Why the audit trail is the real product in financial AI
By Kumar Ujjwal, Founder and CEO, DwellFi
The conversation about AI in finance usually fixates on how capable the model is. In regulated fund operations, that is the wrong thing to measure. The thing that makes an AI answer usable is the trail that proves where it came from. Provenance, meaning every figure traceable to a source and every step reproducible, is what lets an auditor, an LP, or a regulator accept a number an AI produced. A correct answer without that trail is close to worthless, because in this business you cannot rely on a figure you cannot defend.
Key takeaways
- Every number in fund operations already has to be defensible, long before AI enters the picture.
- "The system generated it" is not an answer that survives an audit.
- Provenance moves trust from the person who remembers to the trail anyone can check.
- Speed is what gets marketed. Provenance is what closes the contract.
Why isn't the answer the product?
Because in fund operations the answer was never the bottleneck. A competent analyst can produce the answer. The hard part is proving it, every quarter, to people whose job is to doubt it, and when an auditor asks where a figure came from, what you need is a derivation rather than a result. So an AI that produces brilliant answers and cannot show its work has solved the wrong problem. It automated the easy half and left the hard half, the defense of the number, exactly where it always was.
What does provenance mean in practice?
In practice it comes down to three things working together. Every output figure links back to the source document and the page it came from. Every step in a calculation is open to inspection, so you can see how the number was built rather than take it on faith. And the result is reproducible, so the same inputs produce the same output next quarter and the quarter after that. Underneath, this is a matter of architecture: when the records are event-sourced and append-only, every figure carries its own time-stamped path back to the source receipt it came from, and that is what turns an AI output from a claim into evidence.
How does an audit trail change the reviewer's job?
This is the part that gets under-appreciated, and it is where the real return lives. Without a trail, a reviewer has to re-derive the AI's answer from scratch in order to trust it, which means the AI saved no one any time and may even have added work. With a full trail, the reviewer checks the citations instead, the same way they would check a junior analyst's, following each figure back to its source rather than rebuilding it. That is a fraction of the effort, and it is the whole difference between AI that adds review burden and AI that removes it. For the head of fund services who carries the audit, this is the number that matters: on real fund-ops work, that shift took a NAV-package review from about nine hours down to roughly thirty minutes, with the reviewer on the exceptions and every figure traced to source.
Why does this matter more than model quality?
Model quality is converging, and everyone now has access to strong models, so it is a smaller differentiator every month. What separates AI that ships into a regulated back office from AI that stalls in a pilot is not the model. It is whether the output is verifiable. The engineering that matters now is the trail and the reproducibility and the sourcing, not another point of benchmark accuracy that no auditor will ever ask about. This is also where explainable AI gets misread: knowing why a model leaned one way is useful, but it is not the same as proving where a specific figure came from and reproducing it. Put plainly, verifiable AI is what a regulated institution can adopt, and an unverifiable one is what it has to decline no matter how clever it looks.
So how should an institution evaluate financial AI?
If you own compliance or risk, this reframes the whole vendor conversation. Ask the provenance questions first, before the accuracy ones. Can I trace every figure to its source? Can I reproduce this result next quarter? Can I show an auditor the derivation? If the answer to those is no, the accuracy number is beside the point, because an answer you cannot defend is an answer you cannot use. Speed and accuracy are what vendors lead with. Provenance is what you should actually buy on.
This is what DwellFi is built around. Every figure it produces traces back to its source document and page, the records are event-sourced so the derivation exists before anyone asks for it, and the same inputs reproduce the same result, all inside your own environment. It hands you the answer and the receipt in the same motion, rather than asking you to trust a number it cannot source. Get this right and AI stops being something compliance fears and becomes something operations relies on, which is the moment it finally starts paying for itself.
Frequently asked questions
Have more questions about explainable AI? We got them answered:
Is explainable AI the same as an audit trail?
They are related but not identical. Explainable AI describes why a model decided something, while an audit trail proves where a specific figure came from and lets you reproduce it. Fund accounting needs the second, and ideally both.
What is provenance in AI?
It is the ability to trace every output back to its inputs and its reasoning, and to reproduce the result on demand. In finance, provenance is what makes an answer defensible rather than merely plausible.
Why do AI pilots fail compliance review?
Usually because the output is impressive but unverifiable. Compliance cannot approve a number it cannot trace or reproduce, however good the number looks on the page.
Download The Provenance Checklist: the questions to ask any financial AI vendor before you buy. It turns the test in this piece into a one-page evaluation you can take straight into your next vendor call.
Already convinced? See the trail behind every figure,