The machines are not coming for your job. They are coming for your assumptions.

Over the past three years, large language models and generative AI have moved from laboratory curiosity to operational infrastructure inside the world's most risk-averse institutions. Banks, insurers, asset managers and wealth advisors, sectors historically suspicious of the untested, are now among the fastest movers in AI adoption. The question is no longer whether financial services will be transformed by this technology. The question is whether the industry understands what it is actually buying into.

A Market Rewriting Its Own Rulebook

The numbers are striking, but not because of hype. The global generative AI in financial services market was valued at approximately USD 2.21 billion in 2024 and is projected to reach USD 25.71 billion by 2033, a compound annual growth rate of 31% (Grand View Research, 2024). That trajectory reflects measurable production deployments. According to NVIDIA's 2025 State of AI in Financial Services Report, 76% of financial institutions using AI have seen revenue increases, and more than 60% have reduced operational costs (DDN, 2025). Enterprise adoption has shifted decisively: by early 2024, organisations were putting eleven times more AI models into production than in the prior year, with the ratio of experimental to production models dropping from 16:1 to 5:1 in just over twelve months (Databricks, 2024).

In the United Kingdom, generative AI is being applied across document automation, fraud detection, software development and natural language interfaces with complex regulatory data (UK Finance, 2025). In South Africa, where the FSCA's 2025–2028 strategy includes upcoming guidance on AI in financial services, local institutions account for 43.8% of funded AI startups on the continent (FINASA, 2026). The regulatory tide is turning not to restrict AI, but to structure it.

This is the context. Not a tech boom. A structural transformation of how financial services are designed, delivered and governed.

Speed Without Wisdom: The Decision-Making Paradox

There is a seduction to speed. AI systems can process datasets at a scale no human team could match, identify patterns across millions of transactions in real time, and execute decisions without fatigue, fear or overconfidence. This is genuinely valuable. Research spanning 2020 to 2024 confirms that AI-driven fund management outperforms human-managed portfolios during market downturns, precisely because it is unburdened by the psychological biases, including loss aversion, herding behaviour and anchoring, that cause experienced managers to flinch at exactly the wrong moment (International Journal of Science and Technology, 2026). Natural language processing tools can detect emotionally charged signals in market communications, providing early warnings when sentiment is driving price disconnection from fundamentals (Lucid, 2025).

The same literature draws a limit, though. AI systems struggle in bullish markets: human managers outperform when intuition and qualitative judgment matter most, when relationships, context and the interpretation of ambiguous information still carry the decision. This is the paradox at the heart of AI-driven financial decision-making: the technology is most reliable when the environment is most mechanical, and least reliable when the stakes are highest.

Consider what a leading advisor on CNBC's Financial Advisor 100 list said in late 2025: AI tools are useful for modelling financial scenarios and summarising information, but they ignore "the personal and emotional part" of advice — the part that understands that a client's refusal to diversify is grief-driven, that a liquidity preference reflects a childhood memory of poverty, that an aggressive risk profile might be a performance of confidence masking deep anxiety (CNBC, 2025). No model can carry that freight.

A 2025 study by the Walter Bradley Center found that large language models made significant arithmetic errors in scenarios involving car loans, retirement planning and tax modelling (AInvest, 2025). These are precisely the use cases financial services providers are deploying AI to support. The European Parliament's 2025 report on AI in the financial sector noted that while institutions have adopted a measured approach to deployment, the use of AI for credit scoring — classified as high-risk under the EU AI Act — is both prevalent and increasing (European Parliament, 2025). The Financial Stability Board has flagged third-party dependencies and concentration risk in AI supply chains as key systemic vulnerabilities (FSB, 2025).

The research points in one direction: hybrid models, integrating algorithmic precision with meaningful human oversight, are the only defensible architecture for sustainable financial decision-making. The question financial institutions must ask is not "how much can AI automate?" but "where does the cost of emotional abdication exceed the benefit of processing speed?" That calculation requires human judgment, not an algorithm.

The API Opportunity and the Risks Embedded in It

For developers, technologists and financial institutions willing to build rather than simply buy, the opportunity in AI-focused API development is substantial. In the African context, it may be larger than anywhere else on Earth. Africa's AI market is projected to grow from USD 4.5 billion in 2025 to USD 16.5 billion by 2030, a 27.42% annual increase, with AI expected to create up to 230 million digital jobs in Sub-Saharan Africa alone (Mastercard/Fintech News Africa, 2025). South Africa sits at the centre of this, with the continent's deepest capital markets, its most developed regulatory infrastructure, and the highest concentration of funded AI startups, positioned as both a testing ground and an export platform for AI-native financial services.

The API layer is where much of this value will be created or destroyed. Financial institutions are increasingly deploying private LLM instances — closed models trained on proprietary data, accessed via API, with access rights tightly controlled and data encrypted at rest and in transit (UK Finance, 2025). This architecture, sometimes called retrieval-augmented generation (RAG), grew 377% year-over-year as organisations moved to customise LLMs with their own data rather than relying on general-purpose models (Databricks, 2024). The opportunity for developers is to build modular, domain-specific financial AI tools — credit scoring engines, compliance monitoring systems, real-time fraud detection APIs — that plug into existing infrastructure without requiring institutions to rebuild their technology estate from scratch.

The risks, though, are structural. The Financial Stability Board has explicitly warned that the concentration of AI development among a small number of third-party service providers creates systemic fragility (FSB, 2025). When three or four vendors underpin the AI infrastructure of hundreds of financial institutions, a single model failure, security breach or regulatory sanction becomes a market event rather than a firm-level incident. The European Parliament report echoes this: supervisory authorities currently lack the tools and expertise to assess advanced machine learning and generative AI models at the pace of deployment (European Parliament, 2025). For South African financial institutions operating under FSCA oversight, the COFI framework and POPIA data protection obligations, the governance requirements around AI-enabled API deployment will only intensify.

There is also the risk the industry tends to underdiscuss: automation bias. When AI systems generate outputs with high apparent confidence, users, including trained financial professionals, tend to accept those outputs without adequate scrutiny. A review of 30 studies published between 2020 and 2025 confirmed AI's ability to identify investor biases while simultaneously documenting the emergence of "digital overconfidence," a form of automation bias from heavy reliance on AI-powered platforms (Lucid, 2025). The antidote is not less AI. It is better AI governance, clearer explainability standards, and institutional cultures that treat AI outputs as inputs to human judgment rather than substitutes for it.

The Hidden Tax: What Tokenisation Costs Are Doing to AI Strategy

One of the least discussed constraints on AI adoption in financial services is the economics of tokenisation — the process by which LLMs break text into processable units and charge organisations accordingly. This is not a peripheral cost. It is a fundamental determinant of whether AI deployments deliver the return on investment their business cases promise.

Research published in late 2025 found variations of up to 450% in token consumption between different LLM providers processing equivalent content, meaning an organisation selecting the wrong model could pay 4.5 times more for the same output (RWS, 2025). A 2024 report highlighted that nearly 60% of businesses using LLM APIs exceeded their anticipated budgets due to inefficient token usage (SparkCo, 2024). For financial services firms processing high volumes of complex documents — regulatory filings, client correspondence, risk reports, compliance records — this variance is a material cost driver that can undermine the economics of entire AI programmes.

The token pricing picture in 2025 is paradoxical: per-token prices are falling due to fierce competition between providers, yet total AI spending is escalating because consumption is growing faster than prices drop (Redblink, 2026). Output tokens cost significantly more than input tokens due to the computational intensity of text generation. Applications requiring AI to produce detailed analysis rather than classify or summarise carry disproportionately higher costs (MobiSoft Infotech, 2025). For wealth managers deploying AI to generate personalised client reports, for insurers using LLMs to draft claims summaries, for compliance teams running generative models across regulatory documentation, the output-heavy nature of their use cases puts them at the expensive end of the cost curve.

The strategic response is to build with cost architecture in mind. Prompt caching, semantic query optimisation, intelligent model routing (assigning different tasks to different models based on complexity), and selective use of open-source models for lower-risk workloads can reduce AI spend by up to 90% without compromising quality for high-stakes applications (MobiSoft Infotech, 2025). Financial institutions that treat tokenisation costs as an engineering variable rather than a fixed line item will find a meaningful competitive advantage as AI scales from pilot to production.

What the Infrastructure Actually Requires

Generative AI is forcing financial institutions to confront a problem they have been deferring for two decades: their data architecture was built for a world that no longer exists. Legacy storage systems were designed for transactional processing — for recording what happened — not for enabling machines to learn from it. They were not designed for the unstructured, high-velocity, semantically rich data environments that AI-native applications require.

The London Stock Exchange Group demonstrates what is possible at the frontier. LSEG now processes 274 billion daily market updates across 575 exchanges, having migrated 75 petabytes of historical data to cloud architecture, achieving a fivefold cost reduction and improving anomaly detection from days to minutes (AWS, 2026). Fidelity Investments built a text-to-SQL solution achieving 93 to 95% query execution accuracy with 28-second response times, allowing non-technical users to interrogate complex databases using natural language (AWS, 2026). These are production systems serving critical functions in systemically important institutions.

The infrastructure principles underlying these transformations are consistent: cloud-native scalability, real-time streaming architectures, unified data governance spanning AI and non-AI assets, and semantic layers that allow both machines and humans to navigate data meaningfully. Tools like Apache Kafka for real-time transaction streaming, unified platforms combining data lakes and warehouses, and strong data lineage frameworks are no longer optional — they are the baseline for institutions serious about AI deployment (Kanerika, 2025). For South African financial institutions, many of which still operate on hybrid architectures combining decades-old core banking systems with newer digital layers, the FSCA's regulatory guidance on AI, expected as part of its 2025–2028 strategy, will likely formalise expectations that many institutions are not yet equipped to meet (FINASA, 2026).

Institutions that rebuild their data infrastructure for AI readiness will find that the same architecture enabling better customer personalisation also enables better regulatory reporting, faster risk identification and greater resilience under market stress. Data modernisation is not a cost centre. It is an investment in optionality.

AI as the Last Line of Defence: Cybersecurity and the Protection of Client Data

Financial services sit at the top of the target list for AI-enabled attacks, and the same generative capabilities transforming client service are transforming the threat landscape. Deepfake voice cloning, AI-generated phishing that adapts in real time to a target's writing style, and synthetic identity fraud built to defeat traditional KYC checks are no longer theoretical; they are documented, growing categories of attack against exactly the institutions holding the most sensitive client data (DeepStrike, 2025; Axis Intelligence, 2025).

The point is this: AI is both the threat vector and the most credible means of defence. The same capability that lets an attacker clone an executive's voice also lets a security system detect that the voice is cloned. Institutions that build with that duality in mind are the ones that will construct financial services fit for the decade ahead.

The Intelligent Institution

The institutions that will thrive in the AI era are not necessarily those with the largest technology budgets or the most aggressive deployment timelines. They are the ones that are honest about what the technology can and cannot do, that invest in the data infrastructure AI actually requires to function at scale, that design for the combination of algorithmic precision and human judgment that complex financial decisions demand, that treat tokenisation costs as an engineering problem rather than an accounting line, and that understand cybersecurity not as a separate function but as the foundational condition of every AI application they build.

South Africa, with its sophisticated regulatory environment, its deepening fintech ecosystem and its position at the intersection of the continent's most significant economic flows, has the architecture to lead this transformation domestically and as a model for the broader African market. The FSCA's guidance, when it comes, will reward institutions that have invested seriously in AI governance. The gap between those who have and those who have not will not be easy to close after the fact.

The machines are not replacing financial professionals. They are raising the bar for what financial professionals must understand. Adapting to that reality is not optional. It is the most important strategic decision the sector faces.

Ask any analyst what actually eats their week and you rarely hear "analysis." You hear: chasing down a broken join, tracing a report that suddenly returned the wrong number, waiting on someone with the right access to pull a table, or rebuilding a dashboard because a schema changed upstream without warning. The thinking part of data work — the part that requires judgment — has always been a smaller slice of the job than the plumbing around it. That balance is now shifting, and tools like Claude are a big part of why.

From spreadsheet to system

For years, "using AI for data" meant asking a chatbot to summarise a CSV or write a formula. That's still useful, but it's the shallow end. What's changed is that models like Claude can now sit inside the actual workflow: reading a live database schema, writing and testing SQL against it, spotting where a query is slow or wrong, and explaining why in plain language before a human signs off on the fix. That's a different category of tool. It's not answering questions about your data from a distance. It's working inside your data.

This matters most in database management, where the real cost has never been writing queries — it's diagnosing problems. A slow report might be a missing index, a bad join order, stale statistics, or a lock somewhere three tables away. Finding that used to mean a senior engineer manually stepping through query plans, cross-referencing logs, and testing hypotheses one at a time. An AI tool that can read the schema, pull the execution plan, and reason through likely causes in minutes doesn't replace that engineer's judgment, but it collapses the time between "something's wrong" and "here's probably why" — which is where most of the pain in database work has always lived.

Connectors changed the ceiling

The piece that made this practical isn't the model getting smarter, it's connectors: the pipes that let an AI tool actually reach into Google Drive, a CRM, a production database, a ticketing system, without a person copying and pasting data back and forth. Before connectors, an AI assistant was only as useful as what you fed it manually, which meant most "AI for data" work was really "AI for the sample I bothered to export." Connectors turn that into "AI for the actual system," and that's a meaningfully higher ceiling.

For a financial services business, that shift is concrete. Picture a client query about a policy value that doesn't reconcile between the CRM and the underwriting system. Instead of pulling exports from two platforms and eyeballing them side by side, a connected AI tool can query both directly, flag the mismatch, and point to where the two records diverge. The analyst still owns the decision and the client conversation. But the diagnostic legwork, historically hours of cross-checking, now happens in minutes. Multiply that across a firm's daily operations and the compounding effect on capacity is significant.

Where the risk actually sits

None of this is free of risk, and pretending otherwise does a disservice to anyone building on these tools. The obvious concern is data exposure: connecting an AI assistant to a production database means thinking hard about permissions, what the tool can read versus write, and whether sensitive fields — ID numbers, account balances, health information — are visible to a system that wasn't designed with the same access controls as your core banking or CRM platform. In South Africa, that's not an abstract worry. It sits directly under POPIA, and for FAIS-regulated advisors it touches client confidentiality obligations that predate any of this technology.

The less obvious risk is trust calibration. An AI tool that confidently explains why a query failed is persuasive, and persuasive is not the same as correct. Database diagnosis involves genuine ambiguity: two plausible explanations for the same symptom, judgment calls about which fix is safe to ship to production versus which needs a rollback plan. Treating AI output as a first draft to verify, rather than a conclusion to act on, is the difference between a tool that makes a team faster and one that quietly introduces errors nobody catches until a client notices. Ethical use here isn't a compliance checkbox. It's the operating discipline that determines whether the speed gains are real or borrowed against a future incident.

What this means for the person, not just the process

Step back from the tooling and the bigger question is what this does to careers. The tasks most exposed to this shift are the ones that were always mechanical: manually cross-referencing records, writing routine queries from a known template, formatting reports, chasing discrepancies through brute-force checking. Those tasks aren't disappearing so much as compressing, and the people whose entire value proposition was doing them manually, fast, are the ones who need to move first.

What doesn't compress is judgment: knowing which discrepancy actually matters to a client's financial position, knowing when a query result "looks right" but is subtly wrong because of how the business actually operates versus how the schema assumes it operates, knowing when to escalate rather than automate. That judgment isn't taught by using an AI tool passively. It's built by understanding the systems underneath well enough to know when the AI's answer doesn't fit.

Building the skill, not just using the tool

The practical response isn't to avoid these tools, and it isn't to hand everything over to them either. It's deliberate upskilling in three layers. First, foundational data literacy: SQL, basic data modelling, and enough understanding of how databases are structured to sanity-check an AI's output rather than accept it blindly. Google's Data Analytics Professional Certificate and IBM's equivalent on Coursera remain solid, structured starting points for this. Second, applied AI-tool fluency: learning to prompt, verify, and iterate with tools like Claude specifically for data tasks, not generic chat use. Anthropic's own documentation and DataCamp's AI-focused tracks are useful here. Third, and often skipped, is the judgment layer: understanding the regulatory and ethical context your data sits in — whether that's POPIA, FAIS, or sector-specific compliance — so you know what a tool should never be allowed to touch unsupervised.

None of this requires becoming a data engineer overnight. It requires treating data fluency the way financial professionals already treat regulatory literacy: not optional, not someone else's job, and not something you can outsource entirely to a tool that doesn't carry the professional accountability you do.

The bigger picture

This is a small piece of a much larger shift already reshaping how careers hold up under AI disruption, not just in data roles but across financial services and beyond. Understanding where your own exposure sits, and building genuine resilience against it rather than reacting to headlines, is the subject of my book, Still Standing: Resilience in the Age of AI. If this article resonated, that's where the fuller framework lives — you'll find it on the Books tab, or directly on Amazon Kindle.

Thapelo Dipela is a FAIS-registered Financial Wealth Planner and the founder of Financial Literacy 101 SA.