Every firm we speak to has been shown an AI feature: a tool that summarises a contract, a box that answers a question, an inbox that drafts a reply. Each can save time on a task. None can carry a matter.
A law firm does not run on isolated tasks. It runs on documents, deadlines, correspondence, permissions, precedents, people and judgment, all moving together. That is why Donna is not built as another AI feature. It is built as an agentic operating system for legal work.
The phrase is deliberate. An operating system sits between raw computing power and the work people need to do. It manages files, memory, permissions and applications so the machine functions as one system. Donna plays the same role between frontier AI models and a legal matter. It gives the model the right information, knowledge and tools, while governing what it can see, what it can do and when it must stop for a person.
A model can answer a prompt. An operating system can carry a matter.
This article explains the seven layers that make that possible, what is live in Donna today and what we are still building.
The platform, layer by layer
The engine. Donna is model-agnostic, selecting the right model for each task rather than forcing every job through one. As the models improve, Donna improves with them, without the firm having to change how it works.
Seven layers. One operating system.
Donna connects intelligence to the matter, the firm and the work. The lawyer remains in command.
1. Large language models: the engine, not the system
At the foundation is a frontier large language model. Donna is model-agnostic, using models from Anthropic and OpenAI and selecting the right one for each task. Different models suit different work, and a firm should benefit as they improve without rebuilding how it operates.
A model can read, reason and produce language with remarkable fluency. On its own, however, it knows nothing about your matter. It does not know which contract is current, which precedent the firm prefers or what the client said yesterday. It cannot safely act in the systems around it.
A model by itself is an engine on a bench. Many failures associated with legal AI begin when that engine is treated as the whole car: confident wrong answers, invented citations and work that cannot be checked. The model matters, but it is not the system. It is the most replaceable layer in the stack.
2. The agentic harness: turning intelligence into controlled action
The agentic harness turns a model from something that answers into something that acts. It accepts an objective, breaks the work into steps, selects the right tools, carries out each action, checks the result and tries again when the work does not hold up.
Ask Donna to prepare a disclosure summary, chase outstanding conditions or build a settlement timeline. The harness carries the request through to completed work. But autonomy without control is not useful in a law firm, so one rule is built into the system: the loop always exits at a person.
Anything that changes a matter in a way others can see stops at an approval gate. Donna shows what it has prepared, what it plans to do and the evidence behind it. The lawyer can approve, change or reject it. Donna acts. The lawyer decides.
Today, that loop works for tasks completed within one working session. We are extending it across longer-running work, including waiting for responses, resuming later and keeping dates and obligations current. The loop may grow longer. The approval gate does not move.
3. Data and integrations: the matter is the ground truth
A model with no data about your matter can only be plausible. This layer is where the matter’s reality lives: its documents and versions, correspondence, parties, dates and the information flowing through the firm’s systems.
When a document enters a Space, Donna recovers its structure, indexes its contents and connects it to the rest of the matter. It is no longer an isolated upload; it becomes working material the system can use.
A contract is not merely text to summarise. It belongs to a matter, with parties, dates, prior versions, related correspondence, obligations and consequences. The system must understand that surrounding structure before the intelligence above it can be trusted.
Document ingestion and matter indexing are live today. We are now expanding connections to practice-management systems, email and other firm infrastructure. Those connectors are being built carefully because the data they bring in must be reliable, permissioned and useful to every layer above.
4. Context and knowledge: grounded, or it does not ship
Data is raw. Context is data made useful at the moment Donna needs to act. This layer retrieves the right passages, assembles what is known about the matter and binds claims back to their evidence.
A Donna citation does not merely point to a file. It opens the exact highlighted region of the exact page behind the claim, so checking the work takes one click rather than one afternoon. When Donna cannot ground a document-based claim, it should not present that claim as fact.
Above retrieval sits the knowledge a firm accumulates. Within a matter, Donna maintains a Space wiki and structured facts layer, so settled facts can be recorded once and reused. Across the firm, Donna draws from the Library: the precedents, clauses, playbooks and positions the firm already trusts.
Together, these layers let Donna work from what is true about this matter and how this firm practises, rather than from what is merely common across the profession. This makes the intelligence trustworthy rather than merely fluent.
5. Legal capabilities: the work a firm recognises
The layers beneath this one are general. This is where the system begins to look like legal work.
Donna can draft from the firm’s precedents, compare documents and surface what changed, build timelines of key dates and carry out multi-step work through Concierge and Workflow Agents. Concierge starts with an outcome in plain language, then works out the steps and prepares the work for approval. Workflow Agents let the firm encode recurring work and run it consistently across matters.
These capabilities are designed to span a matter, not merely optimise one task. A revised contract lands late in the day. Donna compares it with the earlier version and the matter’s precedents, identifies the clauses that changed, prepares the response and puts the work in front of the lawyer. Nothing is sent without approval, but the work can be waiting when the lawyer returns.
A matter is the unit a firm is paid to move. A feature improves one step inside it. An operating system connects the steps and keeps the matter moving.
Drafting, comparison, timelines, Concierge and Workflow Agents are live today and continuing to deepen. Legal research across legislation, cases and other bodies of law is still being extended. We would rather release it slowly than release it without dependable grounding.
6. Products and interfaces: where the firm and client meet
All of that intelligence has to live somewhere lawyers and clients can work. In Donna, that place is the Space: one secure workspace per matter where the firm, the client and Donna work together.
The Space holds the conversation, documents, drafts, timelines, tasks, Workflow Agents and the client’s view of the matter. Everything remains connected instead of being scattered across separate tools and inboxes.
An interface can only make promises the layers beneath it can keep. Citations must resolve to evidence. Drafts must use the right precedent. Proposed actions must stop when approval is required. Each person must see only what they are permitted to see.
The interface is where the layers meet. It should make legal work visible, editable and legible, not theatrical.
7. Security and governance: the layer that wraps the rest
We draw security and governance at the top of the stack, but it behaves more like a shell around all of it. Permissions are enforced beneath the model, at the tool and data layers. An agent cannot read what its user is not authorised to see because the model never receives the keys.
Identity follows the user through every action. Role-based access, ethical walls, matter isolation and audit trails apply throughout the system. Service-to-service access uses managed identity, and data services sit behind private networking. Donna never trains on your matters.
Governance also controls agency. Reading a document, preparing a draft and changing something visible to another person are different actions and should not carry the same permission. This layer determines which tools Donna can use and where the lawyer’s approval is required.
Security that is structural outlives security that is merely procedural. It was built first because every other layer depends on it, even though it is drawn last.
Why the whole stack matters
Stack the seven layers and the product changes in kind, not merely in degree. A model can answer a prompt. An agentic operating system can understand the matter, retrieve the evidence, use the firm’s knowledge, carry out the steps and stop for the lawyer’s approval.
That is the difference between adding an AI feature and becoming an AI-enabled firm. One improves a task. The other changes how the matter moves.
Donna is not finished, and we have been precise about where the system ends today and where we are still extending it. But the architecture is deliberate. Every layer exists because legal work cannot be trusted to a model alone.