Santam's Claims Ledger Leaves the JSE

The most useful public record of how AI changes an African insurer's claims book is about to stop being public. It belongs to a company that has said plainly it intends to put more data and machine learning into pricing, risk selection and claims. It will go quiet at exactly the moment that work starts to show up in the numbers.

The company is Santam, South Africa's largest short-term insurer. On 5 October 2026, Sanlam offered R505 a share in cash for the part of Santam it does not already own, 26.6% above the R399 close on 2 October, according to Eyewitness News. Bloomberg's report, carried by Insurance Journal on 7 October, puts the cheque at about R20.7 billion for the 37.3% minority, or 11.7 times Santam's forward earnings. Santam's independent board backs the scheme. The shareholder meeting is set for 30 November and completion for the first quarter of 2027, per New Era in Windhoek, since Santam is also listed in Namibia. I have not been able to read the SENS announcement itself, so these terms are as reported, and the three outlets agree on them.

1. What the interim booklet shows

Santam's investor booklet for the six months to 30 June 2026 is a 50-slide document, and it is the closest thing South African insurance has to a public lab notebook on data-driven underwriting. The headline numbers moved the wrong way for the half.

Premium grew 10.1% while the margin lost more than three points. The booklet explains much of the gap. Excluding large losses, a reserve change and the start-up of Syndicate 1918, its new Lloyd's operation, the margin was 15.7%. That adjustment is management's, and readers should treat it as such. Still, the table is the right shape for an outsider. It sets the claims ratio, the measure any AI-in-claims effort has to move eventually, against the same line a year earlier, on audited definitions, twice a year.

The technology slides are more specific than most insurers manage. Santam's FutureFit 2030 plan commits to using data and AI to improve pricing and risk selection. A geocoding programme is described as "powering smarter underwriting, risk mapping and claims response". Satellite flood monitoring with ICEYE now tracks events as they unfold, and the booklet attributes roughly R200 million of weather savings in the half to this work. The outlook slide reads: "Accelerate investments in AI and data whilst strengthening AI governance."

The R200 million is the figure I would most like to see tested. It is a company estimate with no published method, and weather savings are notoriously sensitive to the counterfactual you choose. A sceptic would ask what share came from fewer claims and what share from faster ones. The useful thing about a listed company is that the sceptic gets to ask, and the answer, or the silence, goes on the record.

2. Who else is publishing

Plenty of South African insurers use machine learning on claims. Very few report the results on a timetable. ITWeb's survey of insurers on 28 July 2026 gathered the rest. Naked Insurance says almost every windscreen claim it receives is approved automatically, with no human involved. King Price says AI has cut turnaround on routine claims by 50%. Discovery Life says its models help keep fraud below 0.5% of claims. Every one of those is a vendor statement about the vendor's own system, offered to a journalist, with no baseline, no sample size and no definition of what counts as routine or as fraud. Naked has gone furthest on the customer side: in May it became, by Business Day's account, the first South African insurer to issue binding car insurance quotes inside ChatGPT.

Sanlam, the buyer, has made its own view of the business clear. When it named Theo Mabaso Group Chief AI Officer in March 2026, the release quoted him: "Insurance is, in many ways, a data business masquerading as a financial one." The same release lists underwriting and claims processing among the areas where Sanlam expects AI to have its strongest effect. Put that next to the offer and the logic is plain. Sanlam wants all of Santam's data, all of its claims flow and all of its capital, and Paul Hanratty told Bloomberg the deal "creates a clearer platform from which to execute."

3. The case for the deal, taken seriously

The best argument against my concern is the one dealmakers made within hours. Andrew Bahlmann, chief executive of the advisory firm Deal Leaders International, told the Money Show that "it makes sense to buy out the minorities and move into this next chapter." He is right on the commercial terms. Two listed companies with one controlling shareholder carry two sets of governance costs and an awkward conflict over who gets the capital. A 26.6% premium is a fair price for minorities to give that up, and a single owner can fund a multi-year data programme without defending every quarter's margin to outside holders. If AI investment pays off over five years and hurts earnings for the first two, private ownership is arguably the better structure in which to carry it.

I accept all of that, and it does not answer my objection. The deal is about who owns Santam. My concern is about who can see what Santam's models do to policyholders. Once Santam is a wholly owned subsidiary, its claims ratio, weather savings and AI spending become lines inside Sanlam's group reporting, at whatever level of detail Sanlam chooses. That is my inference from how consolidated subsidiaries are usually reported, and Sanlam may prove it wrong. Nothing in the reported terms promises a separate disclosure.

The loss is a particular kind of evidence. The South African market has many AI claims about claims and very few time series. Santam's booklet is one of the few where a reader can line up a stated technology programme against an audited ratio, half by half, and see whether the two move together. Take it away and the country's best public test case for AI in non-life underwriting goes dark, a few months after the company said it would speed that work up.

4. What the regulators already asked for

The Financial Sector Conduct Authority and the Prudential Authority anticipated part of this. Their joint report on AI in the financial sector, released by the Reserve Bank on 24 November 2025, says institutions "should clearly disclose when AI is used in consumer-impacting decisions," and gives insurance pricing as an example. It names bias, discrimination and explainability as the risks and points firms to tools such as SHAP and LIME.

That guidance is about what an individual customer is told. It says nothing about what the market is told in aggregate. This is where my view sits, and I would defend it in front of either regulator: a large insurer that uses machine learning in pricing or claims should publish a small, fixed set of figures every six months whether it is listed or not. These would be the share of claims settled with no human review, the repudiation rate on those claims against the rate on reviewed ones, and the number of automated decisions overturned on complaint or appeal. Those three numbers would turn every press-release statistic in section 2 into something a policyholder, an ombud or a competitor could check. They are cheap to produce for any firm already running the systems it describes.

Santam's delisting makes the gap urgent, and the gap is wider than one company. Most of the insurers making claims about AI in South Africa were never listed. Naked and King Price have never published a claims ratio against which their automation figures could be tested. The Santam booklet was the exception that showed what good looks like, and it was voluntary in the sense that matters: no rule required the AI and geocoding slides.

5. Where this leaves the record

Sanlam is paying about R20.7 billion for full control of a short-term insurer whose own documents commit it to more data and AI in underwriting and claims. Santam's most recent public half shows premium up 10.1%, a claims ratio up from 56.0% to 58.4%, and about R200 million of weather savings attributed to geocoding and satellite monitoring, all on management's own definitions. After the first quarter of 2027 those figures will no longer reach the public in that form. Other South African insurers make larger claims for AI in claims handling, with less evidence behind them. The regulators have asked for disclosure to individual customers and have not yet asked for aggregate figures. The deal is reasonable on its terms. The cost is a public measurement series that had only just become worth reading.

Tarry Singh is the founder and CEO of Real AI, an enterprise AI advisory and deployment firm working with global enterprises on production agent systems, model risk, and AI sovereignty strategy. He also leads Earthscan for Energy AI startup, and is a founding contributor to the EU-funded HCAIM and PANORAIMA programmes for responsible AI education across European universities. He writes at tarrysingh.com.