← All articles

Capitalize Tokens as Symbolic Logic to Deliver Auditable Legal Advisory Systems

John Lord, Chairman & CEO - Neota Logic
·
November 17, 2026

Much of the current conversation about legal AI rests on one premise, which is that a large language model should be asked to interpret the legal problem in front of it, every time that problem comes up.

Harvey, Legora and Claude for Legal have shown how good these models have become at drafting, research and synthesizing complex material, and legal teams are right to use them for that work.

The premise holds up less well for questions that carry an organization’s own judgment. A firm’s playbook positions, a legal department’s policies and the way a team reads a regulation it deals with every week are institutional intellectual property, built up over years. Asking a model to interpret that knowledge on each request means paying repeatedly for an LLM to “re-guess” the same policy or regulatory question with varying drift.

Interpretation and encoded expertise behave differently

AI interpreting the law ad hoc and a system executing encoded expertise produce different kinds of answers. The first offers a probabilistic suggestion. The second is deterministic, in that the same inputs lead to the same result, and the result can be traced back to the reasoning that produced it.

Symbolic AI is the older of the two approaches. Lawyers express what they know as explicit rules, conditions, relationships and exceptions, and the system applies them the same way on every run. For a regulated organization that consistency matters as much as the answer itself, because an outcome nobody can reproduce is difficult to defend.

A gifts and hospitality policy shows the difference well enough. A model asked whether a particular invitation complies will read the policy and give a reasonable view. Asked again next month, or after an upgrade to a newer model, it may give a different one. Once the policy is encoded, the same facts lead to the same outcome, and the rule behind it is on record.

Why encoding expertise has been expensive

Legal knowledge engineering has been practiced for decades, and few people who have worked with it doubt its value. What has held it back is the effort involved. Turning a lawyer’s expertise into a formal model meant working rule by rule through regulations, policies and contracts, and for many use cases the cost was hard to justify, so the expertise stayed with the people who held it.

Large language models change the economics of that work. The same ability that lets them read and summarize long documents makes them useful for the early, laborious parts of authoring, such as identifying the facts that matter and how they relate to one another.

Where AI belongs in the authoring process

Used this way, generative AI speeds up parts of the expert’s work without taking it over, and authority over the rules stays exactly where it was.

  • Structure. A model works through the source material, whether regulations, policies, playbooks or contracts, and surfaces the relevant facts along with suggested variables and relationships for the author to consider.
  • Shape and test. Expert lawyers decide what the rules are. They refine the logic and stress-test it against scenarios, some of them generated with AI, before formally signing it off.
  • Execute. The approved logic, workflows and document automation then run consistently, with a record of how each outcome was reached.

The model contributes speed at the stages where speed helps. It has no say in which rules apply or what any outcome should be.

A worked example

Take an invitation to a vendor’s box at a championship game, valued at $1,200 and sent to a business development lead while that vendor is bidding in an open procurement process.

A model first reads the hospitality policy and the related procurement guidance and identifies the factors that bear on the decision, including the value, the relationship to the host, whether a tender is live and the thresholds that trigger approval.

The compliance lawyer then compares those suggestions with how the policy is actually applied. In practice, any hospitality from a party in an active tender goes to the General Counsel regardless of value, a condition the written policy never quite states. The lawyer adds it, tests the logic against a range of scenarios and signs it off.

When the request comes in, it is assessed against the approved rules and, because a tender is live, referred to the General Counsel with the triggering rule attached. A colleague who submits the same facts next quarter receives the same referral.

Each time a team puts the same policy or regulatory question to a model, it pays for another interpretation, along with whatever drift comes with it. A hundred questions about the hospitality policy cost a hundred readings of it.

Encoding the policy moves that spend somewhere else. Tokens are spent on authoring, once, and what they produce stays in place. Institutional expertise is codified into an asset your organization owns, audits, and maintains, and the hundredth question is answered the way the first one was. This is the sense in which token spend can be capitalized rather than consumed.

Deciding what to encode

Encoding is not the right answer for every question. Drafting, first-pass research and summarizing long documents are usually better served by a capable model and a careful reader. The case for encoding gets stronger when a question has three characteristics.

  • It recurs, so each fresh interpretation is a repeated cost.
  • An inconsistent answer would create real exposure, whether regulatory, contractual or reputational.
  • The outcome may later need to be explained to a regulator, a client, an auditor or a court.

For most legal teams, the strongest candidates sit in policy interpretation, regulatory triage, intake and approvals.

Conversational assistants such as Harvey, Legora, Claude and Copilot will remain central to how lawyers handle intake and research. Alongside them, legal teams need a reasoning and execution layer for the answers that have to be consistent and defensible.

Making individual lawyers faster at research and extraction is worthwhile. Capturing an organization’s collective legal expertise and applying it at scale is the more significant opportunity, and it depends on probabilistic intelligence operating within deterministic, governed workflows.

That is the architecture Neota Logic is building, a legal logic engine in which AI accelerates how expertise is captured and the rules approved by experts determine every outcome.

Ready to make your AI workflows defensible?

Book a demo and we'll walk one of your real processes through Neota.

Book demo