What is an AI agent? A working definition with five SMB examples

The working definition

An AI agent is an AI program that:

1. Observes a situation (an inbound message, a file, a system state)
2. Plans a next step based on what it observed
3. Calls a tool (read a file, send an email, query a database, hit an API)
4. Observes the result of that tool call
5. Decides the next step - loop back to step 2 with the new context

The loop continues until the work is done or it hits a stop condition (success, escalation, error). The defining feature is the loop: an agent acts over multiple steps without needing human direction at each step.

How an agent is different

• Different from a chatbot: a chatbot answers in one shot. An agent acts over many steps.
• Different from deterministic automation: a deterministic workflow runs the same way every time. An agent has an AI model making decisions in the loop.
• Different from AI workflow with a single AI step: a workflow with one AI step is automation with AI inside it; an agent is AI orchestrating the whole workflow.

Five SMB-grade agent examples

1. After-hours call triage agent

A trades business runs a voice agent that answers after-hours calls. The agent listens, classifies (urgent vs schedulable, plumbing vs electrical), asks the right follow-up questions for each path, books the schedulable jobs into the diary by calling the field-service management system, and escalates the urgent ones to the on-call tech by SMS. Loops until the call is resolved or escalated.

2. Inbox classification and queue agent

An e-commerce business runs an agent on the support inbox. For each incoming email, the agent reads it, classifies the issue (shipping, returns, product question, complaint), pulls relevant context from the CRM and order history, drafts a response, and either sends it (for high-confidence routine replies) or queues it for human review with everything ready to go. Loops over the inbox.

3. Invoice extraction agent

An accounting firm or finance team runs an agent that watches a folder for new supplier invoices. For each invoice, the agent reads it, extracts line items, classifies expense categories, checks for duplicates against historical records, and writes the entry into Xero or MYOB. Escalates anything it cannot classify with confidence. Loops over the folder.

4. Monitoring agent

A SaaS business runs an agent that monitors application logs and customer-impact metrics. The agent reads logs, looks for patterns indicating an issue, correlates with deployment history, drafts a Slack alert with diagnostic context for on-call engineers, and links to relevant documentation. Loops continuously.

5. Customer-onboarding agent

A subscription business runs an agent that walks new customers through setup. The agent introduces itself, asks the right configuration questions one at a time, applies the configuration via API as it goes, checks progress against an ideal onboarding state, and escalates to a human only when something it cannot resolve appears. Loops per customer.

What makes production agents work

• Bounded permissions: the agent can only do what it has been explicitly authorised to do. No broad write access to systems it does not need.
• Full call logging: every step, tool call and decision is recorded. Diligence-friendly and debugging-friendly.
• Clear stop conditions: success, escalation, error - the agent knows when to stop, not just when to keep going.
• An obvious off-switch: humans can pause or kill the agent without much warning.
• Eval set: a fixed set of test cases that gets run on every prompt or model change.

What does not fit an agent

• Strategic decisions - keep human-only
• High-stakes irreversible actions (sending money, deleting records, contacting customers about sensitive issues)
• Anything where empathy or emotional intelligence matters more than speed
• Workflows that already work well as deterministic automation - adding AI in the loop creates new failure modes for marginal benefit

What an SMB agent costs

Production-grade SMB agents typically cost AUD $25,000-80,000 for v1 depending on the integrations needed. Ongoing operating cost is dominated by API usage (often AUD $50-500/month for typical SMB volumes) plus any platform hosting cost. Total cost of ownership is well under one full-time-equivalent salary against a productivity uplift several times that.