Legal AI in Practice: Four End-to-End Workflows

Theory ends where the matter file opens.

The first seven articles in this Academy built an argument, layer by layer. Legal AI is a system, not a chatbot. Confidentiality and sovereignty are preconditions, not features. Legal work is document work. Context is the case. Vendors must be evaluated on trust, not demos. In-house teams and law firms have distinct realities that any tool has to respect.

This article puts that argument to work. It walks through four real workflows, end to end, and shows where the AI proposes, where the lawyer decides, and where the system records. It also names the anti-patterns, the ways legal AI goes wrong in practice, so they can be avoided deliberately rather than discovered painfully.

This is the bridge between the foundations you now have and the deeper subjects the rest of the series will cover.

The seven principles, briefly

Before the workflows, a one-line recap of what the previous articles established, because the workflows depend on all of it.

Legal AI is a system of model, prompt, context, and guardrails. Confidentiality, sovereignty, and auditability come first. The unit of work is the matter, not the chat. Context has six layers and every answer needs sources. Vendors are chosen on trust and tested with real pilots. In-house teams run on parallel flows under headcount pressure. Firms run on deadlines, leverage, and knowledge that lives in people.

Hold those in mind. The workflows below are simply these principles, applied.

Workflow one: contract review, in-house

A supplier contract arrives by email, in French with an Arabic annex, needing review before signature.

The system ingests the contract into the relevant matter, identifies it as a draft, and recognises both languages. The AI reviews it against the company's own contract playbook, flags the clauses that deviate, the liability cap below the company's threshold, the unusual termination notice, the missing data-protection clause, and drafts a short note explaining each flag with a citation to the exact clause. The lawyer reviews the flags, agrees with most, overrides one based on commercial context the model cannot see, and decides what to negotiate. The system logs the review, the AI's output, the lawyer's decisions, and the version sent back to the supplier.

The AI did the first pass. The lawyer made every decision that mattered. The system kept the record. That division is the whole point.

Workflow two: hearing preparation, law firm

A commercial dispute has a hearing in ten days. An associate must prepare conclusions.

The system holds the matter: the contract, the amendments, the correspondence, the mise en demeure already sent, the opposing party's last filing, the hearing date with the deadline computed. The AI produces a structured first draft of the conclusions, drawing on the firm's precedent for similar disputes, with every factual assertion linked to a document in the file and every procedural step flagged for confirmation. The associate refines the legal argument, adds the strategic framing, and corrects the points where the model was generic. The partner reviews the draft, checks the citations against the sources, applies senior judgment, and signs. The system records the chain from first draft to signed conclusions.

The associate started at seventy percent instead of zero. The partner reviewed instead of rewrote. The court deadline was tracked, not remembered.

Workflow three: internal legal request, in-house

The sales director sends a one-line message. "Client wants to add a penalty clause, can we accept?"

The system turns the request into a structured matter, attaches the contract in question, and links the relevant prior decisions. The AI retrieves how the company handled similar penalty clauses before, summarises the current contract's relevant terms, and drafts a first-pass answer with sources. The lawyer assesses the commercial and legal risk, decides the position, and writes the response that goes back to sales in their name. The system logs the request, the analysis, and the decision, so the next time this question arises, the answer is already in the institutional memory.

A vague corridor question became a traceable, reusable piece of legal work, without consuming half a day.

Workflow four: source-aware legal research

A counsel needs to understand the rules applying to a specific regulatory question.

The system scopes the research to the matter and the user's permissions. The AI retrieves from the firm's own prior work first, then from published primary sources and curated regional references such as LexisNexis Middle East, and produces a synthesis where every claim carries its source. The lawyer verifies the sources, fills the gaps the model flagged as uncertain, and forms the legal judgment. The system preserves the research so it does not have to be redone next quarter.

The research is faster and, crucially, checkable. The lawyer is verifying real sources, not trusting a fluent paragraph.

The role of the human, made explicit

Across all four workflows, the pattern is identical, and it is worth stating plainly.

The AI proposes: first drafts, flags, retrievals, summaries, syntheses. The lawyer decides: the position, the strategy, the override, the final wording, the signature. The system records: who asked what, what the AI returned, what the human decided, and what was sent.

This is not a compromise or a limitation. It is the correct architecture. It is how a profession built on judgment and accountability uses a tool built on pattern and probability, without surrendering either the judgment or the accountability.

The anti-patterns to avoid

Legal AI goes wrong in predictable ways. Naming them is the cheapest insurance available.

Pasting confidential content into a public chatbot. The single most common and most serious mistake, for the reasons set out in Article 2. The convenience is real. So is the breach.

Accepting an output without checking its sources. A fluent answer with no verified source is not legal work. It is a draft of a possible mistake.

Skipping the human review because the output looks good. Looking good is exactly what language models do best, including when they are wrong. The review is not optional politeness. It is the control.

Treating the AI as the author. The AI is the assistant. The lawyer is the author, and carries the responsibility that authorship implies.

Deploying AI on top of disorganised documents. As Article 3 argued, AI on a fragmented file produces unreliable results. The document layer comes first.

Avoid these five, and most of the real-world failures of legal AI are avoided with them.

In Morocco and francophone practice

These workflows are not generic. In Morocco they carry specific texture.

Bilingualism runs through every workflow. Each of the four examples involved French and Arabic in some form, the contract annex, the court filing, the regulatory source. The trust perimeter and the retrieval must handle both, symmetrically, or the workflow breaks at the language boundary.

Traceability is a professional requirement. The audit trail in each workflow is not bureaucratic overhead. It is how a Moroccan legal professional remains accountable, to the client, to the bâtonnat, and where relevant to the CNDP and other authorities discussed in Article 2.

The cabinet and the in-house team connect. Two of the workflows were in-house, two were firm-side, but in Moroccan practice they are linked. In-house teams instruct cabinets, cabinets serve in-house clients, and the same matter often crosses the boundary. A Legal OS that holds the matter cleanly on both sides reduces the friction at exactly the point where Moroccan legal work most often loses time.

Looking ahead: the rest of the series

This article closes the first phase of the Academy, the phase about what legal AI is and how it works in practice. The phase that follows goes wider and deeper.

The coming articles step back to the foundations of law itself, a short history of justice, the origin of law, how legal systems actually work, the major areas of law, and what it means to think like a lawyer. From there the series turns to the lawyer in the age of AI, to prompting and working well with a legal AI assistant, to ethics and responsibility, to the vendor landscape, to practitioner-led benchmarks, and finally to the future of legal work, closing with a glossary tying the vocabulary together.

The thread that connects all of it is the one this article has tried to make concrete. Technology changes the workflow. It does not change the responsibility. The lawyer remains the author. The system serves the work.

Practical example

Bring the four workflows together in a single matter, followed end to end.

A Moroccan distributor faces a dispute with a foreign supplier. The in-house counsel opens the matter, and the contract, amendments, and correspondence are ingested and structured (Workflow three's intake discipline). She runs a source-aware analysis of the termination and penalty questions (Workflow four), forms a view, and decides to instruct external counsel for the litigation. The cabinet receives the matter, already organised, and prepares conclusions for the hearing using a sourced first draft reviewed by the partner (Workflow two). When the supplier later proposes a settlement contract, the in-house team reviews it against the company playbook before signature (Workflow one).

One matter. Four workflows. Two organisations. French and Arabic throughout. At every step the AI proposed, a qualified lawyer decided, and the system recorded. The dispute was handled faster and more cleanly than the fragmented alternative, and at no point did anyone surrender the judgment or the responsibility that the law places on the lawyer.

This example is illustrative. Any specific legal output must be reviewed by qualified Moroccan counsel.

What this changes for Burhan

Everything in this article describes what a Legal Operating System is for.

Burhan exists to hold the matter, structure the documents, handle French and Arabic, run the AI inside the context and the permissions, keep every output sourced, record the full chain of who did what, and connect the in-house team and the cabinet around the same matter. The four workflows are not a wish list. They are a description of the work a Legal OS is meant to carry.

And in all of it, the design principle is constant. The AI proposes, the lawyer decides, the system records. Burhan is built to make the lawyer's judgment faster to exercise and easier to stand behind, never to stand in its place.

The lawyer does not disappear. The lawyer moves higher in the workflow. That sentence has run through this entire first phase of the Academy, and it is the right note on which to move into the rest of the series.

Key points

• Real legal AI works in defined workflows: contract review, hearing preparation, internal requests, and source-aware research.
• In every workflow the pattern is the same: the AI proposes, the lawyer decides, the system records.
• Five anti-patterns cause most failures: public chatbots, unchecked sources, skipped review, treating the AI as author, and AI on disorganised files.
• In Morocco, bilingualism, traceability, and the cabinet-to-in-house connection give these workflows their specific texture.
• This article closes the practical foundations; the rest of the series widens into law itself, the lawyer's craft, ethics, and the future.