About
Subscribe

Putting AI on the record

When AI-assisted advice is challenged, will South Africa’s financial institutions have the evidence?
Johannesburg, 18 Nov 2026
Putting AI on the record.
Putting AI on the record.

Generative AI is now part of everyday work in South Africa’s banks and insurers. Credit analysts use it to summarise client financials, advisers use it to draft recommendations and operations teams use it to prepare client correspondence.

Microsoft 365 Copilot is a common starting point because most institutions already have a Microsoft 365 agreement in place. Many teams are also using OpenAI’s ChatGPT Enterprise and Anthropic’s Claude Enterprise, building proprietary models on cloud AI platforms or turning to personal AI apps on their phones.

“Many institutions can tell you how many licences they’ve bought, but far fewer can say with confidence which AI tools have touched client information,” says Matt Balcomb, Sales Director SA at Smarsh. “The question South African financial institutions face isn’t whether generative AI boosts productivity. The numbers already prove it does. The question is defensibility.”

Matt Balcomb, Sales Director Southern Africa, Smarsh.
Matt Balcomb, Sales Director Southern Africa, Smarsh.

South Africa doesn’t yet have rules written specifically for AI-assisted advice. That doesn’t give institutions a grace period while they wait for new AI laws, because the existing framework already governs every AI interaction that touches client data or advice. The Financial Advisory and Intermediary Services (FAIS) Act requires records of advice to be kept for at least five years, and its General Code of Conduct expects due skill, care and diligence.

The Protection of Personal Information Act (POPIA) holds the responsible party accountable for personal information even when it is processed inside a third-party AI model. “If a large language model (LLM) helped shape a recommendation, that interaction is part of the advice trail and it sits on the same retention clock as everything else,” explains Balcomb.

The evidence gap

When compliance teams are asked about supervisory oversight of AI, the first answer is usually that audit logs exist. Native AI platform logs often record that an interaction happened, but what they capture varies by product, plan and configuration. They leave out prompt edits, regenerated answers and uploaded attachments. They also can’t prove that advice was given with due care, that a supervisor reviewed it or that policy was enforced across every model. “A log shows that someone used a tool. A record shows what was said and decided,” says Balcomb.

Copilot dominates the conversation because of the Microsoft 365 footprint, but compliance risk is tool-agnostic. ChatGPT Enterprise, Claude Enterprise and even in-house LLMs create the same evidentiary blind spots, and often greater ones. Because these tools are frequently adopted department by department, they leave fragmented data silos outside the e-mail, mobile and voice archives institutions already run for compliance. “The AI models themselves aren’t the main compliance risk,” continues Balcomb. “The risk is the absence of supervisory capture around them, and because AI sits between IT, risk and compliance, that gap often belongs to no one.”

Owning the outcome

Under FAIS, supervisory failure is a personal liability for executives and key individuals. A key individual is responsible for overseeing the financial services a provider offers and must remain fit and proper to do so, and advice that passed through an AI tool is not exempt.

“If AI-assisted advice is challenged and the institution can’t show what was produced and how it was supervised, regulators will want to know whether oversight was adequate,” adds Balcomb. “'The tool got it wrong’ is unlikely to be accepted as an explanation for weak oversight.”

Many institutions have responded with a policy telling advisers to verify AI output before sending it, which holds up only until someone asks for evidence that the check took place. “If a client challenges that advice two years later, you have to prove the review happened. Without the prompt, the output and every revision, there’s nothing to prove it with,” he warns.

Clearing the way

A defensible governance framework has to span the entire enterprise AI ecosystem. Smarsh’s approach is centralised, cross-model capture of interactions from every sanctioned enterprise AI platform, including prompts, responses, uploaded files, edits and regenerated outputs, stored in one regulatory archive alongside e-mail, chat, voice and mobile communications.

“Governance is what gives legal, risk and compliance the confidence to green-light the next AI project without throttling productivity,” says Balcomb.

He suggests risk leaders start with three checks. The first is visibility of every AI tool in use, sanctioned or not. The second is retrieval, testing whether one real AI-assisted client interaction can be reproduced in full and kept for the required retention period. The third is supervision, confirming that a qualified person’s review is on record and linked to the advice.

“Closing that gap doesn’t mean holding AI back,” he says. “It means giving risk officers the visibility they need to let adoption move forward with confidence.”

Read more in the Smarsh blog: Why AI Communications Governance Is a Risk Discipline

Share