Your AI Agent Is the New Intern — And Somebody Has to Manage It
Imagine you hired a new intern. Smart, enthusiastic, very fast. On day one you hand them the keys to your customer inbox, your booking system, and your invoicing tool. Then you go on holiday.
That’s roughly what most small businesses are doing right now with AI agents.
Patricia Moore, Field CTO at Boomi, made this point at Boomi World Tour Sydney, and it landed because it’s exactly right. Businesses are deploying AI agents like they’re installing software, not like they’re bringing on staff. The analogy to an intern isn’t just cute. It’s a diagnostic.
Start small. Actually earn the trust.
When Moore started a communications role early in her career, she sent emails on behalf of the organisation. Her manager reviewed every one. As she demonstrated judgement, oversight relaxed, first for certain audiences, eventually for almost everything.
That’s the right model for AI agents. Start with bounded tasks, review the output, catch mistakes before they reach the customer, and expand authority as the agent proves itself.
For a small business this might look like:
- A booking agent that handles appointment confirmations, but a human reviews any rescheduling request above a certain value
- An email responder that handles FAQ queries, but flags anything involving a complaint, a refund, or an upset tone
- An invoice processor that extracts data and drafts entries, but a human approves before anything posts to the ledger
None of this is glamorous. But it’s how you build a system you can actually trust, one that earns more autonomy because it’s demonstrated it deserves it.
When it goes wrong, someone has to answer for it.
Most SMB deployments have a blind spot: nobody owns the AI.
Moore references the classic RACI framework, Responsible, Accountable, Consulted, Informed. Everyone wants to be Consulted or Informed. Nobody wants to be Responsible or Accountable. When an AI agent makes a mistake, that gap becomes very visible very fast.
“The AI did it” won’t satisfy a customer whose booking was deleted, a supplier whose invoice was paid twice, or an employee whose personal data was exposed. Someone human has to own the outcome.
Australia has already seen what happens when AI systems behave in ways their owners didn’t anticipate, the gym booking hack being one real local example. Australian consumer law is fairly clear on this: if your agent does something, you did it.
Before you deploy anything, answer these three questions: Who owns this agent’s outputs? Who gets the call when something goes wrong? Who has the authority to shut it down? If you can’t answer all three, you’re not ready.
Your “simple” process is not simple.
Moore describes an insurance company that came to Boomi with what they described as a two-step claims process. Submit claim. Approve or deny. Easy.
By the time they actually mapped the workflow, it was closer to twenty steps, document collection, evidence extraction, policy checks, escalation rules, regulatory constraints, notification requirements, payment processing. Each of those steps had its own edge cases.
SMBs fall into the same trap. “Just automate my customer enquiries” sounds simple until you ask: which enquiries? What happens when someone’s angry, when the query involves a refund, when it’s about a product you’ve discontinued? What does the agent do when it doesn’t know the answer?
The full breakdown of what can go wrong when you skip this mapping is worth reading: What Can Go Wrong With AI Agents. The short version, a lot, and most of it was predictable.
Map the real process before you touch the tooling. Every exception, every edge case, every point where a human currently makes a judgement call. That’s where you’ll find the risk, and where you’ll find where the agent actually adds value versus where it creates new problems.
Your team is already using AI without you.
Shadow AI is a real thing. People paste company documents into free ChatGPT, run customer data through consumer tools, use whatever gets the job done, not because they’re reckless, but because they’re trying to be productive and nobody gave them a better option.
Banning AI tools without providing sanctioned alternatives that actually solve the problem doesn’t eliminate shadow AI. It just makes it invisible, which is worse.
Find out why your team is using those tools. If there’s a legitimate need, meet it with a controlled solution. Create sandboxes where people can experiment safely. Build a culture where failed experiments can be admitted, if employees expect punishment, they’ll hide the failures until they become disasters.
Building an actual AI team structure
The intern analogy points to a bigger question: how do you structure this properly? Not just one agent with vague oversight, but a real approach, roles assigned, boundaries set, accountability loops built in, clear triggers for when to keep a human in the decision.
📖 More deep dives: News Deep Dives
📊 Compare AI tools side by side | 💼 Free resources & AI prompt packs
More hot takes: Hot Takes →
If this interests you
- Treasury Just Called Out Australian Small Business. Are You One of the 90%?
- Australian AI Weekly: 31 August-4 September 2026
- 83% of Australian Businesses Use AI. Only 21% Have Any Rules About It.
More hot takes: Hot Takes →
