How G2 Names and Classifies CX AI Agents

When it comes to selecting an AI customer service agent, most conversations we have with CX leaders start with "which vendor should we pick?" We think there's an earlier question: which kind of agent fits your team? When we ran our first CX agent evaluation, we sought to establish a benchmark across the variety of agents currently available. The big challenge was how to establish standards across agents built-in to the software you’re currently using versus those available off-the-shelf or configured onto an AI model. With the proliferation of vendor marketing copy, buyers need neutral labels to compare products. We realized the value in naming and classifying types of agents to help buyers weigh the multitude of factors involved in selecting an agent.

This note lays out the four ways companies get a CX agent today, what we call them, and what to weigh before you choose one.

Agentic AI for customer service budgets are on the rise

Buy or build your AI agent? It’s a common dilemma for AI and software adoption. Picture a retail CX leader whose team runs a help desk SaaS solution. Her reps stay less than a year on average, and holiday volume means growing the team as much as five times. She has budget for an AI agent and three options in front of her. She can turn on the agent her help desk vendor sells, bring in a dedicated agent product, or build one on a platform or model.

According to G2’s 2026 AI Agents Insight Report 8 in 10 B2B software buyers said their companies bought AI agent software or services in the past year, up 25% from 2025. Two-thirds have agents in limited or scaled production. Among organizations using agents, customer support is second (35%) after software development (41%) as the most-used agent category.

In other words, many organizations choose to buy versus build. Only 12% build mostly in-house, about the same as last year. The other 88% work with vendors, through agent-builder platforms (42%), ready-made agents (27%) or a mix (19%).

Budgets are growing too. In five of the six industries G2 tracks, more companies now spend at least $1 million a year on agents, and customer service was the top use case in G2's 2025 study.

Four types of AI customer service agents

After speaking with dozens of CX agent vendors and buyers and reviewing terminology in the market, we’ve settled on four classifications for the agents in the marketplace.

By "harness" we mean everything around the model that makes it useful at a support desk: connections to your systems, the actions it can take, guardrails, and handoff to a human. The question each class answers is who builds that harness and who keeps it running.

Why we chose these names

In our discussions with CX agent vendors and CX leaders, we learned that the word "agent" means different things to different buyers. A program lead at a bank who built her own agent uses it for “a tiny thing that does a very specific task,” but says that “when people use the word agents, they mean they’re ready off the shelf the way that they’re being advertised right now.”

Vendors describe themselves with a range of terms. The top three agents on the G2 AI CX leaderboard (as of October, 2026) include Retell, an “AI platform,” Live Agent, an “all-in-one customer service platform,” and Chipp, a “no-code platform.”

Zendesk sells AI agents inside its own help desk and markets Forethought as "Standalone AI agents for any platform." Fin is “natively” integrated into its own and other CX systems and works “natively” across support channels. Sierra positions itself as a “conversational AI platform” and Decagon skips agent terminology for “AI concierge.”

Analysts and the media have positioned agents with a variety of terms. In its report, Seizing the agentic AI advantage, McKinsey interchanges “agents embedded in software suites” with “off-the-shelf agents” and distinguishes these from “custom-built agents.” Bain’s Building the Foundation for Agentic AI labels them “custom agents on engineering tools,” “prebuilt agents embedded in vendor platforms” and “dynamically generated agents in data platforms.” CX Dive’s Agentic AI Trendline uses the broader terms “AI platforms” and “Agentic AI systems” to describe the range of AI agents used for customer experience.

Brands can and should name their products however they want. Our job is labels that mean the same thing to every buyer and every vendor on the leaderboard.

How to consider agents for your organization

Built-in Agent

Your help desk holds years of workflows, macros, routing rules, ticket fields, tags and customer history. A built-in agent works inside your existing help desk from day one. There's no new vendor to onboard, no new security review, and your team keeps working in the tool they know.

The Head of Customer Experience at a fitness chain shared, “Without integration to our membership platform, the AI agent can't do anything except surface FAQs. It has to connect to the system that drives our business.”

With a built-in agent the setup takes more effort up front. You usually have to codify your policies in the format the vendor prescribes, and that takes time from people who know your operation well. Our working view is that this effort pays off over time. When the agent runs on precise rules and the workflows your team already refined, it tends to handle complex policies better as volume grows. The trade-off is that you're tied to your help desk vendor's AI roadmap. This agent may be best for teams standardized on one help desk who plan to scale the agent over years, not months.

Dedicated Agent

Also known as standalone or off-the-shelf, these agents plug into your help desk or run on their own. These favor speed of implementation over integration. Many dedicated agents get you running quickly from uploaded documents and policies, and they're built only for CX. Most work across help desks and channels, so you're not locked into one system. Some, like Sierra and Decagon, assign their own engineers to handle complex deployments.

The risk is a plateau in performance. A fast start can hide the nuanced rules nobody wrote down, and without that work, performance can level off over the long term. You also take on another vendor and another integration. Of buyers we spoke with, one dropped a vendor over a three-hour sync with their existing system, and another was quoted $100,000 for an integration. Consolidation matters too, with Fin now part of Salesforce and Forethought part of Zendesk.

Agent Builder

A builder gives you the harness and lets you design the agent yourself. That means control over workflows, data and governance, and it's why builders are now the most common path in G2's research.

Setting one up is a project. Someone has to define the use cases, connect your help desk and back-office systems, write the instructions and policies, build the workflows, test against real tickets, and deploy to your channels. Many builders are no-code, but you still need skilled people in-house. Usually that's a CX operations or automation lead who understands your processes, plus an admin or developer for integrations. Some builders sell implementation services or work through partners if you don't have that capacity.

After launch, the builder's support covers the platform, not your agent. When a policy changes, a new product launches or quality drifts, your team updates it. The 2026 G2 AI Agents report puts the bar well. Builders need "the connectors, actions, guardrails, and implementation support that turn a toolkit into a dependable enterprise capability."

In-house agent

Building on a foundation model gives you full control and keeps you current as models improve. It also means you build the whole harness yourself. That takes engineers, not just operations staff. You'll need people who can build integrations, retrieval, guardrails, human handoff, monitoring and testing. In practice, you're running your own forward-deployed engineering team. There's no vendor to call when something breaks. Your team owns uptime, model upgrades, cost control and quality checks for as long as the agent runs. That's a big reason only 12% of buyers take this path.

In our foundation model runs, inexpensive models handled many tasks well but fell behind products on policy compliance. As Frankie Li, our VP of Data Science, put it, "The gap we see is not the models, but in the harness."

A CX leader at a global payments company showed what building looks like at scale. Her team runs about 40 subagents, each handling a different scenario, many shaped by regulation in each market. They test new models as soon as they launch and run their own evaluations before switching. Cost can still favor building. An IT services leader told us two dedicated agents met every functional requirement, but building their own on a general-purpose model cost much less over two years.

How to choose

Start with your stack. Every buyer we interviewed treated the agent as one part of a larger system rather than a standalone purchase. Half of them said that an agent is only as useful as the data it can reach and the place its work lands. A senior customer care leader at an ITSM software vendor expects the winners to be the agents most connected to the tools a company already uses. Test sync speed and integration cost before you commit.

Then map your risk and decide who owns the agent. One executive described their rule: "Everything informational is autonomous. Everything transactional needs a human validating before action." Most teams expand autonomy in stages, from drafts to routine answers to small tasks that can't cause harm, and only then to real transactions. Whether your help desk vendor, an agent vendor, your ops team or your engineers build it, someone has to own the agent after launch.

Finally, price the oversight and ask for proof. Cost is in the eye of the buyer. When comparing a six-month setup resolving 50% of tickets versus a one-month setup resolving 30%, views on pricing split. A GRC software leader said paying per resolution is the only model that makes sense, while a customer care director told us, "Per resolution puts me in a position where the better I configure the agent, the more I pay."

Every buyer we interviewed tested in their own context before committing because that’s where the problems showed up. A program lead at a bank said demos alone wouldn’t have convinced her operations teams to sign a contract, so her team tested the agents on real questions first.

Another buyer ran two agents on 1% of traffic for three to six months. As one Head of Customer Experience put it, "I'd trust the numbers, but I'd ask to speak to one of their customers, ideally without the vendor present."

What this does not tell you

Cost per conversation isn't on the leaderboard yet. The long-term scaling and plateau patterns above reflect what we hear from buyers and vendors, and we're designing longer-running scenarios to test them. Agent Builder results depend on how the agent was set up, and we'll share how we configure builders. Many vendors span more than one class, so we label each by its main product.

Check the evaluations and share your feedback

When it comes to the best AI agents for customer service, the spread within each class is wide. Two agents with the same label can perform very differently. Scores change every quarter, as the models and agent products continue to evolve. A few patterns have held up. Strong results show up in every class. That's why we recommend choosing the class that fits your situation first, then comparing vendors within it.

A few questions we're still working through, and we'd like your take. Does a built-in agent's headstart on your workflows beat a dedicated agent's speed over a full year? How much setup should an agent builder get before we score it? And should agent classes change how the leaderboard ranks agents, or only how you filter it?

Class labels are coming to the CX leaderboard and agent profiles, and the next CX evaluation runs in November. If these names don't match how you think about the choice, tell us. Comment below or write to agent-evals@g2.com.