AI Agents vs Chatbots: Most People Are Buying the Wrong One
The terminology has been deliberately muddied by vendors. Here's how to cut through it, identify what you actually need, and avoid paying agent prices for chatbot capability.
Quick Verdict
90% of what's sold as "AI agents" in 2026 is a chatbot with a better prompt. A real agent takes autonomous action — it doesn't just answer questions, it executes multi-step workflows, makes decisions, and uses tools. If yours can only chat, it's a chatbot. That's fine, but don't overpay for the rebrand.
Side-by-Side Comparison
| Factor | Chatbot | AI Agent |
|---|---|---|
| What it does | Answers questions, follows scripts | Takes actions, makes decisions, uses tools |
| Intelligence | Pattern matching + retrieval | Reasoning + planning + execution |
| Integration depth | API calls to fetch info | Full read/write access to your systems |
| Autonomy | None — responds when prompted | Can initiate tasks, monitor, escalate |
| Error handling | Fails or escalates | Retries, adapts, asks for help strategically |
| Setup complexity | Hours to days | Weeks to months |
| Cost | £50–500/mo (SaaS) or £2–8k (custom) | £5–30k build + ongoing orchestration |
| Best for | FAQ, lead capture, basic support | Process automation, decision-making, operations |
The "Agent-Washing" Problem
In 2024, dozens of SaaS vendors renamed their chatbot product an "AI agent" and raised their prices. The tell is simple: ask whether it can take actions you didn't explicitly programme. If the answer is no — if it can only retrieve information and respond to prompts — it's a chatbot.
A genuine agent has tool use. It can write to your CRM, send emails, query databases, trigger other workflows, and make branching decisions based on what it finds. If your "agent" is only reading and responding, you're paying premium pricing for a chatbot. We've documented how real AI agent architectures work under the hood.
When a Chatbot Is the Right Choice
Customer FAQ, lead qualification forms, appointment booking, and basic support routing don't need reasoning or autonomy. A well-built chatbot here costs 10x less and is 10x more reliable than an agent because it follows rules rather than reasoning — and reasoning can go wrong.
If your use case has predictable inputs, predictable outputs, and no branching logic that requires context, a chatbot is better. Simpler systems fail in simpler ways. That's a feature, not a limitation.
When You Actually Need an Agent
Multi-step processes with branching logic are the clearest signal. Processing a refund is a good example: it requires checking order status in the OMS, verifying eligibility against business rules, cross-referencing inventory if the item needs returning, checking customer history for fraud signals, then initiating the refund and logging the action. No chatbot handles that end-to-end.
Any task requiring access to multiple systems with write permissions — not just fetching data but changing records — needs an agent architecture. So does anything where the "right" action depends on context that changes: customer tier, order value, regulatory requirements, time-sensitive conditions.
The Reliability Gap
Agents are powerful but less predictable than chatbots. A chatbot follows rules. An agent reasons — and sometimes reasons incorrectly. The failure modes are qualitatively different: a chatbot gives a wrong answer, an agent takes a wrong action.
This is why most production agent deployments include human-in-the-loop checkpoints for high-stakes decisions: approval workflows, escalation paths, audit logs of every action taken. An agent without oversight is a liability. Budget for the human oversight layer when you're scoping agent projects. Our RPA vs AI agents comparison covers where agents sit in the automation spectrum.
Our Recommendation
Choose a Chatbot When...
- • Use case is FAQ, support routing, or lead capture
- • Inputs and outputs are predictable
- • You only need to read data, not write it
- • Budget is under £5k
- • You need it reliable and fast to deploy
Choose an Agent When...
- • Process has 3+ steps with conditional branching
- • It needs read/write access to multiple systems
- • The "right" answer depends on changing context
- • Exceptions need to be handled, not just escalated
- • You can build and maintain oversight processes
Not Sure Which One Your Use Case Needs?
We'll scope it honestly — whether that means a chatbot, an agent, or both working together.
Talk to Our Team