CFOtech India - Technology news for CFOs & financial decision-makers
India
The chatbot said yes. Who answers for it?

The chatbot said yes. Who answers for it?

Tue, 17th Nov 2026 (Today)
Nick Lim
NICK LIM Founder Switchboard Finance

When software tells a customer they will be approved, the statement belongs to the business that put the software in front of them. Not to the vendor. Not to the model. That was the position before generative AI, and the guidance regulators have published since points the same way. What has changed is how many pieces of software are now willing to say yes.

Assistants are already handing out verdicts

On 16 November CFOtech Australia reported a new index that scores Australian personal lenders on how five AI assistants portray them. Its authors put 166 unbranded questions to the assistants and recorded which lenders were recommended, cautioned against or cited. An index like that exists because the assistants now answer.

They answer with confidence. In July the broker Savvy put two borrower scenarios to three general-purpose assistants and had its own consultants review the answers. On the car loan scenario, the reviewer noted that none of the tools had asked about credit history, living situation or dependants before giving a figure.

A public assistant that guesses a borrowing limit is nobody's licence. An assistant on a lender's or broker's own website is a different thing entirely. It speaks for the business, to a customer the business invited in.

The law does not ask who typed it

The law on misleading conduct has no carve-out for software. Australia's corporate regulator, ASIC, made the point in 2024 in Report 798, its first review of AI in financial services: licensees should apply their existing obligations to AI now, without waiting for AI-specific law. The chair at the time, Joe Longo, said existing duties "put the onus on institutions".

The advertising rules say the same in more detail. ASIC reissued Regulatory Guide 234 in June 2026. The guide covers a message whatever technology produced it, AI-generated content included, and it asks for balanced messages about the features, benefits, risks and limitations of customer tools, AI tools among them.

So the harder discipline is deciding in advance what a finance chatbot should refuse to do, and writing that list down before the first customer types a word.

The scrutiny is arriving on schedule. On 30 September ASIC wrote to bank boards with its priorities for 2026 to 2027. One is a review of new and proposed uses of AI by banks and the impact on customers. A privacy rule that starts on 10 December will also make businesses say, in their privacy policies, where software has a hand in decisions about people.

Five rules that survive a regulator's questions

The model explains. Something else decides. If the assistant gives a customer a result, a figure or a likelihood, that output should come from fixed rules or from a person, with the language model doing the talking around it. A model that improvises a borrowing limit has made a representation on your behalf.

No promise of approval. A lender makes the credit decision, after an assessment. An assistant cannot approve or decline anything, and it should say so when asked. If it tells a customer they will be fine, the business has made a promise it has no power to keep.

Refuse to help anyone improve the facts. Customers will ask an assistant how to present income, or whether a debt needs to be mentioned. Equifax's Fraud Index for 2025 recorded a 25.5 per cent rise in first-party fraud, where a genuine applicant supplies false information. An assistant that coaches the answer has joined the application.

Keep the record. Every conversation should be logged in a form a person can replay. Human review after the fact is only possible if the record exists. If a customer later says the chatbot told them yes, the business needs to be able to read what it actually said.

Say what it is. Tell the customer they are talking to software and that it gives general information only. Then give them a person at every step. When ASIC reviewed licensees in 2024, nearly half had no policy on consumer fairness or bias, and fewer still had one on telling customers that AI was in use.

The question for finance leaders

Most of the cost of a customer-facing assistant is not the model. It is the fence around it: the rules it must follow, the list of things it must never say, the tests that prove it holds, the log that shows what happened. Vendors sell the model. The fence is the buyer's job.

A finance leader signing off an assistant should ask to see three documents. The list of statements it is not allowed to make. The test results that show it does not make them. And the record of a real conversation, replayed from start to finish.

If those documents do not exist, the business has not bought an assistant. It has hired a spokesperson who never sleeps and has no judgement.

Somebody senior should own that list by name, the way somebody owns the financial statements. Software has no licence to lose. The business does.

The chatbot will say yes to someone this week. Decide now who answers for it.

This piece is general commentary from the lending side. It is not legal advice, and it is not advice on how a particular business should deploy AI.