AI agents for business: what actually works now the hype has cooled

Search interest in "AI agents" in Australia rose steadily through the first half of 2026, peaked in May, and has since fallen back by around 80%. That's a familiar pattern. The hype phase is over, and the people still searching are mostly asking a more useful question: which of this actually works in a real business?

This is our answer, from building AI features into production systems. Short version: agents work well on narrow, well-defined jobs with clear tools, clear limits and a way to check their work. They work badly as general-purpose digital employees.

What we mean by an agent

An AI agent is a model that can take actions, not just produce text. It decides which step to take next, calls a tool (search a database, read a document, update a record, send a draft), looks at the result, and continues until the task is done or it hands back to a person.

That's different from a chatbot, which answers questions, and from traditional automation, which follows fixed rules. An agent sits in between: it follows a goal, not a script, so it can cope with variation that fixed rules can't. That flexibility is the whole point, and also the whole risk.

Where agents work today

The agents we'd confidently put into production share a few traits: the job is narrow, the inputs vary but the outcome is well defined, the tools are limited, and a mistake is either cheap or caught before it matters.

• Triage and routing. Reading incoming emails, tickets or forms, classifying them, pulling out the key details and routing them to the right queue with a summary. Mistakes are easy to spot and fix.
• Document processing with lookups. Extracting fields from invoices, applications or specifications, checking them against your systems (does this supplier exist, does this order number match), and flagging mismatches for a person.
• Research and preparation. Gathering the information someone needs before a decision, such as account history, related cases or relevant clauses, and presenting it with sources. The person still makes the decision.
• Drafting within a workflow. Preparing replies, quotes or reports from your own data for a person to review and send.
• Internal knowledge search. Answering staff questions from your policies, procedures and past records, with citations and an honest "I don't know".

In each case, the agent removes the slow, repetitive part of the job, and a person keeps the judgement.

Where agents still struggle

• Long chains of steps with no checkpoints. Small error rates compound. An agent that's right 95% of the time at each step is right much less often across ten steps.
• Open-ended goals. "Grow our sales" or "handle customer service" gives the agent no clear definition of done and no way to check itself.
• Irreversible actions. Payments, deletions, contract changes and anything sent to customers without review. These need a person to approve them, at least until there's a long track record.
• Messy or inaccessible data. An agent is only as good as the systems it can reach. If the information lives in people's heads or in spreadsheets nobody trusts, fix that first.
• Tool sprawl. Giving an agent access to everything "in case it's useful" makes it slower, less predictable and harder to secure.

Agent or simple automation?

Many problems sold as agent projects are better solved with a simpler tool. A rough guide:

Cheaper, simpler and more predictable usually wins. We'd rather recommend a single, well-tested model call than an agent you can't explain.

How to build an agent you can trust

If an agent is the right tool, these are the things we put in place before it touches real work.

1. One job, written down. What the agent does, what it must never do, and what "done" looks like, agreed with the person who owns the workflow.
2. The fewest tools that do the job. Read access first. Write access only where needed, scoped as tightly as possible.
3. Approval on anything costly or irreversible. The agent prepares, a person approves. Make approving fast.
4. Evals on real examples. A scored test set, run on every change, covering the common cases, the awkward ones and the ones where the right answer is to stop and ask. We've written about how we build evals.
5. A full trace of every run. Which tools it called, with what inputs, and what came back, so any mistake can be traced and fixed.
6. Limits on cost and time. A cap on steps and spend per task, so a confused agent stops rather than loops.
7. Monitoring after launch. Correction rates, cost per task and changes in the kind of input coming in, reviewed on a schedule.

Where to start

Pick one workflow that is repetitive, information-heavy, and costing real hours each week. Find the person who owns it. Then test whether AI can do the job on your actual data before committing to a build.

That's what our AI Opportunity Sprint is for: two to three weeks, A$5k–15k, one workflow, a working prototype on your data, and an eval scorecard showing how well it performs and what it would cost to run. Sometimes the answer is an agent. Often it's something simpler. Occasionally it's "not yet". See what the Sprint includes, and why so many AI pilots never reach production.

FAQ

What is an AI agent in simple terms?

Software that uses an AI model to decide and take steps towards a goal, using tools you give it, such as searching a database or updating a record, rather than following a fixed script.

Are AI agents worth it for small businesses?

For a narrow, repetitive job like triaging enquiries or processing documents, often yes. Start with one workflow and measure the time saved before expanding.

What's the difference between an AI agent and a chatbot?

A chatbot answers questions. An agent takes actions in your systems to complete a task. That makes agents more useful and more risky, which is why they need limits and approval steps.

Is it safe to let an AI agent act on its own?

For low-cost, reversible actions with a good test record, it can be. For payments, customer communications and anything hard to undo, keep a person approving the action.

How much does it cost to build an AI agent?

It depends on the workflow, integrations and review needed. A focused Sprint to prove it works on your data is A$5k–15k. A production build typically runs as a Pilot of A$15k–150k. Our cost guide covers what moves the number.

---

Wondering whether an agent is the right fit for a workflow in your business? Book a scoping call. Bring the workflow, and we'll tell you honestly whether it needs an agent, something simpler, or nothing at all.