself end.
• Fits when: your support workflows are genuinely unusual, you need deep control over the logic, or you already have engineers who will own it.
• Watch for: you own everything now, including the monitoring, the guardrails, and the on-call. This is the most expensive option to run, not to start.

One honest note on the vendor listicles you will find next to this page: almost every one is published by a vendor that ranks itself first. Kore.ai's guide opens with Kore.ai; Fin's puts Fin at the top; Botpress's list starts with Botpress. That does not make their tools bad, it makes their rankings useless as neutral advice. We have no product in this list to protect, which is the whole reason to read ours.

How to choose: six questions that decide it

Once you know the category, the choice inside it comes down to a short list of questions about your situation, not the tools' feature grids.

1. Which channel matters? Chat, email, and voice are different problems. A tool that is excellent at chat can be weak at voice. Start from where your volume actually is.
2. Where does your support data live? The agent is only as good as the knowledge base and systems it can reach. If your answers live in five disconnected places, fixing that comes before buying anything.
3. How is escalation designed? The most important feature is how cleanly the agent hands a stuck conversation to a human, with the context intact. Test the handoff, not the happy path.
4. What stops it from making things up? Ask exactly how the tool keeps answers grounded in your content and what it does when it does not know. "It uses AI" is not an answer; "it only answers from these sources and escalates otherwise" is.
5. How does it price? Per resolution, per seat, and per conversation bill very differently at your volume. A per-resolution model can punish you for doing well; a per-seat model can hide the real cost. Model your actual numbers.
6. How will you measure deflection honestly? Decide before you buy how you will count a real resolution (closed without a human, and not reopened within a few days), so the vendor's dashboard cannot grade its own homework.

The part the demos skip: escalation, containment, measurement

The demo always works. The three things that decide whether the tool works in production are the three the demo skips.

Escalation design. Customers forgive an agent that says "let me get a person" far more than one that loops. Every conversation needs a clean exit to a human with the full history attached, and a person needs to own the queue those exits land in. Hallucination containment. An agent grounded only in your approved content, that escalates when it is unsure, is safe; one free to improvise is a liability, especially anywhere near billing, returns, or anything a regulator could ask about. Keep it on a short leash and log every answer. Honest measurement. Track resolution rate the way you would audit it, not the way a sales deck presents it. The number that matters is contacts fully closed without a human and not reopened, and it is almost always lower than the headline deflection figure a vendor quotes.

Build vs buy, honestly

We build custom systems for a living, so take this as the opposite of a sales pitch: most companies should buy. The off-the-shelf agents are good, they are getting cheaper, and they already handle the escalation, logging, and channel plumbing you would otherwise build from scratch. Buying gets you to deflection in weeks; building gets you there in months.

Building your own is the right call in a narrower set of cases: when your support workflows are genuinely unusual and no product fits them, when you need control over the model and guardrails that the products will not give you, or when the agent has to reach deep into systems the vendors cannot integrate with. A custom build is priced like any scoped production system, $50,000–$250,000, and that band is a single scoped build; multi-system, multi-year programs run seven figures and beyond, and scale like that is work we take. The cost that surprises people is not the build, though. It is ownership: the monitoring, the guardrail tuning, the on-call for when the agent is wrong at 2am. Every AI agent is wrong sometimes, and a bought one comes with a vendor on the hook for that, while a built one puts you on the hook. Weigh that before the feature list.

Either way, the work of wiring an agent into your real automation and systems is the same kind of job, whether the agent is bought or built. That integration, not the tool choice, is where these projects usually succeed or stall.

Where we fit

We are a senior team, an engineer and an operator, and we do not sell a customer service product or take commissions from anyone who does. What we do is help established companies pick the right category, run a fair comparison against their own tickets, and then implement the choice, buy or build, so it reaches production with escalation and guardrails that hold. That work has a name on our site, AI agent consulting, and it sits inside the broader intelligent automation we do. It starts with a fixed-scope assessment of your support operation at $20,000 to $80,000, so the tool decision rests on your real numbers rather than a vendor's demo.