Every support tool now claims an "AI agent." The term covers everything from a 2019 keyword bot in new clothes to systems that genuinely resolve most incoming questions. In plain language: what these systems actually do, what to demand before trusting one with customers, and what they cost across the market as of July 2026.
Disclosure: we build Clickalong, an AI tutor for SaaS products, so we have opinions. The mechanics below are vendor-neutral.
Three different things wearing one name
- Chatbots follow authored decision trees. Predictable, rigid, and infuriating off-script.
- Copilots help your agents by drafting replies and summarizing threads. The human stays in the loop (Intercom and Zendesk both sell this as a per-seat add-on).
- AI agents talk to customers directly and are expected to resolve, not deflect. This is the category that matters, and the one with real failure modes.
How a grounded agent actually works
The honest version, minus the marketing:
- Your docs, help articles, and approved answers are indexed, typically both as keywords and as embeddings (numeric representations of meaning) so "cancel my plan" matches a doc titled "How billing works."
- When a customer asks something, the system retrieves the most relevant passages.
- The language model is instructed to answer only from those passages, which is what "grounding" means, and to say so when they don't cover the question.
- A well-built agent cites which documents it drew from and escalates with full context when confidence is low.
The quality of that pipeline (retrieval, grounding discipline, escalation judgment) is the entire product. The base model matters less than vendors imply.
Six things to demand
- Grounded answers with citations. If the agent can't show which doc an answer came from, you can't audit it.
- The ability to refuse. "I don't know, let me get a human" is a feature. An agent that always answers is an agent that sometimes lies.
- Escalation with context. The human should inherit the whole conversation, the page the user was on, and what was already tried, so nobody restarts the explanation.
- Guidance, not just prose. For "how do I…" questions, the best answer is often showing: a guided tour over the real UI beats four paragraphs of steps.
- Transparent metering. Know exactly what a "resolution" costs before the invoice teaches you.
- Data boundaries. Your docs shouldn't train someone else's model; visitor PII shouldn't leak into prompts.
The metric that matters (and its trap)
The headline number is deflection, meaning conversations resolved without a human. Measure it honestly: resolved without human touch ÷ all resolved conversations, not "conversations the bot replied to." A bot that answers and gets abandoned didn't deflect anything; it just delayed the churn. Watch CSAT on AI-resolved conversations separately, because high deflection with cratering satisfaction is a silent alarm.
What the market charges
- Intercom Fin: $0.99 per outcome, on top of $29 to $132 seats (the full math).
- Help Scout AI Answers: $0.75 per resolution, on top of $25 to $75 seats.
- Freshdesk Freddy: 500 sessions included, then $49 per 100 sessions.
- Tidio Lyro: from $32.50/month for 50 AI conversations (~$0.65 each).
- Chatbase: message credits by tier ($32/500 up to $400/15,000); advanced models consume multiple credits per reply.
- Clickalong: included in the plan, at 200 answers/month free and 3,000/month on the $49 flat Pro plan. (Ours, and yes, the model is the point: improving your deflection rate shouldn't raise your bill.)
Where this is going
The next step is agents that can see and act, not smarter prose. Clickalong's tutor already reads a sanitized skeleton of the user's current screen to point at the right element, and its tours fall back to AI guidance when the UI shifts underneath them. The agents worth buying in 2026 treat "answer the question" as the floor, and "help the user finish the task" as the job. If that framing matches your product, try it free: 200 answers a month, your real docs, ten minutes to install.