Token Governance: The Quiet Tax
Legal judgment has always been metered at consumption. Tokens are the first time mismanagement bites on the supply side. Here's what governance actually requires.
The tokenmaxxing epoch has predictably given way to token anxiety, and the legal tech world is feeling the whiplash. The pricing model for legal AI has shifted three times in eighteen months. Harvey and Legora built a category on point-solution pricing — purpose-built legal tools commanding a premium, justified by practice-area depth. Then frontier models arrived with general-purpose capability at subscription rates that undercut the category. Now, vendors like Legora and Ivo are moving to consumption-based pricing, with more likely to follow. Some teams are still evaluating their first tool. Others are asking whether the tool they chose six months ago is still the right one — at a price they can predict.
This problem is not as new as people think. Legal teams were already running on a consumption-based model for the thing that mattered most — their own attention. As any lawyer who has spent the first two hours of their morning sorting email and Slack knows: the color-coding rituals, the draft replies left open in six tabs, the starred messages that will never be answered today. The NDA that somehow took thirty minutes. The major product review that got fifteen minutes between calls. The quiet tax exacted by constant distractions and context switching. The billable hour. Legal judgment has always been metered at consumption. Legal teams have been mismanaging it for decades — they just never had to build explicit architecture for it, because the cost was invisible.
Even if the problem isn’t new, the fear is justified. What if consumption costs spike? What if the model I built on gets expensive? What if I’m at the mercy of a vendor’s pricing decision and there’s nothing I can do about it?
The fear is right. The framing isn’t. “Will prices go up?” is anxiety with a budget line. The actual question — the one most legal teams haven’t asked yet — is: what is the shape of my AI consumption, and what levers do I control?
Fear Has No Architecture
Most legal teams have identified the problem. Almost none have parameterized it.
Parameterization means this: you know what AI workloads you’re running, you know what model tier each one hits, and you’ve made deliberate choices about which tasks warrant which tier. Without that map, you’re running on defaults — usually the model the vendor recommended, or the one the most vocal power user prefers — and hoping the bill stays manageable. That is far from a strategy.
Ask a legal team which model tier each of their AI skills is currently running on. In most rooms, the answer is a long pause — the look of someone being asked about infrastructure they assumed someone else owned. Nobody did the mapping. Nobody was asked to.
Both Anthropic and OpenAI publish model selection guidance. They’re not on most legal teams’ radar. Simplified for legal work, the framework is consistent across both: high-volume, low-judgment tasks — clause extraction, contract classification, NDA first pass — belong on the small, fast model. Standard legal work — contract review, research summaries, policy drafting — goes on the mid-tier. Complex multi-step reasoning, novel legal questions, multi-jurisdictional analysis — justifies the premium tier. Long autonomous agentic workflows requiring a lot of discretion belong at the top. Your mileage varies depending on your risk appetite, your work shapes, and work volume across practice areas.
The teams routing every task to the premium model because it “feels safer” are paying two to three times the necessary rate on the majority of their workload. The teams routing everything to the small model to save cost are producing output that requires human correction — which costs more than the token savings. Neither team has done the mapping. They’ve made a default decision and called it governance.
Parameterization is not complicated. It is, however, work — and it starts before the first deployment.
Know Your Exoskeleton
Which brings me to the prerequisite.
“Inside the Exoskeleton” described the architecture: the layered AI system built around the lawyer — handling deflection, context, judgment calls, and learning — so that attorney attention stays on the work that actually requires it. The exoskeleton is the infrastructure of a well-run legal function in the AI era. It handles what can be handled so the human can handle what can’t.
Like any infrastructure, it has a cost to run. And like most infrastructure costs, it tends to go unexamined until a bill arrives that someone can’t explain.
The prerequisite to consumption governance is an inventory of what you’re actually running. Most legal teams building exoskeletons cannot answer a basic question: which model tier is each skill hitting right now? The skill was built. It worked. It was deployed. Nobody went back and asked what it’s spending.
The second prerequisite to consumption governance is standardization. Most legal teams are reinventing the wheel with individual users creating their own skill. Ignore the other governance issues of inconsistency and entropy over time, and you still have the token waste coming from the same skill being created five different ways.
This is the inventory problem — and it sits underneath every governance conversation about consumption pricing.
Before you can map your token use by task type, you have to know your exoskeleton. Not conceptually. Operationally. Which skills are running? Against which model? At what volume per week? And critically: is model selection hard-coded into the skill, silent or user-selectable? Because user-selectable means uncontrolled. Every time a lawyer routes a high-volume extraction task to the premium model because they’re uncertain which to use, that’s a quiet governance failure — unreported, accumulating, and invisible until the invoice.
The exoskeleton was built to protect lawyer judgment. Token governance is how you ensure it doesn’t quietly undermine its own purpose.
Model Selection Is Not a Runtime Decision
Most legal teams treat model selection as a prompting preference. Someone opens a tool, starts a task, and either accepts the default model or makes an ad hoc choice in the moment.
Model selection is a design-time decision. It belongs in the architecture of each skill — not in the hands of each individual user at each individual run. And it doesn’t get made once. It gets revisited at every phase of a skill’s lifecycle.
At design, the questions are: what tier does this task genuinely require? What’s the acceptable latency? Haiku is fast, Opus is not real-time, and that difference matters for workflows that run synchronously. Does this skill chain to others — and if so, does the team understand that cost compounds at every step in the chain? Is prompt caching built in — cached prompts cost a fraction of uncached, and most teams leave that lever entirely untouched? Is there a hard cap on output length, or is the model free to be as verbose as it chooses? Does the data flowing through this skill have confidentiality requirements that constrain which tier can be used?
But the design question most teams miss entirely is whether the human stays in the loop synchronously — or only at approved escalations. At the Claude for Auckland meetup last week, I showed four hundred people a two-stage NDA workflow I used to reduce review time by 99% at Omnidian. Version 2 was an expert skill: consistent, trusted, embedded with legal judgment, cutting NDA review time from twenty minutes to two. But the needle didn’t move for turnaround time, because every NDA still ran through me first.
The skill thought exactly like me — and still made me the bottleneck.
Version 3 added governance scaffolding at design time, such as AI identity, hard stops, a signing threshold, and an escalation path. That’s what allowed legal to be entirely out of the loop except for defined escalations, with sales running it in two minutes instead of waiting for legal to review. The design question isn’t whether the AI can do the work. It’s whether you’ve built a system that can run without you for the ninety-five percent that doesn’t need you. This has token ramifications because keeping the human in the loop for unnecessary tasks invariably costs more tokens for the human to context-load, ask questions and execute judgment for something the AI could do. Governance scaffolding is not just about removing yourself from the loop. It is what tells you which tier belongs in the loop.
At piloting, one question most teams skip: are you testing on the same model tier you’ll deploy on? Otherwise, you haven’t tested what you’re shipping. The Omnidian NDA workflow ran two weeks in pilot before going live — every output checked against the normal process, escalation thresholds calibrated against real edge cases. That’s a pilot. Testing capability on the model you won’t deploy is a rehearsal.
At deployment, model selection should be hard-coded or have the skill reference a model selection policy. Override permissions should exist, and they should be logged. Budget caps by team, matter type, or user are not punitive — they’re governance signals.
At monitoring, the first question is who owns this specifically. In most legal teams the answer is either nobody, or legal ops. Ideally every team member building with AI should know the basics of model economics and which model to choose when. Token ROI — the ratio of output quality to cost per task — isn’t tracked. Cost is tracked, sometimes. Quality against cost, almost never. And drift — where a skill built for one narrow use case quietly gets adopted for a broader set of tasks it wasn’t designed for — goes undetected until the bill arrives.
At training, does your team know which model for which task, and how to prompt differently by tier? A prompt calibrated for Opus will underperform on Haiku. Most teams train on the premium model and wonder why outputs degrade when cost pressure forces a shift.
At iteration and deprecation: when a skill is upgraded, does anyone re-evaluate model selection? When a vendor deprecates a tier, who owns the migration? The answers are typically: no, and nobody. The fortunate part is that a human need not sit through endless reviews for every skill: skill maintenance skills can run on a regular basis and escalate only the edge cases for human review.
Both Curves Are Consumption-Based
Here is what most consumption-pricing conversations miss: AI tokens are not the only finite resource your legal team is spending. Legal teams were already running on a consumption-based model for the thing that mattered most — their own attention. The cost was always there, in the misdirected efforts and context switches and sustained focus that never quite assembled. They just never had to account for it explicitly, because there was no invoice.
Token governance is the first time legal teams have been forced to build explicit cost architecture for intelligence. That’s not a burden. It’s an opening.
The teams that build it — who know their exoskeleton, code the model selection in, instrument the lifecycle, and design for batching — are not just managing AI spend. They’re getting a second shot at the attention problem they never solved the first time.
The architecture isn’t complicated. But it has to be built before you sign the contract — not reverse-engineered from the invoice.
Natalie Kim
Natalie is the founder of Inflection Advisory and works with organizations on AI acceleration and governance. As VP Legal at Omnidian she led a full-arc enterprise AI adoption.
The Judgment Layer publishes on AI governance, board accountability, legal intelligence, and the judgment no algorithm is taking from you.