The Next AI Shift Isn’t Chatbots — It's Tables

Tabular AI is cheapening everything from reporting to prediction modeling

It might look less spectacular than ChatGPT, but it might just unlock tons of value. Image generated with Leonardo AI

Most insurance and finance teams don’t suffer from a lack of data. They suffer from a lack of timely, reliable decisions. The core data that drives the enterprise—actuarial triangles, claims histories, policy attributes, IFRS 17 cash flows—lives in tables, yet the process of turning that data into a forecast remains a multi-month marathon of brittle pipelines and manual effort. Actuarial and data science teams spend the vast majority of their time preparing data, not interpreting it.

This has been the accepted cost of doing business for decades. But a new class of artificial intelligence models is emerging that targets this bottleneck directly, and it has nothing to do with generating text or images. It’s about understanding the language of the balance sheet.

Prediction Is Becoming Cheap

For as long as the profession has existed, building a predictive model—for claims, for lapses, for market risk—has required specialized teams, custom-built data pipelines, and long, expensive development cycles. That fundamental constraint is beginning to disappear.

Just as large language models learned to understand grammar and semantics from the entire internet, a new category of foundation models is being trained on the native structure of enterprise data. These models can ingest raw business tables and produce forecasts directly, dramatically compressing the time it takes to get from data to a prediction. The ability to forecast is becoming a utility. Prediction is becoming infrastructure.

This is more than an incremental improvement; it is a phase change in the economics of analytics. When the cost of generating a forecast collapses, the entire organization can begin to operate at a higher frequency.

What This Means for Insurance & Finance

The most immediate impact will be a radical acceleration of existing analytical cycles. But the strategic consequences are far more profound. When prediction is cheap and fast, the feedback loop between the business and its data tightens from months to days. This enables outcomes that were previously out of reach:

• Faster, more accurate reserve estimates, allowing for more dynamic capital allocation and earlier detection of deteriorating portfolios.

• Proactive lapse management, with models that can identify at-risk policyholders in time for intervention.

• Quicker pricing feedback loops, where the performance of a new product is understood in weeks, not years.

• Operational fraud prevention, with the ability to score claims for fraud risk in near-real-time.

Strategically, this translates to shorter product development cycles, earlier and more reliable risk signals, and tighter, more responsive capital management. It lowers the dependency on a small number of scarce data science experts and empowers domain experts—the actuaries and financial analysts—to explore more scenarios and ask more questions of the data themselves.

Prediction Is Not Decision

This is the point where the hype must be met with executive discipline. These new models are exceptionally powerful at one specific task: telling you what will probably happen next, based on what has happened in the past. They are not designed to tell you what you should do.

They optimize for correlation, but they don’t understand causation. They don’t encode regulatory logic. They don’t provide governance. And they don’t have a concept of your firm’s unique risk appetite. An organization that treats a powerful predictive model as a decision-making system will, eventually, get burned.

The risks are real: silent model drift leading to unexplained reserve movements, decisions based on spurious correlations that break under stress, and significant regulatory exposure from using an unexplainable black box for a material assumption.

The Right Mental Model for Leaders

To harness the power of this technology without inheriting its risks, leaders must think in terms of a layered architecture that separates prediction from control and decision.

1. The Prediction Layer: This is where the new foundation models live. Their job is to ingest raw data and produce powerful forecasts, handling the complex statistical heavy lifting.

2. The Control Layer: This is the governance engine. It sits on top of the prediction layer, validating its outputs, comparing them against simpler, more transparent models, flagging anomalies for human review, and providing a complete audit trail.

3. The Decision Layer: This is where the business lives. It takes the validated outputs from the control layer and applies business rules, regulatory constraints, and expert judgment to make a final, auditable decision.

Winning organizations won’t replace actuaries with AI. They will give their actuaries, analysts, and executives radically better instruments. The goal is not to automate the decision, but to industrialize the production of the evidence required to make it.

What Leaders Should Do Now

For the CIO, CFO, or Head of Actuarial, this is a strategic moment. The correct response is not a series of isolated proof-of-concepts, but a deliberate architectural shift.

• Treat tabular AI as strategic infrastructure, not a science experiment. The ability to generate fast, reliable predictions from core business data is a foundational capability for the future.

• Start with internal analytics before tackling regulated reporting. Use the technology to accelerate internal pricing studies, experience analyses, and operational dashboards to build confidence and experience.

• Always pair prediction with governance. Never deploy a predictive model without a control layer to validate its outputs and a human in the loop for key decisions.

• Invest in the control and decision layers, not just the models. The long-term competitive advantage will come from the quality of your governance and decision-making frameworks, not from having a slightly better predictive model.

The Future Is Structural

The next wave of enterprise AI isn’t conversational. It’s structural. It’s about fundamentally rewiring the process by which data is turned into insight and insight is turned into action.

The winners won’t be the companies that predict better. They’ll be the companies that decide better.

---

Reads of The Week

• This sharp critique of Cursor by Devansh argues that while the AI coding tool has gone viral among startups and influencers, it’s dangerously ill-suited for enterprise software development. The author details serious concerns—from security lapses and hallucinated code to unreviewable pull requests, messy multi-file edits, and unreliable customer support—suggesting that Cursor’s design clashes with the stability, compliance, and workflow rigor large organizations require. This piece is a sobering reminder that productivity gains mean little if they come at the cost of trust, governance, and long-term maintainability.

• If you’re trying to figure out which AI coding tools are actually worth your time, this creator roundup by Jeff Morhous and colleagues is insightful. Fourteen builders share what they really use—often some mix of Cursor, Claude Code, Replit, Lovable, v0, and CLI agents—and, more importantly, how they use them in practice. For Wangari Digest readers tracking how AI is reshaping real software workflows (not just hype cycles), this piece offers a grounded look at the emerging patterns—and a chance to steal a few battle-tested setups for yourself.

• For years, tabular machine learning has belonged to XGBoost, LightGBM, and CatBoost—but this piece by Duong Nguyen and Nicolas Chesneau argues that era may be ending. It introduces the rise of tabular foundation models like TabPFN, CARTE, and TARTE, explaining how synthetic pretraining, in-context learning, and semantic knowledge injection could fundamentally change how we build and update models on structured data. It’s a thoughtful look at how foundation-model thinking is finally coming for the “last bastion” of classical ML.