What is an AI agent? A plain-English answer for small business owners
What is an AI agent? A plain-English answer for small business: how it differs from a chatbot or automation, what it can't do, and who answers for it.
An AI agent for a small business is software that is given a goal and a set of rules, then plans the job, uses your existing tools and checks its own work, while a person approves anything that reaches a customer.
That last part is our rule, and the technology doesn't require it. Keep that in mind when someone is selling you one.
What the companies building them say
Anthropic, which makes the Claude models, says "there is no agreed-upon definition of what an agent is". So when two vendors say "AI agent", they may not mean the same product. Source: Anthropic · checked 2026-10-31.
The working definitions do share one idea. Anthropic describes an AI model that "directs its own processes and tool use when accomplishing a task", instead of following a fixed script. Google Cloud calls agents "software systems that use AI to pursue goals and complete tasks on behalf of users". Sources: Anthropic, Google Cloud · checked 2026-10-31.
In plain terms, you say what done looks like and leave the steps to the software. It can reach things outside the chat, like a calendar or an inbox. And it works in a loop: in Anthropic's words, it "plans, acts, observes the result, adjusts, and repeats" until the job is finished or it needs a person.
Neither definition requires anyone to approve anything. Software that emails your customers unchecked can still be called an agent. Anthropic does describe agents pausing "for human feedback at checkpoints", so a checkpoint is a design choice. We put one in front of anything a customer will see. Source: Anthropic · checked 2026-10-31.
Where an agent sits on the AI ladder
We use a four-level AI ladder to explain this. The question at each level is who decides.
| Level | What it looks like | Who decides | | :-- | :-- | :-- | | 1. By hand | A person does the job from memory, with no written rule | A person, every time | | 2. Automated rules | Software follows a fixed if-this-then-that rule and stops when an input falls outside it | A rule a person wrote once | | 3. AI-assisted workflow | A person designs the steps, and AI does one heavy step, like reading or drafting | AI drafts, a person approves | | 4. AI agent | You give the goal and limits, and it plans the job and uses your tools | The agent, inside your limits. You approve what reaches a customer |
Most of the confusing words on vendor sites land somewhere on those levels:
Rule automation is level 2. Same input, same result. It's fast and predictable, and it stops at the first thing nobody wrote a rule for.
"AI automation" is usually level 3. The term is loose, so ask who designed the steps. Anthropic draws the line this way: workflows run "through predefined code paths", while agents direct their own process. If a person mapped the steps and AI fills in one of them, that's a workflow, however it's labelled. Our ladder page covers what it takes to move up.
A chatbot isn't a level at all. It's the window a customer types into, and anything can sit behind it. Google Cloud describes bots as "typically following pre-programmed rules", which is level 2. A chat window that drafts answers for staff to check is level 3. We go through the chat side in what an AI chatbot actually answers.
One job, four levels
Example. A hair salon takes booking requests by web form and text.
Level 1. The owner reads requests at 9pm, checks the diary and replies to each one.
Level 2. A form entry lands in a spreadsheet and sends "we got your request". Nothing gets booked, and a text message doesn't trigger the rule at all.
Level 3. The owner has laid out the steps. For each request, AI drafts a confirmation in the salon's voice, and the owner taps approve.
Level 4. The owner gives it one goal: keep the bookings calendar clean. It reads new requests from both channels, checks availability, drafts the confirmations and puts them in a queue for approval. A client wants a colour in a slot that's too short? It flags that one with a note saying why.
Nothing reached a customer on its own at level 4 either. What changed is the owner's job. They stopped working through steps and started checking a queue.
What an agent can and can't do today
An agent is worth considering when a job has several steps and the steps vary. It can work through those steps with the tools you give it, and it can use the result of each step to decide the next. It can also stop and ask when it's stuck, if it's been built to.
The jobs that suit it share three traits: they come up often, a mistake is cheap, and the result is easy to check.
It falls short in four places:
- Checking its own work is no guarantee, because it's the same software marking its own homework. Anthropic warns that agents' autonomy brings "higher costs, and the potential for compounding errors". One wrong step early can carry through everything after it. Source: Anthropic · checked 2026-10-31.
- It can't run a rule nobody has written down. If you can't explain the job to a new hire in five minutes, an agent has nothing solid to work from.
- It isn't good at people. Google Cloud says agents "can struggle with nuanced human emotions". A complaint from an upset regular still needs you. Source: Google Cloud · checked 2026-10-31.
- It's often the wrong tool. Anthropic's advice to builders is to find "the simplest solution possible". If a fixed rule does the job, use the rule. It's cheaper and it behaves the same way every time.
Who answers when it gets something wrong
Your business does. Australia's guidance on adopting AI says "your organisation is ultimately accountable for how and where AI is used", and we've covered what vendor contracts say about that in what an AI employee actually means. That post is also where you'll find the pitch for software that acts with no sign-off. Our ladder calls level 4 "an AI employee you manage" for a reason: the managing stays with you.
So settle three things before an agent touches a real job:
- One named person owns it. The same guidance says to "make a specific person accountable for every AI system". In a small business, that's probably you. Source: ai.gov.au · checked 2026-10-31.
- The limits are written down: what it can read, what it can change, and what always waits for approval.
- There's a log you can read. When a booking goes wrong, you need to see what it did and why, without ringing the vendor.
Then ask where customer details go. The Office of the Australian Information Commissioner says privacy obligations apply to personal information put into an AI system, and to AI output that contains personal information. It also recommends against entering personal information into publicly available AI tools. Source: OAIC · checked 2026-10-31. Check how this applies to your own business and get legal advice if you're unsure, because we're not lawyers.
The approval queue also limits the damage from a compounding mistake, since it's the last place you can catch one before a customer sees it.
Should you get one?
Probably not for most jobs yet. If your business is on level 1 or 2, you're in good company. Nobody jumps straight to level 4. You climb one job at a time, and a job earns its way up with a track record on real work.
If someone is pitching you an agent, ask these five questions:
- Does it choose its own steps, or follow a fixed set?
- What can it read and change in my systems?
- What waits for my approval before it goes out?
- Where's the log, and can I read it myself?
- Which AI service sits underneath, and where does my customers' data go?
A straight seller will have answers ready. We'll add one thing about ourselves: we have no paid AI client results to show you yet, which is why the salon above is labelled as a scenario.
Questions we get asked
What do I need before an AI agent can start?
Three things: the job written down well enough to explain to a new hire in five minutes, written limits on what it can read and change, and one named person who owns it. Without the first, it has nothing solid to work from.
Can an AI agent send things to customers on its own?
Technically, yes, if it's given access. Whether it should is up to you. When we build one, anything that reaches a customer waits for a person to approve it, and that rule doesn't change after handover.
How much does an AI agent cost?
There's no standard price, because it depends on the job, the tools it connects to and what you pay for the AI underneath. Ask for ongoing usage costs as well as the build. At Whale, you start with a fixed-price diagnosis day, and any build after it is quoted per job, before work starts.
Got one job you'd hand over first?
We build AI agents one job at a time, with limits you set and your approval on anything that reaches a customer. See how AI Agents work at Whale
Sources (all fetched 2026-10-31). Anthropic, Measuring AI agent autonomy in practice: https://www.anthropic.com/research/measuring-agent-autonomy ("there is no agreed-upon definition of what an agent is"). Anthropic, Trustworthy agents in practice: https://www.anthropic.com/research/trustworthy-agents ("directs its own processes and tool use when accomplishing a task"; "plans, acts, observes the result, adjusts, and repeats"). Anthropic, Building effective agents: https://www.anthropic.com/engineering/building-effective-agents ("through predefined code paths"; "for human feedback at checkpoints"; "higher costs, and the potential for compounding errors"; "the simplest solution possible"). Google Cloud, What is an AI agent? (page last updated 04/02/2026): https://cloud.google.com/discover/what-are-ai-agents ("software systems that use AI to pursue goals and complete tasks on behalf of users"; "typically following pre-programmed rules"; "can struggle with nuanced human emotions"). Australian Government, Guidance for AI adoption, Foundations: https://www.ai.gov.au/staying-safe-and-responsible/essential-ai-practices/guidance-ai-adoption-foundations ("your organisation is ultimately accountable for how and where AI is used"; "Make a specific person accountable for every AI system"). OAIC, Guidance on privacy and the use of commercially available AI products: https://www.oaic.gov.au/privacy/privacy-guidance-for-organisations-and-government-agencies/guidance-on-privacy-and-the-use-of-commercially-available-ai-products (privacy obligations on input and output; best-practice recommendation on publicly available tools). Whale wording: https://whaledigitalconsulting.com.au/custom-ai-os and https://whaledigitalconsulting.com.au/business-process-automation#ai-ladder.
