The distinction in one line
A chatbot produces a response. An agentic system produces an outcome. If your problem is "people cannot find information", you need the first. If it is "people spend hours doing repetitive multi-step work", you need the second — and they cost very differently.
The terminology is being used loosely enough that buyers cannot tell what they are being sold. Here is a practical way to distinguish them and decide which you actually need.
What separates them
| Chatbot / assistant | Agentic system | |
|---|---|---|
| Output | An answer | Completed work |
| Steps | One exchange | Plans and executes many |
| Tools | None, or retrieval only | Reads and writes to real systems |
| Self-checking | No | Verifies output before proceeding |
| On failure | Says it cannot help | Retries, replans, or escalates |
| Human role | Reads the answer | Approves at defined checkpoints |
| Measured by | Answer quality | Task completion rate |
The same request, handled two ways
Take a refund request arriving by email.
Chatbot
- Reads the message
- Explains the refund policy
- Tells the customer what to do next
- A person still processes the refund
Agentic system
- Reads the message and identifies the order
- Checks eligibility against policy and order data
- Verifies its own conclusion against the rules
- Pauses for one-click human approval
- Issues the refund and confirms to the customer
- Logs the whole chain
The chatbot saved the customer a search. The agent saved your team the job — and paused before doing anything irreversible.
What agentic systems require that chatbots do not
- Tool design — narrow, validated operations against your real systems. Badly scoped tools are the most common cause of unreliable agents.
- Planning — decomposing a goal into verifiable steps, with the ability to re-plan when reality differs.
- Verification — checking output against rules before it proceeds or is delivered.
- Approval gates — irreversible actions pause for a human, calibrated to your risk tolerance.
- Observability — traces showing what ran, what it used, and why it decided what it did.
- Failure handling — retries with feedback, and clean escalation when retries do not help.
This is why the cost difference is real and unavoidable. A chatbot that gives a wrong answer is embarrassing. An agent that takes a wrong action has changed something in your business. The engineering above is what makes that difference acceptable — and skipping it produces a system nobody will authorise to run.
Scoring your own workflow
Score each — three or more "yes" answers suggests an agentic system is worth costing:
- Does the work involve multiple steps across more than one system?
- Is it high-volume — daily or many times weekly?
- Can the rules be written down, including the common exceptions?
- Does it currently consume meaningful staff hours?
- Is the input reasonably structured?
- Would completing it correctly be verifiable by a machine?
If most answers are no — particularly the third and sixth — you likely need a grounded assistant rather than an agent, and you should be pleased about that. It is cheaper and lower risk.
Where agentic systems genuinely pay back
| Workflow | Why it fits |
|---|---|
| Document extraction and filing | Structured, verifiable, high volume |
| Reconciliation across systems | Rule-based, tedious, error-prone manually |
| Recurring report compilation | Mechanical multi-source gathering |
| Lead qualification and routing | Consistent rules, works after hours |
| Support ticket triage and drafting | Human approves rather than composes |
| Onboarding sequences | Many small steps across several systems |
Start narrower than feels ambitious. One workflow, human approval on everything initially, autonomy widened only as the logs prove reliability. Teams who deploy broad autonomy immediately either scare their stakeholders into shutting it down, or discover an expensive mistake — both of which set the programme back further than starting small would have.
A practical recommendation
Most organisations should build a grounded assistant first. It is cheaper, delivers value quickly, and — importantly — teaches you where your documentation and data are weak. Those weaknesses would have undermined an agentic system anyway.
Move to agentic once you have a specific multi-step workflow with measurable hours attached and rules you can write down.
Not sure which your workflow needs? Describe what your team repeats — we will tell you honestly if a simpler assistant would do. See our agentic AI service, AI agent service, and back-office workflows worth automating.