Industry · Insuretech

AI and insurtech software for South African insurers

Claims that settle themselves when they should, fraud rings found by the bank accounts they share, and an audit trail your supervisor can read. Built for FSCA and Prudential Authority scrutiny, POPIA and IFRS 17 — not retrofitted for them.

Discuss your project
8%
of South African insurers reported using AI, against 52% of banks
Source: FSCA & PA, AI in the SA Financial Sector (2025)
R3.5bn
potential industry loss to fraud and organised crime if left unchecked
Source: Insurance Crime Bureau, 2024 Annual Report
59
bank accounts used by one syndicate claiming on 69 policies across multiple insurers
Source: Insurance Crime Bureau, 2024 Annual Report

Why insurance automation is a governance problem first

Insurers were automating decisions long before anyone called it AI. Actuarial tables are models; underwriting rules are automated decisioning. What has changed is not whether a machine decides, but whether anyone can explain how — and South Africa's regulators have moved from watching that question to acting on it.

That reframes the buying decision. The hard part of an insurtech build is almost never whether a model can read a claim form. It is whether you can show a supervisor the lineage behind a declined claim eighteen months later, whether the personal information in that claim was minimised before a model saw it, and whether a human was accountable for the outcome. Teams that treat those as a compliance workstream bolted on at the end rebuild the platform. Teams that treat them as architecture ship once.

We build the governed layer underneath: ingestion and lineage, classification and masking before inference, model and prompt versioning, human-in-the-loop gates on anything that affects a policyholder, and audit trails that answer a question rather than store a log. The AI sits on top of that, not instead of it.

Our work in adjacent regulated ground is the closest proof: we built a FICA compliance and fraud-scoring platform for a South African neo bank using document intelligence for KYC verification and a low-latency endpoint for real-time transaction scoring. The compliance shape of that problem — automate the clear cases, route the uncertain ones to a human with a structured summary, evidence every decision — is the same shape as claims.

Case study: FICA compliance and fraud detectionCase study: regulatory-grade data platform for a neo bankOur fintech practice
How it works

How a claim moves through an automated pipeline

This is the shape of most insurance automation we would build. The AI does the reading and the checking; the rules you set decide what settles on its own; a person decides everything else.

  1. STEP 01

    Claim arrives

    Email, WhatsApp, broker portal or a scanned form — every channel lands in one queue with a reference number.

  2. STEP 02

    Read and classified

    Document intelligence pulls out the policy number, incident, dates and amounts, and classifies the claim type.

  3. STEP 03

    Personal information masked

    ID numbers, medical and banking details are masked before any model reasons over the claim. POPIA by design, not by policy document.

  4. STEP 04

    Checked against policy and fraud graph

    Cover, waiting periods and excess are checked, then the claim is tested for shared bank accounts, addresses and providers across the book.

Automated

Clear case: settled straight through

Within limits you set — claim type, amount, risk score — it is approved without anyone touching it.

Human decides

Uncertain case: routed to an assessor

With a one-page summary and the exact reason it was referred, so the assessor sees a decision, not a pile of PDFs.

Every decision evidenced

What the model saw, which rule and model version decided, and who approved it — retrievable by query when a supervisor or the Ombud asks.

Guides

Insurance AI, explained: how to do it and why it is hard

The questions insurance leaders actually ask us, answered plainly. Each one starts with a short answer you can take into a meeting, then the steps or reasons behind it.

01 · How to

How do you automate insurance claims processing in South Africa?

Short answer

Start with one high-volume, document-heavy claim type. Automate the reading and checking, let limits you set decide what settles straight through, and send everything else to an assessor with a summary. Keep a person accountable for any decline, because POPIA section 71 restricts decisions based solely on automated processing.

  1. 1

    Pick one claim type with high volume and clear rules, and measure today's cycle time and how many hands touch each claim. Without a before-number there is no business case.

  2. 2

    Bring every channel — email, WhatsApp, broker portal, scanned forms — into one queue with a reference number.

  3. 3

    Use document intelligence to read and classify each claim, and send low-confidence fields to a person instead of guessing.

  4. 4

    Mask ID numbers, medical and banking details before any model reasons over the claim.

  5. 5

    Agree the straight-through limits (claim type, amount, risk score) with your claims team, not the vendor, and start them conservatively.

  6. 6

    Route everything outside those limits to an assessor with a one-page summary and the reason it was referred. Declines always go to a person.

  7. 7

    Log what the model saw, which rule and model version decided, and who approved — so a supervisor or the Ombud gets an answer, not a reconstruction.

Source: POPIA section 71 — automated decision making
02 · How to

How do you detect insurance fraud rings, not just fraudulent claims?

Short answer

Treat claims as a network rather than a list. Connect claimants, identities, bank accounts, addresses, providers and policies, then look for links no single claim shows. The Insurance Crime Bureau described one syndicate using 98 illegally obtained identities to claim on 69 policies at multiple insurers, paid into 59 bank accounts.

  1. 1

    Resolve entities across the book first: the same person, account or address recorded three different ways is still one node.

  2. 2

    Build a graph of shared attributes — bank accounts, phone numbers, addresses, providers, brokers and beneficiaries.

  3. 3

    Score the patterns that matter: many policies paying into few accounts, fast-track claims soon after inception, one provider across unrelated claimants.

  4. 4

    Give investigators a short, ranked queue with the links that triggered each alert, not a raw score.

  5. 5

    Feed confirmed and cleared cases back in, so the model learns your book rather than a generic one.

Source: FAnews on the Insurance Crime Bureau 2024 Annual Report
03 · Why

Why are South African insurers slower to adopt AI than banks?

Short answer

In the FSCA and Prudential Authority survey, 8% of insurers reported using AI, against 52% of banks. The regulators read this as a cautious approach, and named a shortage of skilled AI professionals as a major constraint. Of the insurers already using AI, 41% planned to spend less than R1 million on it in 2024.

  • Skills: the regulators' own report names the shortage of skilled AI professionals — in industry and in the supervisors themselves — as a major constraint.

  • Budget: of the insurers already using AI, 41% planned to spend less than R1 million in 2024, so most programmes are still pilots.

  • Decisions land on people. A declined claim or a repriced premium affects a policyholder directly, so explainability is not optional. (This and the next two points are what we see in the work, not survey findings.)

  • Data sits in legacy policy administration systems, many with no API, and is rarely clean enough to train on without work.

  • Insurance decisions are examined years later. A life claim can be disputed long after the policy was written, so the evidence trail has to outlast the model.

Source: FSCA & PA, AI in the SA Financial Sector (2025)
04 · How to

How do you use AI in underwriting and stay POPIA-compliant?

Short answer

Minimise the data, document every feature, and keep a person in decisions that materially affect a policyholder. POPIA section 71 restricts decisions based solely on automated profiling unless appropriate measures exist — including a chance to make representations and enough information about the logic behind the decision.

  1. 1

    Write down the purpose of each data field and drop the ones the model does not need. Minimisation is cheaper at design time than at audit time.

  2. 2

    Treat health and other special personal information as its own category, with its own conditions — not as one more feature.

  3. 3

    Keep feature lineage, so any input to a price can be traced to its source and explained to a supervisor.

  4. 4

    Route declines and material loadings to an underwriter, with the model's reasons attached.

  5. 5

    Be able to explain a decision to the policyholder and record their representations, as section 71 requires.

  6. 6

    Monitor outcomes over time for drift and unfair results across groups — the fair-treatment test your conduct supervisor will apply.

Source: POPIA section 71 — automated decision making
05 · How to

How do you get insurance data ready for AI and IFRS 17?

Short answer

Build one governed, lineage-tracked data layer and run both reporting and AI from it. IFRS 17 already forced insurers to trace reported figures back to source records. The same tested tables, agreed definitions and lineage are what an AI model needs, so the work is shared rather than done twice.

  1. 1

    List every source: policy administration, claims, reinsurance, finance and any broker or UMA feeds.

  2. 2

    Land raw data unchanged, then clean and standardise it in tested layers, so every change is visible and repeatable.

  3. 3

    Agree business definitions — earned premium, incurred claims, contract boundaries — with actuarial and finance before anyone builds a model.

  4. 4

    Add data quality tests that stop a bad load from reaching a report or a model.

  5. 5

    Track lineage from each submitted figure back to the source record.

  6. 6

    Classify and control access to personal information, then expose curated features to models from the same platform.

Challenges We Solve

The hard problems Insuretech faces

Manual Claims Processing

High-volume, paper-heavy claims workflows slow settlement and drive costs. AI document extraction and classification automate straight-through processing for eligible claims.

Underwriting Accuracy

Thin data and legacy models lead to adverse selection. We build feature stores and ML pipelines that surface richer risk signals for more accurate pricing.

Fraud Detection

The Insurance Crime Bureau puts the industry's exposure to fraud and organised crime at R3.5 billion if left unchecked. The expensive cases are coordinated — shared identities, bank accounts and addresses across insurers — which per-claim rules structurally cannot see.

Regulatory Reporting

SAM, IFRS 17 and the Prudential Authority's returns demand precise, auditable data. We build the governed data foundations and automated reporting that make a submission traceable to its source records.

By line of business

Where AI fits, line by line

Insurance is not one industry. The use case that pays back first is different for a motor book, a funeral book and an underwriting manager.

First win: Claims volume

Short-term: motor, home and commercial

High-frequency, document-heavy claims are the obvious first target: first notice of loss, quotes and invoices read automatically, simple claims settled straight through, and recoveries and salvage tracked as data rather than as email threads.

First win: Claim verification

Life and funeral

Funeral and life books carry the fraud the Insurance Crime Bureau keeps reporting: fake dependants, borrowed identities and claims shortly after inception. Cross-checking beneficiaries, documents and bank accounts across the book is where models earn their keep.

First win: Reporting to insurers

UMAs and MGAs

Underwriting managers live between their own systems and their insurers' reporting requirements. Pipelines that produce reconciled bordereaux and binder reporting from source data remove a monthly manual exercise and the errors that come with it.

First win: Admin and records

Brokers and intermediaries

Renewals, policy schedules, quote comparisons and FAIS record-keeping are repetitive and rules-based. Automating the admin frees advisers for advice — while keeping the records the FSCA expects to see.

Our Solutions

What we build for you

01

AI-Powered Claims Automation

LLM agents that read, classify, and triage incoming claims documents — routing simple cases to straight-through settlement and flagging complex cases for human review.

02

Risk & Pricing Models

ML pipelines that ingest telematics, claims history, geospatial, and external data into risk signals that sharpen pricing for granular segments — with every feature documented so the actuary can defend it.

03

Governed Data Platform

A centralised, lineage-tracked data environment where every data point can be traced from source to regulatory submission — reducing audit risk and remediation cost.

Regulatory context

What South Africa's regulators have actually said about AI

Insurers keep being told AI regulation is coming. In South Africa it has already started, and the documents are public. Here is what exists as of September 2026, with dates, so you can check rather than take our word for it.

  1. Published 24 November 2025

    Joint FSCA and Prudential Authority report on AI

    The Financial Sector Conduct Authority and the Prudential Authority jointly published their inaugural report, Artificial Intelligence in the South African Financial Sector — the first comprehensive picture of AI, machine learning and generative AI adoption across South African financial institutions. It is based on a survey of the sector, which means your supervisor now has a baseline for what is normal and what is not.

  2. Stated for July 2026 — confirm current status

    A joint discussion paper, and then an instrument

    The FSCA, the Prudential Authority and the Reserve Bank said they would follow that survey with a joint discussion paper setting out recommendations for the safe and responsible adoption of AI, explicitly with a view to developing a formal joint regulatory instrument. We have not been able to confirm the paper's publication status independently, so treat the date as intent rather than fact — but the direction is unambiguous, and building for an instrument that does not exist yet is cheaper than retrofitting one.

  3. Prudential Authority communication, 24 April 2026

    Preparing for Frontier AI and AI-Accelerated Cyber Risk

    A direct communication to the financial sector whose stated purpose is to raise awareness of a material shift in the cyber-risk landscape arising from rapid advances in AI. Worth reading at board level: it frames AI as something that changes your threat model, not only your operating model — which matters when you are about to give an agent access to policyholder data.

  4. Introduced in the National Assembly, not yet enacted

    Conduct of Financial Institutions (COFI) Bill

    COFI consolidates market-conduct requirements from across the sectoral laws into a single Act, on an activity-based licensing model with principles- and outcomes-focused requirements. For insurers the practical consequence is that fair-treatment outcomes become demonstrable obligations rather than policy statements — and demonstrating an outcome is a data problem.

  5. In force

    POPIA and automated decision-making

    POPIA governs any automated processing of personal information, and automated decisions that significantly affect a data subject carry specific obligations. In claims and underwriting that is not a hypothetical: a declined claim or a repriced premium is exactly such a decision. Design for minimisation and human review from the start, because consent and masking are close to impossible to retrofit across a live book.

  6. In force

    IFRS 17 and reporting lineage

    IFRS 17 made insurance contract measurement far more data-hungry and far more auditable. Practically, every figure in a submission should be traceable to its source records without a manual reconciliation pack. Insurers that built that lineage for IFRS 17 already own most of the foundation an AI programme needs; those that patched it with spreadsheets will pay for that twice.

  7. Finding of the November 2025 report

    Skills, not models, are the binding constraint

    The regulators' own finding was that AI is not yet deeply embedded across South African financial institutions, and that a major constraint is the shortage of skilled AI professionals — in industry and in the supervisory authorities themselves. That cuts both ways: capability is scarce, and so is the supervisory capacity to assess it, which is precisely why documented, explainable design carries weight.

Where we start

Engagements we run in Insuretech

01

Straight-through claims for the clear cases

Document intelligence reads and classifies incoming claims, settles the unambiguous ones automatically, and routes the rest to an assessor with a structured summary and the reasons for referral — so the human sees a decision to make, not a pile of PDFs.

02

Underwriting models with defensible features

Feature stores and pipelines that combine claims history, telematics and geospatial data into risk signals, with feature lineage documented so a pricing decision can be explained to a supervisor rather than attributed to the model.

03

Fraud rings, not just fraudulent claims

Graph-based anomaly detection that finds relationships between claimants, providers, addresses and bank accounts — the coordinated patterns that per-claim rules structurally cannot see — tuned so investigators get a workable queue rather than noise.

04

Regulatory reporting that reconciles itself

A governed platform where IFRS 17 and statutory submissions are generated from tested, lineage-tracked tables, so month-end is a pipeline run and an audit question is a query rather than an archaeology project.

05

POPIA-safe use of policyholder data

Classification and masking applied before any model sees a record, retention and deletion enforced by the platform rather than by policy documents, and human-in-the-loop gates on decisions that affect a policyholder.

06

Legacy policy administration systems with no API

Where the book sits in a system that cannot be integrated, we automate against the interface with audit logging and session replay — deliberately, for processes whose volume justifies automation but whose integration cost does not.

Proof

Work we have shipped in regulated financial services

Same regulators' expectations, same compliance shape: automate the clear cases, route the uncertain ones to a person, evidence every decision.

Automating FICA Compliance for a South African Neo Bank with Google Vertex AI
Case study · Fintech

Automating FICA Compliance for a South African Neo Bank with Google Vertex AI

90%
Reduction in fraud detection latency
75%
Of KYC reviews fully automated
Read the case study
Cloud Data Architecture for a South African Neo Bank on Google Cloud
Case study · Banking

Cloud Data Architecture for a South African Neo Bank on Google Cloud

10×
Data processing throughput improvement
99.9%
Data quality SLA achieved at gold layer
Read the case study
Key terms

Insuretech terms, in plain language

SAM (Solvency Assessment and Management)
South Africa's risk-based prudential regime for insurers, supervised by the Prudential Authority. It is the local counterpart of Europe's Solvency II.
IFRS 17
The international accounting standard for insurance contracts. It replaced IFRS 4 for reporting periods beginning on or after 1 January 2023, and made insurance reporting far more data-intensive.
Straight-through processing (STP)
A claim or transaction handled end to end by systems, without a person touching it, within limits the insurer sets.
FNOL (first notice of loss)
The first report of a claim to the insurer — usually where automation starts, because it is where the documents arrive.
UMA (underwriting management agency)
An intermediary given authority by an insurer to underwrite, and often administer claims, on its behalf. Internationally the closest term is MGA (managing general agent).
Bordereaux
Periodic schedules of premiums or claims that an underwriting manager or cedant reports to an insurer or reinsurer.
TCF (Treating Customers Fairly)
The FSCA's outcomes-based conduct framework for fair treatment of customers. The COFI Bill is set to carry its fairness outcomes into law.
Insurance Crime Bureau (ICB)
The South African non-profit body, funded by its member insurers, that combats insurance fraud and crime — including by analysing data across members.
FAQ

Insuretech questions

From our blog

Document workflow automation: beyond the PDFPOPIA as a catalyst for better workflow automation5 data quality pitfalls silently killing your automationAI proof-of-concept red flags

Sources & further reading

Related Services
AI AgentsAI AutomationCloud Data PlatformsData EngineeringDashboards & ReportingAI Consulting

Ready to transform your Insuretech data strategy?

Let's discuss the specific challenges your business faces and design a solution that delivers real impact.

Book a free AI assessmentView our services