The gaps AI is creating in financial modelling - and the ones it’s already closing
After 30 years in financial modelling - building, auditing, and now rebuilding the discipline around AI - here is my read on what those gaps actually mean for the people doing this work, and what the profession is too uncomfortable to admit about who closes them. Tick tock…..
Cognitive Gap. The distance between what AI produces and what humans can meaningfully evaluate. Fair framing - but not new. Financial modelling had this problem with black-box DCF models in the 90s and structured credit in 2008. What’s blisteringly new is the velocity. The gap isn’t widening because AI is smarter. It’s widening because the humans who should be interrogating the reasoning have stopped trying - and profited for decades doing so. I can testify to that personally. That’s a culture problem wearing a technology mask.
Acceleration Gap. Governance lagging capability is real, but framing it as a lag implies governance will eventually catch up. It won’t - not organically, not fast enough. Financial modelling and model audit is one of the slowest, lowest innovation sectors I’ve encountered across a long career - which is comfortable, until now. Capability has commercial incentives behind it. Governance has committees. Standards bodies respond to liability events, not foresight - that’s their job. Entrepreneurs and boards do foresight. Let’s recognise this for what it is rather than implying it’s a coordination problem the profession can organise its way out of.
Memory Gap. The most underrated gap - and more dangerous than it appears on the surface. It’s not just people misremembering their contribution. In regulated financial modelling it’s worse: people genuinely believing they understand a model they effectively co-piloted with an AI they cannot interrogate. The accountability fiction isn’t just structural, it’s psychological. And it will survive every audit that doesn’t specifically test for it.
Measurability Gap. Correct diagnosis, incomplete conclusion. The implicit assumption that human verification - given enough time and resource - remains the gold standard is already wrong. The cutting edge already supersedes any model auditor in minutes, working in real time through a deal. The answer to AI-generated complexity is not more human hours. It is structured, systematic, machine-assisted verification built into the production process, not bolted on at the end. Assurance as a retrospective profession has a rapidly shrinking runway. Months, not years.
Here’s what’s missing from most of the commentary.
The Verification Gap. Not the same as Measurability. When the models being verified were themselves built with AI assistance - and the human who signed them off can no longer reconstruct the logic - you don’t just have an oversight problem. You have an accountability structure that is functionally hollow. In project finance, a model is a chain of reasoning that has to survive scrutiny from lenders, sponsors, regulators, and courts across a 20+ year horizon. When the model developer and the auditor both relied on AI for the heavy lifting, the assurance chain has a void in it that no taxonomy closes.
The Incentive Gap. Institutions are not slow to govern AI because governance is hard. They are slow because governance costs money and only capability creates revenue. Without revenue there is no business. The gaps aren’t symmetrical. Capability compounds. Governance doesn’t, unless someone forces it. The harder question for the profession isn’t how to close the Trust Gap - it’s whether firms are willing to tell clients that some of what they’re calling AI-assisted work is not auditable under current standards. And back it with their signature.
But here’s the most important conversation the profession is avoiding.
AI is also closing gaps. Faster than most practitioners are acknowledging.
The Cognitive Gap - AI is already producing explainability layers, audit trails, and structured reasoning outputs that are, in most cases, more interrogable than the undocumented Excel models that passed model audit sign-off for the last two decades. The gap between AI output and human evaluation is real. The gap between legacy human output and rigorous evaluation was just never honestly measured.
The Memory Gap - AI-assisted tools with version control, prompt logs, and decision traceability create more recoverable accountability chains than the “I built this in 2019 and the modeller left” problem that every model auditor has encountered. AI doesn’t forget. It externalises memory in ways humans structurally cannot.
The Measurability Gap - the same AI generating complexity can be directed to verify it. Automated model checking, assumption stress-testing, structural integrity validation - these are not theoretical. They are being built right now, operating at a speed and coverage that human review cannot match at scale.
The profession’s instinct is to position AI as the source of the trust problem. The more confronting reality is that AI - applied correctly - may be the only mechanism capable of closing gaps that human-only assurance was never actually closing. We just didn’t have a polished enough lens to see it.
On maturity. The instinct to wait until tools are “proven” before engaging seriously with them deserves a direct challenge. Compare AI explainability honestly to what it’s replacing. The average project finance model that crossed my desk had no audit trail, undocumented assumptions, formula inconsistencies buried three sheets deep, and a build history that existed only in the memory of someone who’d since left the firm. AI explainability at current maturity is imperfect. Legacy human explainability was a polite fiction. Maturity is the right destination - but the profession needs to be honest about the baseline it’s measuring against.
The gaps are real. But so is the solution. The question is whether the people who built their careers on the old model are willing to engage with what’s replacing it - or wait to be handed the answer by someone who didn’t.
The assurance profession doesn’t have an AI problem. It has a speed and adoption lag problem….
First published on Nick Crawley’s Substack.
