A great unbundling is happening in finance: the unbundling of the production of financial models from their verification. In the antediluvian period before large language models (LLMs), building a financial model and understanding how it worked were largely the same activity. A CFO, an analyst, an associate, or a banker who built a financial model generally knew why each cell held this number and not that number, why such and such a formula was used, what its assumptions meant and what each schedule should reconcile to. This did not guarantee that the model was correct relative to specified evidence, definitions, rules and tolerances, but the knowledge required to construct it was accumulated in the process of building it.
Now, one can ask an LLM to build a financial model, without needing to understand that model, or having to build any portion of it. While the responsibility for the financial model has not shifted, production of financial models has become detached from the knowledge required to verify it. The person approving the financial model is now responsible for work they did not construct and cannot readily defend. As a consequence, a new, meaningful and rising cost of verification has emerged. The “asymmetry of verification”1 is such that it is now considerably cheaper to produce a financial model, but not proportionally cheaper to verify the correctness of that financial model.
The economic consequences are palpable. Productivity gains from using LLMs are eroded by verification costs, as professionals must trace formulas, reconcile figures, locate source documents, reconstruct assumptions and acquaint themselves with the logic of work they did not build. When Macabacus surveyed 75,000 of its users, it found that 62% believe they have shipped a financial model or presentation that contains an error generated by LLMs, and 45% believe they probably have, “but no one noticed”.2 A lofty 24% of respondents report having comprehensive guardrails against such errors. The further up the corporate hierarchy, the less confident respondents were about the accuracy of these LLM-built financial models. In that same survey, respondents report spending at least 30 minutes verifying each financial model or presentation built by an LLM.
Finance faces a choice between relying on human review of algorithmically generated financial model, thereby eroding the productivity and financial gains from algorithmic production; increasingly elaborate attempts to make LLMs check themselves; or, accepting work whose correctness cannot be confidently established. This race between the falling cost of algorithmic thought and the cost of verifying algorithmic production is one bounded by human attention. The resulting “measurability gap” as Christian Catalini, Xiang Hui and Jane Wu call it, will lead to verification becoming “a primary production technology” rather than a compliance function.3 A fourth choice is open: finance can develop independent, automated systems capable of establishing whether financial work is correct. The objective is not to replace professional judgement, but to free it from the mechanical burden of checking what can be checked automatically.
This is not the first time an expansion in production has required a corresponding expansion in verification. The growth of long-distance commerce was accompanied by the development and dissemination of double-entry bookkeeping, described by Luca Pacioli in 1494.4 The expansion of joint-stock companies brought new systems of statutory reporting, audit and professional accountancy. Mass manufacturing gave rise to statistical quality control, developed by Walter Shewhart in the 1920s. The growth of securities markets was followed by the Securities Acts of 1933 and 1934 and a new framework of standardised disclosure and financial accountability. The revolution that has ushered in this age of algorithmic thought will not be any different. Every major expansion in the scale and complexity of production has required new institutions and technologies for establishing the reliability of what is produced.
The next stage of financial automation will not simply be about producing financial models. It will be about verifying them. Ultimately, it will make possible something finance has never had at scale: proof-carrying financial models.
Notes
- 1. Jason Wei, ‘Asymmetry of Verification and Verifier’s Rule’, Jason Wei, 15 July 2025, https://www.jasonwei.net/blog/asymmetry-of-verification-and-verifiers-law.
- 2. Macabacus, GenAI in Financial Services: Velocity and Verification (Macabacus, 2026), 14, https://macabacus.com/lp/ai-report.
- 3. Christian Catalini et al., ‘Some Simple Economics of AGI’, arXiv:2602.20946, preprint, arXiv, 25 February 2026, https://doi.org/10.48550/arXiv.2602.20946.
- 4. Luca Pacioli, Summa de Arithmetica, Geometria, Proportioni et Proportionalita (Paganino Paganini, 1494), https://icaew.onlineculture.co.uk/ttp/?id=b2f80d6f-2e46-4cd5-9e8f-ac7a00ef4390&type=book.
Cite this essay
Chicago Manual of Style
- Bibliography
- Noko, Joseph. “The great unbundling: LLMs have separated financial production from verification.” The Mirandolan (blog), November 18, 2026. https://josephnoko.com/the-great-unbundling-llms-have-separated-financial-production-from-verification/.
- Note
- Joseph Noko, “The great unbundling: LLMs have separated financial production from verification,” The Mirandolan (blog), November 18, 2026, https://josephnoko.com/the-great-unbundling-llms-have-separated-financial-production-from-verification/.
- BibTeX
@article{noko2026the, author = {Noko, Joseph}, title = {{The great unbundling: LLMs have separated financial production from verification}}, journal = {The Mirandolan}, year = {2026}, month = {nov}, day = {18}, url = {https://josephnoko.com/the-great-unbundling-llms-have-separated-financial-production-from-verification/}, urldate = {2026-11-18} }