AI assistant, automation and agent
A basic AI assistant typically responds to instructions and produces content. Ordinary automation follows defined rules, such as sending a reminder three days after a due date. An AI agent may use AI to decide which steps or tools are needed to reach a goal. In practice, products can combine these approaches, so ask what the system can actually do rather than relying on the label.
A chatbot that drafts a customer reply is different from a connected system that sends it, updates the customer's record and changes an appointment. The second system has authority to change the business, not just suggest wording.
Where an agent might help
Suitable supervised pilots might include organising approved meeting notes, drafting follow-up tasks, categorising fictional support enquiries or compiling a review of records the business has authorised it to access.
Before giving an agent access to email, customer information, finances or shared documents, consider whether an ordinary workflow or a restricted assistant would achieve the same outcome with less risk.
Unlock the rest of this lesson
You're halfway through. Enter your email to read the full lesson, use the exercises and quizzes, and unlock the rest of The Business of Growth library.
Already signed up through The Business of Growth? Use the existing signup to return and unlock this browser. No payment required.
Worked example: the service inbox
Fictional example. A maintenance business receives requests about new bookings, cancellations and job progress. It wants an agent to handle the inbox.
A first pilot lets the agent classify fictional messages and draft responses without sending anything. Staff review whether the categories and proposed actions are correct. A second, separately approved stage might connect to a limited knowledge source and place drafts in a review queue. It would not autonomously cancel jobs, promise dates or access payments.
The team records errors, missed exceptions and how long approvals take. If the agent fails to recognise urgent issues, the business should stop that use or add stronger controls rather than simply increase its autonomy.
Four controls I would require
- Least privilege: Access only the systems and records required for the task.
- Human approval: Confirm consequential changes and external communications before they happen.
- Traceability: Keep records of the instructions, information used, proposed actions, approvals and changes made, subject to privacy rules.
- Stop and recovery: Provide a way to pause the system, revoke access and correct mistakes.
Also test how the agent handles malicious or misleading instructions inside emails and documents. Content it reads may contain text trying to redirect its behaviour. Treat external material as untrusted and test failures before deployment.
Why 'set and forget' is a poor business plan
Systems, policies, pricing and customer needs change. An agent that worked on one workflow may fail after a small change to forms or permissions. Someone needs to monitor results, update the rules and remain responsible for the decisions.
Your next action
Your notes stay on this page unless you choose device saving, download them or submit them to Greg below.
Keep names, private customer details and confidential information out of your notes. Device saving is optional. Only the separate Send my lesson work button below sends anything to Greg. There is no account or cross-device sync.
Two questions before you move on
These checks support learning. They do not assess business health, award a qualification or send a result to Greg.
Ask Greg
Ask a question about this lesson or a challenge in your own business. I read submissions and may reply or cover common questions in future lessons. I cannot promise an individual reply to every question.
Lesson: AI Agents for Small Business: What Can They Actually Do?. Please do not include confidential client, financial, health or staff information.
Questions this lesson answers
Can an AI agent run parts of my business?
It may assist with specific authorised workflows. It cannot take legal or commercial accountability away from the business, and broad unattended control creates additional risk.
Is an AI agent different from a chatbot?
Often. A chatbot may only answer questions, while an agent may plan steps and use tools. Check the product's real capabilities and permissions, since marketing terminology varies.
Work out where your business is getting stuck
The free Growth Diagnostic asks ten questions across acquisition, conversion, relationships, delivery and commercial control. It gives you an area to investigate and three things you can do next.
One complete sample lesson in each stream is fully open. Read halfway into every other lesson, then enter your email to unlock the remaining content, exercises and quizzes across the library. Ongoing marketing emails are optional.