Everyone calls their chatbot an "AI support agent" now. It's become a label you slap on a widget that answers questions from a knowledge base. That's not wrong, exactly — answering questions is part of the job. But it's one job out of four, and the other three are where support actually gets resolved instead of just acknowledged.

We think an AI support agent worth the name has to be able to do all four: answer, show, act, and escalate. Miss one and you've built a smarter FAQ, not an agent.

Job one: answer, but with the right scope

Answering is the easy part now. Any model with your docs in context can produce a plausible response. The hard part is knowing what it doesn't know and saying so instead of guessing. A support agent that confidently invents a refund policy or a pricing tier is worse than no agent at all — it just moves the damage from "unanswered ticket" to "angry customer with a wrong answer in writing." We've written before about why most AI customer support tools are optimized to lie to you — the incentive to always sound confident is baked into how most of these tools are graded.

The fix isn't a bigger model. It's grounding answers in documentation that's actually current, and building the humility to escalate when the docs don't cover the case. That's the whole premise behind the Brain: documentation that flags itself as stale or contradicted and rewrites itself from real conversations, instead of quietly going out of date while the agent keeps citing it.

Job two: show, don't just describe

A huge share of support tickets aren't knowledge problems, they're navigation problems. The user isn't confused about what a feature does — they're confused about where it is. Telling someone "go to Settings > Billing > Payment Methods" in a chat window is asking them to translate your words into clicks on a screen they're not looking at while they read.

An agent that can actually see the product should just show them: highlight the button, walk the cursor to the field, point at the toggle that's currently off. That's the difference between Show and a chat transcript with instructions in it. It's a small design shift with a big effect on resolution time, because you've removed the translation step entirely.

Job three: act, with the user's hand still on the wheel

This is the one most tools skip entirely, because it's the one with real risk attached. Answering wrong is embarrassing. Acting wrong — canceling the wrong subscription, refunding the wrong invoice, deleting the wrong seat — is a support ticket about your support agent, and those are the worst kind.

But skipping it means every "how do I" ticket that could've been a self-serve action instead becomes three more back-and-forth messages while a human confirms details that were already sitting in your database. Do exists because we think the right answer isn't "never let the agent touch anything," it's "let it prepare the action and show the user exactly what's about to happen before it happens." Confirmation isn't a compliance checkbox, it's what makes acting safe enough to actually ship. We went deeper on this distinction in AI customer service should do things, not just say things.

Job four: escalate like it means it

The fourth job is the one that gets treated as a failure state instead of a feature. It isn't. An agent that escalates cleanly — with the full conversation, the account state, what it already tried, and why it's punting — saves more support-team hours than an agent that never escalates and just loops the customer through the same three canned answers until they give up and email support anyway.

A good escalation is judged on:

  • How much context the human gets without having to ask the customer to repeat themselves
  • Whether it happens at the right moment, not after five frustrated messages
  • Whether the handoff includes what was already tried, so the human doesn't retread it
  • Whether the same gap gets fed back into the documentation so it doesn't happen twice

That last point matters more than it sounds. An agent that escalates without feeding the miss back into its own knowledge is going to have the exact same conversation next week. This is also where stuck-point analytics earn their keep — they tell you which questions are actually escalation-worthy versus which ones the agent just hasn't been taught yet.

Why this framing matters when you're evaluating one

If you're comparing tools, ask each vendor to show you all four jobs, not just a demo of the chat window answering a question well — any of them can do that. Ask to see it point at a real button in a real product, prepare an account action with a confirmation step, and hand off a conversation with context intact. We laid out a fuller checklist in how to judge an AI support agent before you buy one, and if you want the specific comparison, Peeve vs Intercom Fin walks through where the four-job framing actually plays out in practice.

The category name "AI support agent" has gotten loose enough to mean almost anything. We'd rather it mean something specific: a system that answers, shows, acts, and escalates — in that order, with confirmation before anything real happens. Anything less is a chatbot with better branding.