An AI agent is a program that uses a language model to decide what to do next: which tool to call, what information to look up, when a task is complete. Unlike a chatbot that answers one message at a time, an agent works toward a goal across many steps. That power is exactly why agents need clear boundaries.
Agents vs chatbots vs automations
- A chatbot responds to a message. It has no tools and no goal beyond the current reply.
- An automation follows a fixed sequence of steps decided in advance.
- An agent is given a goal and a set of tools, and decides the sequence itself — searching, reading, calling APIs, checking results and retrying.
Most real "AI agents" in business are somewhere between automation and agent: the overall flow is fixed, but individual steps allow the model to reason and choose.
Anatomy of a business agent
- Goal and instructions — what the agent is for, what it must never do.
- Tools — functions it can call: search the CRM, read a document, create a ticket, send a draft for approval.
- Memory and context — the information it needs for the task, retrieved from your systems.
- Control loop — the code that runs the model, executes tool calls, enforces limits and stops.
- Observability — logs of every step so you can inspect why the agent did what it did.
Types of agents businesses actually deploy
Research agents that gather and summarise information; sales assistants that prepare account briefs and draft follow-ups; support agents that resolve known issues and escalate the rest; operations agents that reconcile data across systems and flag discrepancies.
Guardrails that matter
Give agents read access widely and write access narrowly. Require approval for anything irreversible. Cap steps and spend. Keep a human in the loop where the cost of an error is high.
Where I stand
Agents are the most exciting and the most over-sold area of AI. The winners will be teams that pick narrow, high-volume goals and instrument everything. A boring, reliable agent that saves two hours a day beats an impressive demo every time.
Business use cases
- Sales research agents that prepare account briefs before calls
- Customer support agents handling known issues with escalation paths
- Internal operations agents reconciling data between systems
- Research and monitoring agents summarising markets, competitors or regulations
- Personal productivity agents that draft, schedule and follow up (with approval)
- Data-quality agents that detect and propose fixes for inconsistent records
Limitations to be honest about
- Agents can loop, stall or pursue the wrong sub-goal; step and budget limits are essential.
- Every tool you expose is a potential action the agent can take incorrectly.
- Latency and cost grow with the number of steps.
- Debugging requires full trace logs; without them you cannot explain behaviour to a stakeholder.
- Agents amplify the quality of the underlying data — poor data yields confident nonsense.
How I approach it
- I begin with a single, measurable task and design the tool set around it, with read-mostly permissions. The agent runs in "suggest" mode first, producing recommended actions that a human approves. Only when its suggestions are accepted at a high rate do we allow selected actions to execute automatically.
Frequently asked questions
What is the difference between an AI agent and an AI chatbot?
A chatbot answers a message. An agent pursues a goal across multiple steps, using tools such as search, databases and APIs, and decides what to do next.
Are AI agents safe to use in a business?
They can be, if scoped correctly: limited tools, approval for irreversible actions, step and cost caps, and complete logging.
How do you build an AI agent for a business?
Define one task, list the tools it needs, build the control loop with limits, run in suggestion mode, measure acceptance, then automate gradually.
Do agents need a lot of training data?
No. Most business agents use a foundation model with well-designed instructions and tools. Examples help refine behaviour but training is rarely required.
Last updated 11 September 2026