An AI agent is AI that takes steps in your systems, not just answers in a chat window. It might read an email, look up the customer and job, draft a reply and create a task for someone to approve. For an established, growing business, agents are real today for narrow, repetitive back-office jobs: sorting incoming work, drafting, handling documents and answering staff questions from your own information. They only work when they can reach your systems through proper connections, the data is clean, their access is limited, a person approves anything that matters, and every step is logged. The "AI employee that runs your back office" is still mostly hype. The sensible start is one workflow, in a small pilot, with a fixed end date.
What an AI agent actually is
A chat assistant such as ChatGPT or Copilot waits for you to ask something, then answers. You carry the answer to wherever it's needed.
An agent is given a goal and a set of tools, and works through the steps itself. The tools are the important part: "look up a customer in the CRM", "read the attachments on this email", "create a draft purchase order", "add a note to this job". The AI decides which tool to use next; the tools do the actual work in your systems.
For example, if your team receives order emails with PDF attachments, an agent could:
- Read the email and the PDF.
- Look up the customer and their price list in your sales system.
- Check stock in your stock system.
- Prepare a draft order with any problems flagged, such as an unknown product code or a price that doesn't match.
- Put it in a queue for a person to approve.
Each of those steps needs access to a system. That's why most agent projects turn out to be integration projects with AI in the middle.
According to the National AI Centre's AI adoption insights for December 2025 to February 2026, agentic AI is among the uses that "remain largely untapped" by Australian small and medium businesses. That's not a reason to rush. It's a reason to start in the right place.
What an agent needs before it can do anything useful
Access to your systems through proper connections
An agent can only act in systems it can reach. That means integrations through each system's API (the official way for software to talk to other software) or another supported connection. If your job system, stock system or industry software has no way in, or only a half-working connector, that has to be solved first. Our API integration page explains what that involves.
Clean, consistent data
If a customer appears three times under slightly different names, or product codes differ between your sales and stock systems, an agent will pick the wrong one or stop. People work around messy data from experience; agents can't. Cleaning the data first, and agreeing which system holds the master record (a single source of truth: one place everyone takes the number from), makes the difference between a useful agent and an expensive one.
Limited permissions
Give an agent the least access it needs. If its job is to draft orders, it can read customers and stock and create draft orders, and nothing else. It shouldn't be able to approve payments, change bank details or delete records. Start with read-only access where you can.
A person approving what matters
The AI does the first pass; a person approves. Anything that affects money, stock, pay, customers or staff goes through a person before it's final. The federal Guidance for AI Adoption makes the same point about keeping people accountable and in control.
Logs of every step
You need a record of what the agent read, what it did, what it was unsure about and who approved the result. That's how you check its work, fix mistakes, answer an auditor and decide whether it's earning its keep.
What's realistic now
These uses work today in a growing business, provided the conditions above are in place.
- Triage: sorting a shared inbox or a queue of requests, working out what each one is (order, query, complaint, invoice), attaching the right customer or job, and routing it to the right person.
- Drafting: preparing replies, quotes, purchase orders or job notes from your own templates and data, for a person to check and send.
- Document handling: reading supplier invoices, delivery dockets, timesheets and order forms, and preparing entries for approval. Our post on AI reading dockets, orders and invoices goes into detail.
- Internal questions: answering staff questions from your own procedures, price lists and policies, with a link to the source so people can check the answer.
- Checks and exceptions: comparing an invoice to the order and the delivery, and flagging only the ones that don't match.
Notice what they have in common: they're repetitive, the input is messy, a person can check the result quickly, and a mistake is caught before it costs anything.
What's still mostly hype
- "An AI employee for your back office." Agents handle narrow tasks well. A whole role, with judgement, context and accountability, is a different thing.
- Set and forget. Agents need monitoring, and their instructions need adjusting as your business changes.
- Agents that pay, approve or commit you to things on their own. Possible to build, but not wise without a person in the loop.
- "Connect everything and the agent will find the value." Without a defined task and clean data, it finds confusion.
- Customer-facing agents as a first step. We don't build customer chatbots. Back-office work is where the risk is lowest and the gain is easiest to measure.
The risks to plan for
- Wrong actions, made quickly. A person who makes a mistake makes one. An agent with the same mistake in its instructions can repeat it a hundred times before anyone notices. Approvals and limits catch this.
- Access that's too wide. An agent with broad access can be misused or confused. Limit what it can see and do.
- Instructions hidden in documents. An email or PDF can contain text designed to trick an AI into doing something else. Agents that read outside documents must not be able to take high-impact actions without approval.
- No record. Without logs, you can't explain a decision or fix a pattern of errors.
- Privacy and automated decisions. If an agent helps make decisions about people, such as customers or staff, privacy rules apply. New Privacy Act transparency requirements for automated decisions start on 10 December 2026; our post on automated decision-making disclosure explains them. We're not lawyers, so get advice for your situation.
- Running costs that creep. Usage-based AI costs rise with volume. Set limits and watch them during the pilot.
How to start small
- Pick one workflow. One task, one team, high volume, easy to check. Order emails, supplier invoices and delivery dockets are common first choices.
- Map how it works today. Who receives it, what they look up, where they type it, what goes wrong. You'll often find a step that a plain integration can remove without any AI. If you're unsure whether you need AI at all, our companion article AI, automation or integration? Which one your business actually needs and the AI fit check help you decide.
- Connect only what that workflow needs, with limited permissions.
- Run it alongside the current process for a fixed period, with a person approving every output.
- Measure and decide: extend, change or stop.
Our back-office AI pilot runs exactly this: one workflow, real documents, a person approving and a fixed end date. If you'd like to run a pilot yourself, our guide to running an AI pilot safely sets out the steps.
What it looks like when it's working
On Monday morning, your customer service lead opens a queue instead of an inbox. Weekend order emails have already been read, matched to customers and checked against stock. Most are ready to approve with one click. A handful are flagged: an unknown product code, a price that doesn't match the customer's agreement, an order over their credit limit. Each flag says why. Nothing was entered into the sales system without a person approving it, and every step is in the log.
That's an agent doing what agents are good at: the reading, looking up and preparing, with a person deciding.
The technical side
For whoever looks after your systems:
- Tools are API calls. Each agent tool wraps one operation in one system, ideally through a service account with only the permissions that tool needs. Open standards such as the Model Context Protocol make tools easier to plug in, but they don't remove the need for sensible permissions.
- Draft, don't commit. Have tools create drafts or queue items, and let your existing approval steps finalise them.
- Log inputs, tool calls and outputs, with the approving user, in a place your team can search.
- Use business-grade AI services whose terms state how your data is handled, and check those terms for the plan you use.
- Test with real documents, including messy and unusual ones, before trusting the results.
Our AI workflow automation page covers how we build these workflows into the systems you already run.
Frequently asked questions
What is an AI agent in simple terms?
AI that takes steps in your systems to finish a task, such as reading an email, looking up the customer and preparing a draft order. A chat assistant answers questions; an agent does steps.
Are AI agents ready for small and mid-sized businesses?
For narrow, repetitive back-office tasks with a person approving, yes. For running whole processes unsupervised, not yet. Start with one workflow.
Do we need to connect our systems before using an agent?
Yes. An agent can only act in systems it can reach, so connections to your job, stock, sales or accounts systems come first. Often that integration is the biggest part of the work, and sometimes it fixes the problem without AI.
How do we stop an agent making expensive mistakes?
Limit its access, have it create drafts rather than final records, require a person to approve anything that affects money, stock, pay or customers, and log every step.
Where should a business start with AI agents?
With an independent look at where AI fits. Our AI readiness assessment identifies which workflows are worth trying first, and our AI consultant for small business page explains how we work.
If you have a workflow in mind and want an honest view of whether an agent, a plain integration or something simpler is the right fit, book a free initial chat.