Why Australian Firms Are Pulling Their Customer AI Agents

Eighty-four per cent of Australian firms that put an AI agent in front of customers have since pulled it back out. Not paused a pilot — shut down or rolled back a live, customer-facing agent because something went wrong with how it was controlled. The figure comes from Sinch's survey of 2,527 senior decision-makers, and Australia's rate sits ten points above the global average of 74 per cent. Here is the part that should change how you buy: among the organisations with the most mature governance in place, the rollback rate was higher, not lower — 81 per cent.

The distinction matters, because "AI agent" is now stamped on almost everything a software vendor sells a trades business or a professional firm. A chatbot answers a question. An agent takes an action on its own: it books the job, sends the quote, replies to the complaint, updates the record, moves the money. Sinch surveyed large companies — 1,000 employees and up, across financial services, retail, telco and professional services — and 63 per cent of the Australian ones had already moved a customer-communications agent into live production.

Then most hit the same wall. The reasons firms gave for pulling their agents were mundane and predictable: privacy concerns, no way to audit what the agent had actually done, and hallucinations — the agent confidently telling a customer something untrue. These aren't exotic failures. They are the ordinary ways autonomous software goes wrong when it faces the public.

Want to know what this means for your business? We'll map it in one call.

The instinct, reading that, is to go buy a governance framework. The data says that is not the lesson. If the best-resourced companies — the ones with AI committees, model registers and compliance teams — are the most likely to pull a customer-facing agent, the problem is not a shortage of policy. It is that letting software act unsupervised on your customers is genuinely hard, and the disciplined firms are simply the ones honest enough to switch it off when it misfires. A ten-person trades business or a boutique firm has none of that scaffolding. It carries the same risk with less to catch it.

And the risk lands on you, not the vendor. When an agent quotes the wrong price, promises a callback that never comes, or tells a customer something untrue, the business wears it — we've written before about who is on the hook when an AI tool misleads a customer under Australian Consumer Law, and the short answer is you. It is the same governance gap we flagged when SAP's July numbers landed: only 22 per cent of Australian businesses feel ready to govern the AI they already run, and in that same report 49 per cent said their AI agents had already taken incorrect actions. Half. That is not an edge case you're insuring against — it's the base rate.

None of this is an argument against agents. It's an argument for where you point them first. The failures cluster on one surface: the customer-facing one, where a wrong action is public and sometimes contractual. Point the same agent at your own back office — drafting the quote for a human to send, summarising the job notes, reconciling invoices before someone signs off — and the cost of a mistake is a corrected draft, not a misled customer. That's the Admin Leverage win without the Customer Intelligence risk. If you're weighing where an agent could safely start in your business, that's the first question to answer.

You don't need a governance department. You need three boundaries. First, keep a human on the send button for anything customer-facing — let the agent draft, a person approves, at least until you've watched it work for a month. Second, cap what the agent can do alone: reading your calendar and drafting a reply is low-risk; confirming a price, locking in a booking or issuing a refund without sign-off is not. Third, keep a log — "no way to audit it" was one of the top reasons firms pulled theirs, and if you can't see what the agent did and why, you can neither fix it nor defend it. Start the agent inside the business, prove it there, then let it face a customer.

Key takeaways

Common questions

Related notes

Thinking about letting an AI agent handle customers?

The difference between an agent that sticks and one you pull in a fortnight is where you point it and what you let it do alone. A short conversation will map where an agent can safely start in your business — and the boundaries to set before it faces a customer. Book a call to talk it through.