The Assets AI Builds That Nobody Records

This summer I automated a finance workflow at a public company. A task that took ninety minutes took about three. I sized the annual saving at roughly $784,000, presented it, and then went looking for where it appeared in the financial statements. It does not.

What actually happened is that an operating expense line stopped growing. There is no asset called “the encoded judgment of the accounting team.” Under US GAAP, purchased intangibles are capitalized and internally developed ones are largely expensed as incurred, with narrow exceptions for certain software development costs. So the tool was expensed while it was built, and the thing it produced is visible only as an absence: headcount never added, hours never spent, a line that stayed flat while revenue did not.

The most valuable thing a company builds with AI shows up in its accounts as something that did not happen.

None of this is new in kind. Brand, process and institutional knowledge have always gone unrecorded. What is new is the rate. A workflow that used to take a decade of tribal knowledge to grind down can now be specified in a quarter, and the gap between a company’s recorded assets and its real productive capacity opens faster than the accounting was designed to track.

Why this is a valuation problem

If you value a company on reported margins, you are measuring the output of these assets without ever seeing the assets. Two companies with identical income statements can hold very different amounts of encoded judgment. The one holding more will show operating leverage your model does not predict, and it will arrive as a pleasant surprise rather than as something you underwrote.

The obvious objection is that “invisible asset” is precisely what every management team claims when it wants credit for something it cannot prove. That objection is correct, and it is why the claim needs a footprint rather than a narrative.

The footprint I would look for is boring and checkable. Is revenue per employee rising across several periods while the relevant expense line stays flat, with no headcount reduction that explains it? If yes, something real is compounding. If the only evidence is a slide with the word AI on it, nothing is.

The part I actually care about

There is a second-order version of this that interests me more. The scarce skill is no longer operating the tool. It is being able to specify a judgment precisely enough that it can be encoded at all.

That is not a technical problem. To automate the invoice review I had to understand why the accountant flags the invoices she flags, which meant understanding the business before I understood the workflow. The model could not tell me that. Nobody could except her.

So this is the bet I am making with my own time. The tools will keep changing and the specification problem will not. I would rather be the person who can describe the judgment than the person who can operate the model, which is also why I would rather learn finance properly than learn another framework.