Without defensibility, the other three don't matter
Why defensibility is the fourth non-negotiable requirement for legal AI.
Defensibility01
Eras02
Stack03
Context04
Duties05
Receipt06
Part 01The claim
The Fourth Requirement
Without defensibility, the other three don't matter
Every product team balances three things: speed, accuracy, cost. Legal adds a fourth, and it is not a feature you bolt on at the end. A defensible product survives the question “Show me what actually happened and prove it.”
Not survive a marketing presentation. Survive the moment a regulator, an opposing party, or a disciplinary board asks for the record.
The customer is a fiduciary. A fiduciary-native product treats the evidentiary requirement as a core capability, not a compliance overlay.
Move the sliders. The three pillars are table stakes. Remove the fourth requirement and the product stops being adoptable.
Part 02Analysis
The Four Eras of Legal Tech
The duty does not move when the work does
In the first era the lawyer was the integration layer. Everything lived in one person's head, and the record was whatever that person chose to write down. The system was the lawyer.
Contract lifecycle management gave the work filing cabinets. Copilots automated stare-and-compare. Each era moved capability without moving the duty.
Agents pull their own context, make decisions nobody pre-authored, and trigger actions downstream. The lawyer becomes the author of the rules the agent operates inside.
Part 03Building
Four Stages, One Stack
Each stage moves the work upstream
Prompt engineering asks what to say. Context engineering asks what to know. Intent engineering asks what to want. Each is a step upstream from the last.
Intent alone is a promise without an enforceable contract. Spec engineering asks what the output must satisfy, and it is the only stage that produces something a third party can check.
The work moves from instruction to authorship
Legal is the leading edge here because the stakes are highest and the standards are the most formally defined. Finance, healthcare, and regulated product development follow the same path.
Part 04Gathering
The Ring of Context
Context without proof is just a story with no receipt
The inner ring is agreement-adjacent and structured: master agreements, amendments, orders, playbooks, signing limits, indemnity ceilings. The outer ring is conversational and informal: Slack, procurement email, CRM notes, signing authority buried in an HRIS.
For thirty years the only system that could integrate the outer ring with the inner ring was the lawyer's brain. Aggregation finally does that job, and it makes the agent better informed.
The lawyer with rich context can think better. The lawyer cannot yet prove what the system did on their behalf.
Part 05Human stakes
The Fiduciary Rules
Delegation does not transfer duty; automation does not suspend it
ABA Formal Opinion 512 reads Model Rule 1.1 competence to cover the tools a lawyer uses, and Rules 5.1 and 5.3 to cover supervision of nonlawyer assistance. An agent is nonlawyer assistance that moves faster than the supervision designed for it.
Four duties map onto four specifications: loyalty to authority binding, care to consensus verification, disclosure to human review attribution, confidentiality to hash-only enrollment.
The spec is the governance. Same activity, two vocabularies, one product capability.
Part 06Proof
The Defensibility Chain
Logs are not receipts. A log tells you a story. A receipt is evidence.
A log is editable and engineering-oriented. It was built to help a developer debug a system, and it can be rotated, truncated, or rewritten without anyone noticing. A fiduciary receipt is signed and immutable, and it is addressed to someone outside the system that produced it.
Configure the mandate in the workspace and run it. The hash is computed in your browser over the payload you selected, exactly as it stands.
Attest is a working prototype of the agent action receipt.
Coda
The Defensibility Imperative
Fast, accurate, and cheap are table stakes. Defensible determines category membership.
A non-defensible AI product in any regulated profession is not a worse product. It is a product in a different category, the category practitioners pilot enthusiastically and quietly abandon once the liability questions get serious.
The claim
The moat is the discipline. The discipline is the product.