Industry

The Cost No One Puts in the Budget

AI budget is read as subscription lines. What that reading misses often costs far more.

When a legal department or an investment committee examines its AI budget, it looks at subscription lines. They are visible, quantified, comparable, negotiated, renewed. That is the apparent cost of legal AI, and it captures most of the budgetary attention. This attention is misplaced, not because these lines are false, but because they divert the gaze from the only question that truly decides an AI investment’s value: not what it costs, but what it allows or prevents doing.

One must distinguish two ways of spending badly. The first, the most discussed, is paying too much for what one buys. The second, far costlier and almost never measured, is misdirecting what one buys: putting one’s budget where it produces no durable value, and withdrawing it from where it would. The budget read as subscription lines protects against the first error and blinds one entirely to the second. One negotiates the price of each tool fiercely, and never asks whether the money is placed in the right spot.

The real budget question is not what AI costs. It is what the misplaced budget prevents you from doing.

The account no budget makes

What the subscription-line budget never surfaces is the cost of what one does not do because one spent elsewhere. Each euro committed to yet another generation tool is a euro not committed to what the organization really lacks: the layer that would hold its tools together, the memory that would retain its decisions, the governance that would make its AI deployable at scale. That cost appears on no invoice, because it is the cost of an invoice one did not issue, for a thing one did not buy.

This type of cost has a name in economics: opportunity cost. It is the value of the best option one gave up by choosing the one taken. In legal AI, it is massive and systematically ignored, because it is by nature invisible: one sees what one bought, never what one could have bought instead. An organization that spent its whole AI budget on generation tools has, in the same move, decided not to build the layer that would have made those tools coherent, without ever formulating that decision, nor measuring its price.

One measures the scale of this invisible cost by a simple sign: the promised gains do not materialize. An organization that invested heavily in tools and finds, two years later, that its productivity has not moved to the level of the promises has not made a bad purchase in the strict sense; each tool, taken alone, keeps its promise. It made a bad allocation: it put its money into capabilities it already had in abundance, and not into the coherence it lacked. The budget was sufficient; its orientation was wrong.

You see what you bought. You never see what you could have built with the same budget.

This opportunity cost has a feature that makes it graver still: it compounds. The organization that, year after year, reinvests in abundant capabilities rather than in the missing layer does not merely make a one-off error; it accumulates a growing lag. While its tools commoditize, the layer it did not build keeps being missing, and the gap between what it has and what would have been useful widens with each budget cycle. One year’s opportunity cost becomes the base for a larger opportunity cost the next year, until catching up becomes a project in itself.

The cost of the complexity you sustain

Alongside what one does not build, there is what one sustains unwittingly. Each added tool is not only a spending line; it is a commitment to maintain, to govern, to make coexist with the others. This upkeep has a cost, which is not in the subscription but which accumulates with each additional tool, and which weighs all the more as the stack grows. An organization that adds tools with no layer to hold them does not merely spend more; it commits to maintaining a growing complexity, whose upkeep cost ends up exceeding the value of the tools themselves.

This upkeep cost has an unfavorable trajectory: it increases over time, while the value of the tools tends to commoditize. At the moment of purchase, the tool brings a net benefit and its upkeep seems negligible. Two years later, the tool has become a commodity, its advantage eroded as the market caught up with its capability, but its upkeep cost has remained, even grown with the complexity of the stack. One finds oneself expensively maintaining tools whose competitive edge has vanished, without ever having decided to.

The bad budget allocation is therefore not only a bad initial placement; it is a commitment to a trajectory of rising costs for declining value. It is the exact opposite of what an investment should produce. And this trajectory is invisible as long as one reads the budget year by year, line by line: each year taken in isolation seems reasonable, it is their sequence that reveals one has committed to a slope where one pays more and more for less and less distinctive value.

Why reading in subscriptions misleads

Reading the budget as subscription lines is not neutral: it steers the decision toward what compares easily, and away from what really matters. Two subscriptions compare: same nature, posted prices, listed features. An infrastructure layer does not compare to a subscription, because it is not of the same nature: it is not bought by the feature, it is judged by what it holds together. By filing everything in the same “subscriptions” column, the budget renders infrastructure invisible, or makes it appear as one more subscription, more expensive and less demonstrable than the others.

This distortion has a direct consequence on the decision. Faced with a table of comparable lines, the reflex is to choose the line with the best feature-price ratio, that is, one more tool. Infrastructure always loses this comparison, because it does not play in the same category: it does not do a task better, it makes the whole coherent, which does not fit in a column of features. The budget, as read, is therefore structurally biased against the only expense that would solve the underlying problem.

One must therefore change the way the budget is read before being able to spend it well. As long as one lines up subscriptions, one optimizes within a category that is not the right one. The moment one asks “where does our budget produce durable value, and where does it destroy it by sustaining a complexity that keeps growing,” one leaves subscription comparison for a real allocation trade-off. It is this shift, from the price of tools to the orientation of the budget, that distinguishes a legal department that endures its equipment from one that steers it.

A budget that lines up subscriptions optimizes the wrong category. The right question is not price, it is allocation.

Reorienting, not merely spending less

The practical consequence is not to spend less, but to spend elsewhere. A lucid organization does not first seek to reduce its subscription lines; it seeks to reorient part of its budget toward what it structurally lacks, the layer that makes its tools coherent, governable, endowed with memory. It is not one more expense added to the others; it is a displacement of the expense, from the abundant component to the rare layer, from the volatile to the durable.

This displacement has a property that mere cost reduction does not: it pays off over time instead of depreciating. One more subscription produces a benefit that plateaus and a cost that repeats; an infrastructure layer produces value that accumulates, because it capitalizes context, memory, governance, instead of letting them scatter. To reorient the budget toward the layer is to stop renting capacity and start building an asset. And an asset, unlike a subscription, belongs to the organization and appreciates over time.

It is this displacement that a layer like MAX makes possible, not by adding to the stack as an extra expense, but by reorienting the investment toward what was missing: the coherence, the memory, the governance of the whole. The right budget trade-off in legal AI is therefore no longer to compare subscription lines, nor even to reduce them. It is to measure what the absence of architecture costs, in uncaptured value, and to decide, knowingly, to reorient the budget toward what will produce durable value rather than toward what produces less and less.

One objection remains: a legal department does not always have the latitude to freely reorient its budget, caught in multi-year commitments and equipment habits. That is true, and it is precisely why budget lucidity must come early. The longer an organization waits, the larger the share of its budget committed to sustaining complexity grows, and the less margin it has left to build what is missing. To recognize early that the real cost is not in the subscriptions, but in what they prevent building, is to give oneself the room to maneuver that time, otherwise, closes off.

The right decision is not to spend less on AI. It is to stop renting capacity and start building an asset.