A chatbot, a workflow and an AI agent can use the same language model while behaving like three different products.
The difference is not intelligence. It is who chooses the next action, which tools are available and how tightly the path is constrained.
NIST launched an AI Agent Standards Initiative in 2026 to support interoperability, security and trustworthy adoption. That attention reflects a practical reality: once a model can act through tools, architecture matters as much as the answer itself.
A chatbot handles a conversation
A chatbot receives a message and returns a response. It may use instructions, conversation history and a knowledge source, but the user normally drives the sequence.
This is suitable for questions, guided intake, explanations and drafting. It is less suitable when a business needs guaranteed steps, durable state and a complete audit trail across systems.
- Best for: questions, discovery, guided forms and internal assistance.
- Main strength: simple interaction and low integration effort.
- Main limit: conversation is not the same as a controlled business process.
Workflow automation follows a defined path
A workflow is a sequence designed in advance: trigger, validation, action, exception and completion. The same input follows the same rules.
AI can be one step inside it. A model may classify an email or extract fields, while ordinary code validates required values, updates the system and records the result.
Many of the tasks in 15 Business Processes You Can Automate with AI need exactly this architecture, not a fully autonomous agent.
- Best for: repeatable work with stable rules and integrations.
- Main strength: predictable behavior, easier testing and clear ownership.
- Main limit: exceptions outside the designed path require a new rule or a person.
An AI agent chooses actions within limits
An agent receives a goal, examines the current state, chooses among available tools, observes the result and decides what to do next.
That flexibility is useful when the exact sequence cannot be known in advance. It also expands the failure surface: the agent may choose the wrong tool, repeat an action, trust unverified content or spend more time and money than expected.
The UK Competition and Markets Authority describes agentic AI as a shift from using tools to delegating outcomes in its 2026 consumer report. Delegation makes limits and accountability essential.
- Best for: multi-step research, exception handling and tasks with changing paths.
- Main strength: adapts the sequence to the situation.
- Main limit: higher uncertainty, security exposure and monitoring cost.
The five questions that separate them
If fixed code can express the path, a workflow is usually cheaper to test and safer to operate. Add an agent only where flexible planning creates measurable value.
- Who selects the next step: the user, fixed code or the model?
- Can the system write to external tools or only suggest an answer?
- Is the full path known before execution begins?
- What happens when the model is uncertain or a tool fails?
- Can every action be explained, reversed and attributed?
Compare cost, risk and maintenance
Chatbot
Usually the quickest to launch. Cost grows with conversation volume, retrieval and human review. The main maintenance work is content quality and response evaluation.
Workflow
Integration work is higher, but execution is predictable. Maintenance focuses on APIs, rules, schemas and exception paths.
Agent
Tool calls, retries, planning loops and observation increase variable cost. Maintenance includes permissions, action logs, limits, evaluation and incident review.
Use a layered architecture instead of one label
A useful business system often combines all three: a chat interface collects intent, a workflow controls the normal path and a constrained agent handles a small set of exceptions.
The model should not own authentication, arithmetic, permissions or irreversible writes. Those controls belong to deterministic services around it.
This is why selecting a model is only one component of the system, as explained in When AI Models Disagree.
A practical decision rule
The UK AI Playbook recommends first deciding whether AI is the right tool and avoiding unsuitable high-risk use cases. The architecture should follow the work, not a fashionable product label.
- Use a chatbot when the user remains in control and the output is advice or a draft.
- Use a workflow when the path is known and reliability matters more than flexibility.
- Use an agent when the path genuinely changes and the benefit justifies stronger controls.
- Use a hybrid when interaction, predictable execution and limited planning all matter.
How I separate these roles in practice
In the multilingual blog system, chat alone was useful for discussing an article but could not reliably preserve its operational state.
Trello and the publication API formed the deterministic workflow. AI handled research synthesis and language adaptation. The agent-like part was limited to choosing tools for reading the queue, preparing assets and checking live pages.
Publication still required explicit authority, while code enforced the checklist. That division grew from the same controlled-pilot principle described in AI Automation for Small Business.
Questions and answers
Is every chatbot an AI agent?
No. A chatbot may only answer messages. An agent selects actions and uses tools toward a goal within defined limits.
Is an agent better than a workflow?
Not automatically. A fixed workflow is often more reliable and easier to audit when the path is known.
Can a workflow use a language model?
Yes. Classification, extraction, drafting and summarization can be model steps inside deterministic control logic.
What is the biggest agent risk?
Giving flexible planning more permissions than the task requires without logs, limits, confirmation and rollback.
Where should a small company start?
Begin with a narrow workflow in draft mode. Add agent autonomy only after exceptions and measurable benefits are understood.