AI Agents vs Chatbots: What Operators Should Know (2026)

A chatbot answers from what it was indexed on — help articles, a FAQ, maybe a scraped PDP. An AI agent uses tools against your systems, cites what those tools returned, and can be stopped before money moves. If it cannot use your systems, and cannot be stopped, it is a demo with a new label.

On June 17, 2026, AgentCore Harness reached GA (What’s New). Hosting a loop got cheaper. That did not turn a help-center widget into an agent. Agents Classic is in maintenance for new customers after July 30, 2026 — the wrong net-new path either way.

This is part 3 of AI Agents for Business. It is not a rewrite of AI agent vs workflow automation. That post is who owns the state machine. This one is chatbot vs worker.

First-party signals we reuse (not client KPIs) — Gateway ~180 ms → ~95 ms median tool round-trip on a B2B CRM assistant (12 tools, ~8k turns/day) — Gateway post. Support-style AgentCore 50K sessions/mo ~$791/mo platform + model (decision guide). Browser left on for lookup-only turns: roughly 3× Runtime-shaped spend — support field-guide post.

Reproduce this — Compare your current widget to the allow-list in support-agent-tool-policy.md. If the only “tool” is search, you have a chatbot. Series folder: ecommerce-ai-agents-series/. HITL rules: human-in-the-loop post.

Opinionated take: do not relabel a chatbot as an agent to win a budget meeting. Trade-off: the honest chatbot ships faster and will not touch OMS writes. The agent takes longer because tools, identity, and a stop button are the product.

The distinction that matters in a meeting

The hub states it in one line: a worker with tools — and a stop button.

FactualMinds is an AWS Select Tier Services Partner. We have zero published AI-agent case studies. A chatbot testimonial from a GenAI engagement is not agent-delivery proof.

Two comparisons people mash together

1. Chatbot vs agent (this post) — search vs tools + stop.
2. Agent vs workflow (field-guide post) — who executes tax, capture, labels. Hybrid default: the workflow executes; the agent returns a structured decision.

If you only remember one substitute: App-style FAQ widget → keep it as search. Do not bolt createRefund onto it to “make it an agent.”

What broke — Help-center search wired to Admin API keys so the slide could say “agent.” Detection: the widget either could not act (buyers angry) or it could (refunds without Policy). Recovery: split the surfaces. Search stays search. The agent gets a tool catalog and HITL. The Browser-hot ~3× spend is what happens when you leave an expensive tool on to look like agency.

Named substitutes

• Question already in the help center → chatbot / site search. No agent.
• Question needs order + shipment facts → support agent, week-one reads — how agents reduce support work.
• Money already has a correct state machine → workflow, not either UI — agent vs workflow.
• Evaluating vendors who say “our chatbot is agentic” → ask for the tool list, the stop rule, and the audit trail. If those are slides only, it is a chatbot.

If You Only Do One Thing

Write two sentences on a slide: “Our chatbot searches N articles. Our agent may call these M tools and must stop before refund / price / PO.” If you cannot fill M and the stop, you do not have an agent yet.

What to Do This Week

1. Inventory every “AI” surface on the storefront and in admin. Label each chatbot or agent using the table above.
2. Remove write credentials from anything labeled search.
3. If a real agent is in scope, open the hub and pick one family. Do not start a sixth “chatbot 2.0” project.

What This Post Doesn’t Cover

It does not retell Cedar, Harness, or Strands. It does not claim a conversion lift from replacing a widget. We have not A/B tested chatbot vs agent for a named merchant in a published study; the numbers above are platform and field-guide benchmarks. The evaluation table is AI agent vs chatbot.

Primary next step: AI Agents hub. For the table: AI agent vs chatbot.