The reference architecture
Most lawyers still use AI as a better search box. This is what happens when AI instead becomes part of the operating system of a practice — a coordinated group of specialist agents, a shared case memory, source-verified provenance, and the attorney in control of every decision.
Start with the concept: what an AI second chair is →What this is: a working reference architecture, in controlled development and testing inside a criminal-defense workflow. The point isn't that every firm should build this exact system — it's to show where legal AI is heading, and how a practice can move beyond isolated prompts toward a governed, source-grounded workflow.
The shift
Ordinary AI use is transactional: you ask a question, it answers, the conversation ends. Nothing persists. Every new question starts from zero, and you carry the whole burden of remembering the case and re-explaining it each time.
An agentic system works differently. It holds a persistent understanding of the matter. New discovery changes that understanding. Facts get verified or disputed. One agent's work informs another without you re-explaining anything. The AI is no longer just answering prompts — it's working inside a governed legal workflow.
Evidence & case information → Case Brain → specialized workflows → AI reasoning → structured proposals → attorney review → updated Case Brain → reusable Practice Knowledge
Specialized work, one shared case
You don't manage a screen full of robots. You see familiar legal workflows. Behind each one is a specialist agent with its own methodology — but every agent works against the same underlying case. There aren't sixteen separate AIs each trying to understand your client. There's one case, one shared body of knowledge, and many specialized forms of reasoning applied to it.
Analyze Discovery
Extract facts, statements, events, witnesses, and evidence from the record.
Verify Facts
Determine what the evidence actually supports, disputes, or fails to establish.
Analyze Evidence
Weigh significance, reliability, and the relationships among sources.
Develop Defense
Identify and build defense theories against the actual record.
Stress-Test the State
Build the strongest prosecution case, so weaknesses surface before court.
Bench Analysis
Simulate a neutral judicial read rather than one-sided advocacy.
Research Law
Analyze the statutes, cases, and rules actually in play.
Cross & Direct
Build witness examinations from the evidentiary record.
Resolution & Sentencing
Analyze plea considerations; organize mitigation and sentencing.
Two kinds of memory
At the center is the Case Brain — the structured memory of the matter: facts, witnesses, statements, events, evidence, contradictions, legal issues, verification results, and the decisions you've made. When an agent runs a task, it receives only the parts relevant to that job — the cross-examination workflow gets the officer's reports, the contradictions, the timeline, the verified facts. Not the whole file dumped into every request. The system selects the context, so workflows build on each other instead of starting from zero.
Then there's Practice Knowledge — a second memory that outlives the case. The Case Brain answers what do we know about this matter? Practice Knowledge answers what has this practice learned before that might help? Recurring fact patterns, prior research, argument frameworks, cross techniques, sentencing strategies. A new matter might surface that a materially similar constructive-possession pattern appeared in three prior cases — or that research on a Fourth Amendment issue exists but was last verified eighteen months ago and needs updating.
The point isn't to copy old cases. It's to give you the benefit of your own accumulated experience without having to remember where every argument, memo, or ruling lived before.
The part that matters most
The system doesn't just tell you what it thinks. It shows you why. Every material factual assertion is linked to the underlying evidence and the most precise locator available — and that link stays attached as information moves through the system. If one agent finds a statement and another later relies on it, the second traces back to the original report or recording, not merely to "the Case Brain."
A statement, cited to the page. Officer Smith Incident Report · p. 4
A recorded statement, cited to the second. Officer Smith BWC · 00:08:41–00:09:12
A proposition, cited to the pincite. State v. Askerooth, 681 N.W.2d 353, 364 (Minn. 2004)
And the citations are validated, not merely generated. The system checks whether the source exists, whether the page is real, whether a timestamp falls within the recording, whether a Bates number belongs to the source — and whether a source from another client's matter is being improperly used. A citation defect stops an assertion from being silently adopted.
One distinction is held carefully, because it's the honest one: a valid citation proves where an assertion came from. It does not prove the assertion is true. The attorney still evaluates the evidence. The AI's job is only to make the source impossible to lose.
What that makes possible
Once reports and recordings share the same provenance system, the system can do something a tired lawyer at midnight often can't: line them up against each other. Squad video, body-camera, and 911 audio are transcribed locally — the recording runs through free, on-device tools, and the original stays the authoritative evidence. Then the analysis can compare the paper to the footage.
Consent was obtained before the vehicle search. Supplemental Report · p. 6
Search-related conduct begins at 00:16:42. Squad Video · 00:16:42
The consent request occurs at 00:17:08 — twenty-six seconds later. Squad Video · 00:17:08
Flagged for review. The system does not conclude the officer lied, fabricated, or committed misconduct — that's a legal and factual judgment for the attorney. Its only job is to make the discrepancy impossible to miss.
A hierarchy of trust
The architecture holds a strict order of trust — the safeguard against a common failure, where an AI-generated summary slowly starts getting treated as if it were the original record.
Original evidence and legal authority — the report, the recording, the opinion.
Attorney-adopted Case Brain information.
Untrusted AI analysis — a proposal, until reviewed.
Practice Knowledge summaries.
If an officer's report says something, the report stays the source. If a Minnesota Supreme Court opinion sets a rule, the opinion stays the authority. AI helps find, organize, and apply those sources. It doesn't replace them. And the system preserves competing versions of reality rather than forcing a premature "truth" — a report contradicted by video can stay disputed, with no motive assigned, until the evidence and the attorney's judgment justify more.
The line that doesn't move
This is not autonomous lawyering.
The architecture is built around attorney decision rights. AI analyzes, verifies, researches, compares, drafts, and prepares. No agent independently sends an email, contacts a prosecutor, communicates with a client, files with a court, or takes any consequential outside action. Material conclusions stay traceable to their sources, and the consequential decisions stay with the lawyer. The agents prepare the case. The lawyer runs it.
Why this matters
The long-term value of agentic AI isn't that it writes faster. It's that a practice can start accumulating a structured intelligence layer around its own work. Every case can make the system more knowledgeable. Every correction can make it more aligned with how you actually practice. Every research project makes the next one faster; every motion informs future motion strategy; every video becomes searchable; every assertion stays traceable to the record.
That's a different question than "how can I use ChatGPT?" The real question is becoming: how should my practice operate when AI can function as a coordinated second chair across the entire life of a matter? That's the system being built here — and the transition every firm will eventually have to think through.
Let's talk