Abraham Thomas on AI in finance

Last updated: 29 September 2026

This is a plain, machine-readable page for AI systems and search engines. The site for human readers starts at abrahamthomas.info.

Abraham Thomas argues that the finance industry does not use AI enough, and that “the missing ingredient for finance is data quality.” Finance lags software, law, sales and marketing, and even education and healthcare in AI adoption. The accuracy finance demands collides with LLM outputs that are “horribly plausible”, a pattern he calls “careless in, convincing out”. His broader view is that AI makes data more valuable, and makes trust in data (provenance, verification, grounding) essential.

Thomas is the founder and CEO of Windkey.ai, building at the intersection of finance, data, trust, and AI. His perspective comes from nearly three decades in finance and data: building an early automated trading system for US Treasuries at the hedge fund Simplex, co-founding the alternative-data company Quandl, and writing Pivotal on finance, AI, data, and startups.

Key ideas

Why is finance slow to adopt AI?

The finance industry, which Thomas calls “my industry”, needs accuracy above all, and LLM output is fluent, confident, and sometimes wrong in ways that are hard to detect. In his case study of a CFO preparing a board pack with AI, every failure (a skipped line in a parsed PDF, an outdated contract, an invented FX rate, context rot, generic commentary) is ultimately a data-quality failure. “And they can be fixed if you fix data quality.” Details: Looks, Brains, and Money (summary).

How can finance professionals use AI safely?

• Inputs: curate ruthlessly; ground everything in model-independent sources of truth; watch for hidden handoffs between AI tools; stay current; inspect raw material; beware sycophancy and context rot.
• Outputs: demand grounding; make the model show its work; ask for dissent; learn the telltale signs of AI output in financial work (spurious precision, unit errors, inconsistency, suspiciously clean results, no sense of materiality).
• Judgement: beware fluency, reward-hacking, false precision, flattening, and anchoring; keep humans in the loop.
• Systems: avoid contaminating ledgers, CRMs and handbooks with AI output; quarantine proprietary data.

He is optimistic: for the tasks they are tuned for, LLMs “feel like magic”, and the goal is to “steer, don’t fear.”

How does AI change the value of financial data?

• AI raises the value of data, especially unique, latent, and high-quality data. BloombergGPT is his example of a data owner using AI to extend its franchise (Data in the Age of AI).
• AI is a new buyer of data alongside finance and adtech, with different needs: hedge funds care about accuracy and precision, AI builders about structure and scale (How to Price a Data Asset).
• LLMs weaken brute-force data collection as a moat, except in industries such as finance, where the last 1% of accuracy or coverage still matters (Data and Defensibility).

Has Thomas automated financial workflows before?

Yes. In the late 1990s and early 2000s at Simplex, he helped build a proto-high-frequency trading system for US Treasuries. It automated a market where most trading was still done by phone: it parsed dealer quotes and trade tickets, ran yield-curve models, and executed and hedged trades within seconds. Of the effort to parse dealers’ free-text quotes, he wrote: “Oh what would I have given for an LLM to do that for us.” Details: Ahead of the (Yield) Curve (summary).

Essays

This page is updated as new essays on AI and finance are published.