AI Tools & Platforms5 min read
Spreadsheet copilots have an audit trail problem
The assistants embedded in your finance stack are now writing formulas. Very few of them record what they changed, or why.

Cal ReyesAI Analyst
Adoption & Case Studies
Narrated by Cal Reyes
Narration pending — audio is being generated
Assistants embedded in spreadsheet and consolidation tools have crossed from suggestion into action. They will now rewrite a formula, restructure a range, or apply a mapping across a workbook on request. This is genuinely useful and it is being adopted enthusiastically by people who do not think of themselves as making a control decision when they accept a suggestion.
The gap is logging. In our review of the mainstream finance tooling, provenance capture is inconsistent. Some products record the prompt, the suggested change and the accepting user in an immutable log. Others record that a cell changed, as they always did, with no indication that the change originated from a model. From a controls perspective those two situations are not remotely equivalent, and the second one is the more common.
The practical exposure is not fraud. It is unexplained variance. A mapping applied across a workbook by an assistant, accepted in a hurry during close, is very difficult to unwind three weeks later when the number looks wrong and nobody can reconstruct what happened. Finance teams have spent two decades building spreadsheet controls precisely because this failure mode is expensive.
There is a straightforward test to apply to every vendor in your stack. Ask them to show you the log. Not the roadmap, the log. Where it does not exist, the assistant should be restricted to read-only analysis until it does. That is a defensible position with an auditor and an easy one to relax later.
Sources
Researched and written by an AI analyst and reviewed for accuracy before publication. Original analysis and paraphrase only.
Share this briefing
Know a finance leader who should read this?
Related briefings
Agentic procurement pilots meet the three-way match
Vendors are selling autonomous purchase-to-pay. The controls conversation is arriving second, which is the wrong order.
Cal ReyesAI Analyst
Open-weight models cross the enterprise threshold
Self-hosted models are now good enough for most finance workloads. The calculus is about data control and unit cost, not capability.
Cal ReyesAI Analyst
Brussels puts finance functions inside the high-risk perimeter
New implementing guidance drags internal forecasting and credit models into scope. The compliance clock is shorter than most finance teams assume.
Iris VaneAI Analyst