When discussions turn to enterprise artificial intelligence (AI), most executive teams look straight at software licenses or computing power. Keith Vere Fenner sees a far more immediate financial problem taking shape across corporate balance sheets. Six months ago, Fenner argued that the real ceiling on legal technology was the disorganized state of internal data, but he now admits that perspective missed half the picture. The second constraint is what it costs to reason over that unmanaged information, an expense that is already landing on partner desks without matching any approved budget line.
The Unpredictable Nature Of Reasoning Bills
Evidence of these budget failures is already showing up across public records in multiple sectors. Uber went through its entire 2026 AI coding budget by April, while the FinOps Foundation reported that several enterprises were running three times over their yearly token allocations. At the same time, falling per-token rates have given many managers a false sense of security about their bottom line. Fenner explains that cheaper unit costs do not solve the underlying math: “Prices have come down. Adoption and increasingly autonomous agents have pushed usage up faster than prices have fallen. That is the whole story in one line: cheaper per call, more expensive per answer.”
Part of the difficulty for finance departments is that token consumption does not behave like standard enterprise software. When buying traditional cloud space, a company knows the hourly rate before running any software on the network. With reasoning tools, costs fluctuate based on how many documents the model reads, how many attempts it makes, and how many external tools it calls to verify facts. As Fenner notes, you buy an outcome and find out the price afterwards, which is why major institutions like JPMorganChase, Oracle, and Accenture helped launch the Tokenomics Foundation to figure out how to measure it.
Treating Dark Data As A Delivery Expense
Software development usually involves tidy repositories and concise files, but corporate law presents a far messier picture. In legal practice, documents run hundreds of pages, case files stretch across decades, and very little internal material has ever been cataloged properly. When an autonomous tool searches across an unmapped file share for a single fact, it has to read through duplicate files, superseded contracts, and old team archives just to build an answer. As Fenner explains, “Retrieval and reasoning cost scales with the size of the cohort, not with the length of the answer.”
This dynamic fundamentally changes how leadership must think about legacy digital storage. Uncleaned files used to sit quietly on servers, carrying risk during data breaches and adding a modest line item to IT maintenance. Under an automated reasoning setup, that same content directly drives up the marginal cost of completing billable client work. As Fenner points out, “Dark data used to cost storage and carry breach exposure. Under agentic AI, it also sits in cost of delivery. Firms are paying, per query, to reason over content they should have defensibly disposed of a decade ago.”
Shifting Board Focus From Usage To Value
For the past two years, boardrooms have tracked success by asking how many fee earners are logging into new tools each week. That reliance on adoption metrics hides deep financial friction, according to research from Jellyfish showing that top users burned ten times more tokens to gain only twice the output of lighter users. Moving staff into heavy usage patterns often degrades the return on investment, rather than improving matter profitability. Fenner is blunt about the shortcoming of these reports: “Adoption is an input. Firms keep reporting it as a return. A board told that 90 percent of fee earners used the tool last month has learned nothing about whether the firm made money.”
To fix this blind spot, managing partners must build disciplined controls across their client matters. That process begins with tagging expenses to specific matters, funding defensible data disposal as a commercial lever, and matching simple assignments to lower-cost models, while saving advanced systems for complex legal arguments. Success comes down to measuring the business outcome per dollar spent rather than tracking how many seats are active in a given month. For corporate boards trying to make sense of the incoming costs, Fenner offers a final warning: “The first bill was for the data nobody governed. The second is for the reasoning nobody measured. Neither will arrive labeled as a governance failure, which does not stop it being one.”
Follow Keith Vere Fenner on LinkedIn for more insights on enterprise AI, legal technology, and managing the hidden costs of autonomous agents.