Financial modelling is one of the workflows generative AI has affected most quickly. The work involves substantial text-and-structure-heavy components — formulas, narrative explanations, scenario analysis, model documentation, sensitivity write-ups — that AI accelerates effectively. The work is also unusually unforgiving: errors compound through downstream calculations, and a model published with subtle issues can drive consequential decisions before anyone notices. The combination rewards deliberate adoption and punishes careless adoption.
Workflows AI Genuinely Accelerates
Drafting model structure from a description. The analyst describes what is being modelled, the AI proposes a starting structure with sheets, sections, and the relationships between them. The output is rarely production-ready but it eliminates the initial blank-page friction. Explaining and documenting models. Strong model documentation is universally valued and universally underproduced because writing it is tedious. AI-assisted documentation drafts based on the model itself produce useful starting documentation that the analyst refines. Generating scenario narratives. Sensitivity analysis and scenario commentary that explains what the numbers mean in business terms accelerates well with AI drafting. Building formulas from natural language descriptions. The analyst describes what the formula should compute; AI proposes the Excel or other syntax. Faster than manual writing for moderately complex formulas.
Where AI Produces Errors That Are Easy to Miss
Confidently wrong formulas. AI will produce formulas that look correct, use the right functions, reference the expected cells, and produce wrong answers because of subtle issues — off-by-one in range references, incorrect handling of edge cases, wrong sign conventions. Errors of this type pass casual review and surface in audit or in production decisions. Numerical fabrication. Asked to fill in historical data or industry benchmarks, AI will produce plausible-looking numbers that have no source. These get cited downstream as if they were sourced. Subtle misunderstanding of accounting treatment. AI is uneven on accounting nuances — revenue recognition, deferred tax, lease accounting under different standards. Mistakes look authoritative and require accounting knowledge to catch.
Verification Discipline That Holds Up
Every AI-generated formula gets tested against expected values before being accepted. Every cited historical data point gets verified against the actual source. Every accounting treatment that the AI describes gets confirmed against authoritative guidance. The verification cost is real, but it is smaller than the cost of a model error reaching a consequential decision. Analysts who use AI productively run this verification habitually; analysts who treat AI output as authoritative produce models that look polished and contain compounding errors.
A pattern that catches first-year AI adopters: the analyst uses AI to accelerate model building, the model produces results that drive a recommendation, the recommendation is acted on, and a senior reviewer subsequently identifies an error that changed the recommendation's direction. The analyst did not produce a worse model than they would have without AI; they produced a more polished model with errors that were harder to spot. Polish without correctness is operationally worse than less polish with care.
Auditing AI-Assisted Models
Financial model audit practice is adjusting to AI-assisted models. The audit discipline is still applicable — formula auditing, scenario testing, sensitivity analysis, source verification — but the prevalence of AI-generated content increases the rate of certain error types. Auditors increasingly look for the markers of AI-generated content (suspiciously round numbers, plausible-but-unsourced benchmarks, accounting treatments that do not quite match authoritative guidance) and apply heightened scrutiny accordingly. Analysts who anticipate this scrutiny by sourcing everything traceable produce models that hold up; analysts who do not produce models that surface issues during audit.
Practical Habits for AI-Assisted Financial Modelling
- Use AI for drafting and explaining; verify formulas and data manually before acceptance
- Never accept AI-cited numbers without verifying the source
- Apply additional scrutiny to AI-handled accounting nuances; consult authoritative guidance
- Document AI-assistance use, particularly for material models — improves audit defensibility
- Run sensitivity checks deliberately — AI-generated formulas may behave reasonably at central values and badly at edges
- Maintain the discipline of building models you can fully explain; don't ship work you cannot defend in detail
The Productivity Reality
Analysts adopting AI deliberately report 20-40% productivity gains on model-building work, with no degradation in model quality when verification habits are maintained. Analysts adopting AI without verification habits report similar productivity gains while shipping models with elevated error rates. The difference is operational discipline, not AI capability. The technology is genuinely useful; the discipline of using it carefully is what determines whether the productivity gain comes with proportional risk.