What is an AI agent?
Plain-English definition · Updated 2026-10-07. Numbers dated; verify with the vendor.
AITop is an independent guide. Prices and features change fast — always check the vendor's page.
An AI agent is a language model that has been given tools and a loop — so instead of answering one question, it can decide to look something up, call something, check the result and try again until a stopping condition is met. The practical line is action: a model that reads and summarises is a feature; a model that can cancel your order is an agent. In 2026 the word also stopped being only a description and became a price tag: major AI support and coding products now bill per completed "outcome" or "resolution" at roughly $0.50–$2.00 each, which makes the definition of "done" the most expensive sentence in the contract.
What it actually means
Take a chat model and give it three extra things and it becomes an agent. Tools: a set of callable functions — search, a database query, an API that moves money — described in a way the model can read. A loop: the model's output is fed back in as the next input, so it acts on what it just learned instead of producing one final answer. And a stopping condition: the rule that says the task is finished, or that the agent has failed and should hand over.
That third part is the one buyers underestimate. A chat model always terminates — it answers, and you are done. An agent decides for itself when it is finished, which means someone has to define "finished" in advance, and that definition is where cost, quality and risk all land. It is also why "agent" is a spectrum and not a checkbox: the same underlying model is a summariser, a copilot that drafts a reply for a human to approve, or an autonomous system that closes a customer's ticket, depending entirely on its tools and its stopping condition.
Two labels are worth keeping separate. Agent-assist puts the model beside a human — it reads the thread, drafts the answer, and a person sends it. Autonomous agent lets the model finish the job by itself, and escalates only when its stop condition says it cannot. Vendors use one word for both, and the price difference between them is roughly an order of magnitude.
Why it matters when you're picking a tool
Because "agent" is now doing the work that "AI-powered" did in 2023: it appears on every landing page and distinguishes almost nothing. Three questions cut through it, and they are the same three every time.
What is it allowed to do? Reading and acting are different products. An agent that summarises your tickets is low-risk and low-value; an agent that can issue refunds or delete records has a blast radius, and you should expect permissions, audit trails and a sandbox to be part of what you are buying. Where does it stop? Ask for the escalation rule and the confidence threshold in writing — "it hands off when unsure" is a marketing sentence, not a testable one. What is it billed against? This is the question that has changed most: the industry spent 2026 moving agents off seat-based pricing and onto per-outcome pricing, so your bill now scales with the agent's success rate rather than with your headcount. That is a genuinely better shape for a growing team, and a worse one for a high-volume operation that is already stable.
The 2026 reality check
The connector problem got solved, which changed what you can buy. Anthropic donated the Model Context Protocol (MCP) to the Linux Foundation's Agentic AI Foundation in December 2025, with OpenAI, Google, Microsoft, AWS, Cloudflare and Bloomberg among the supporters. By March 2026 the protocol was reporting roughly 97 million monthly SDK downloads and 10,000+ published servers. The practical effect for a buyer is that the tool surface stopped being proprietary: a growing number of agents from different vendors can now plug into the same server, so "which tools does it connect to" is a weaker reason to pick one vendor than it was a year ago.
The word turned into a meter. Roughly in the space of one year, per-outcome pricing went from a curiosity to the default model in AI customer service. Analysed in date order: HubSpot cut its customer agent to $0.50 per resolved conversation on 2026-04-14; Gorgias charges $0.90 per resolved interaction on annual billing ($1.00 monthly); Fin (formerly Intercom) charges $0.99 per outcome; Zendesk bills $1.50 committed or $2.00 pay-as-you-go for automated resolutions beyond a small allowance. Same structure, four different names — and none of those names means the same thing.
The performance numbers are vendor-reported, and they cluster. HubSpot publishes a 65% resolution rate across the 8,000-plus customers who switched its agent on. Fin's average is quoted at around 76% of support volume resolved end to end. Tidio advertises that Lyro handles up to 67% of repetitive queries unattended. These are not far apart, which is reassuring — but they are self-reported, measured on each vendor's own definition of a resolved conversation, and there is no equivalent of an independent crash-test score. Treat them as a vendor's claim about a typical workload, not as a specification.
And the category consolidated fast. Salesforce closed its roughly $3.6 billion acquisition of Fin on 2026-09-10, moving the product into Salesforce AI Labs with prices unchanged at the close. Zendesk completed its purchase of Forethought on 2026-03-26 and resells it as Advanced AI Agents through sales. Two of the most respected agents here changed owners inside seven months. Worth knowing before you build a workflow onto one.
Quick checklist
- Separate "reads" from "acts." If the agent cannot change anything in your systems, you are buying a search box with better manners — price it accordingly.
- Get the stop condition in writing. What counts as done, what counts as failure, and what happens on the boundary between them.
- Ask who defines a "resolution." On per-outcome contracts this one word sets your invoice. A customer who stops replying is not necessarily a customer whose problem was fixed.
- Check the tool surface before the model. MCP-compatible means you are not locked to one vendor's agent; a proprietary connector list means you are.
- Demand the reporting, not the claim. "Up to 65–79% resolved" is a marketing range. Ask for your own automation rate after 30 days and a way to audit what the agent closed.
Where you'll hit it
Agents are the reason three of our pages read the way they do: the best AI customer service tools ranking (where "agent" is literally the unit you are billed in, and the definition of resolved is the whole argument), the best AI coding assistants of 2026 (where an agent edits files rather than answering questions), and Claude Code vs Cursor (the clearest example of the assistant/agent split — one drafts while you watch, the other is meant to be handed a task and left alone). Related terms: AI token — the fuel every agent burns, and why reasoning multiplies it — and context window, which is the hard ceiling on how much an agent can hold in mind while it works. Full list on the AI tools glossary.
The bottom line
An agent is a model plus tools plus a loop plus a stop condition — and only the last two decide whether it is useful or expensive. In 2026 the market answered the "is it real?" question by starting to bill for it: $0.50 to $2.00 per completed outcome, seat prices quietly retreating to second place. So the useful skill is no longer spotting which vendors are bluffing. It is reading the definition of done, because that sentence is the price.