Your team is copying information between tools, chasing approvals or rewriting the same material. Someone suggests an AI agent. Someone else wants a dashboard. Both might be useful. Neither is the starting point.
Use automation when the steps are predictable. Use an AI agent when the work involves interpreting information and choosing a next action. Build custom software when people need a better way to see, use or manage the work. These approaches can sit inside the same system.
For an Australian agency or growing service business, the useful question is: what is the smallest dependable change that would make this job easier?
The three options, in plain English
Scroll table
| Approach | What it does | A sensible first example |
|---|---|---|
| Workflow automation | Moves work through steps and rules you define | Create a project and notify the delivery lead after a signed agreement |
| AI agent | Uses a model and permitted tools to interpret a request and choose actions | Read an enquiry, consult approved service information and prepare a response for review |
| Custom software | Gives people an interface and business logic designed around their job | A client portal for approvals, files and project progress |
An agent is more than a character or chat window. n8n’s documentation describes agents as decision-makers that use language models and tools. That distinction matters: a fixed sequence with one AI writing step is not automatically an autonomous worker. See n8n’s explanation of agents.
Start with a real example of the work
Take one recent job from beginning to end. Find the original request, the information someone looked up, the decisions they made and the final output. Mark where the work waited for somebody.
Suppose a client asks for a proposal. The team needs to understand the brief, find relevant experience, choose a scope, check prices and send a document. Those are different kinds of work. A useful system might gather approved material automatically, use AI to prepare a first draft and give a person a clear place to check it. The person still approves the scope, price and promise.
This is an illustrative design, not a claim that a proposal can safely write and send itself.
Choose automation when the rule is already clear
If the team can describe a reliable rule, start there. A completed form can create a record. A signed agreement can start an onboarding checklist. A missing field can send the work back to a person.
Before connecting anything, agree which tool holds the authoritative record. Decide what happens if the same event arrives twice or one system is unavailable. A workflow that needs somebody to quietly repair it every Friday has not removed the job; it has moved it.
A useful first version has a visible status, a named owner and a way to retry or finish the work manually.
Add an agent when interpretation earns its place
An agent can be helpful when inputs vary: different enquiry formats, inconsistent notes or a question that needs several sources. Start with a narrow task and explicit permissions.
For example, let an assistant read approved project information and suggest an answer. Keep sending messages, changing commercial terms and editing important records behind a person’s approval until the behaviour has been tested and the authority deliberately agreed.
Test ordinary examples and awkward ones: an incomplete brief, contradictory instructions, an out-of-date price and a request outside scope. Agree what a good answer looks like before judging the demo.
Build software when the interface is part of the problem
An internal tool or client portal can be the right choice when people cannot see the status of work, find the current version or understand what they need to do next. The value comes from the shared view, the workflow and the decisions it supports.
That software may contain AI. It may also contain ordinary forms, permissions and integrations. A clear approval screen can be more useful than another chat window.
If an existing product already handles the job well, use it. The build-or-buy guide explains when a custom layer earns its cost.
Give the first build a boundary
Write down five things:
- The person who will use it and the job they need to finish.
- The input it may read and the output it should produce.
- The decisions it may make and the actions a person must approve.
- The exceptions that should stop it or send it to someone.
- The result you will compare with today’s process.
If the work involves personal information, add a data review before connecting real records. The OAIC recommends due diligence when selecting AI products, including considering access, privacy risks and human oversight. Read the OAIC guidance.
What should an Australian business do first?
Bring one actual example, not a complete archive of company data. A short walkthrough can reveal whether the next step is a rules-based connection, a bounded AI assistant, an interface or simply a clearer process.
Agree a small delivery scope and a result check. Then decide whether the next improvement is worth doing. See the four-stage delivery process.
Do we need all three?
No. Choose the smallest combination that handles the job dependably. Complexity creates work of its own.
Can an agent work inside tools we already use?
Potentially. The available integrations, account permissions and review controls need checking before promising that it will work inside a particular chat platform or business system.
Does custom software always cost more?
The answer depends on scope and ongoing ownership. Compare the complete job, including subscriptions, maintenance, review time and the cost of keeping the current workaround. Use the project-budget worksheet.
Bring a real example.
Bring one workflow to a conversation. We can work out whether it needs a small connection, an AI assistant or a proper software product.
Let’s talk