Most investigation teams we speak to run a matter across six or seven tools. One processes the data, one hosts the review, one cleans the ledgers, a notebook runs the tests, a dashboard shows the client and email carries the rest. Each tool is good at its step. None of them can carry a matter from the first mailbox to the final report.
An investigation does not run on isolated steps. It depends on evidence, transactions, custodians, chronology, privilege, deadlines and the people who have to sign the findings, all moving together. So at Donna we did not set out to add AI to one of those tools. We built an agentic operating system for investigations and eDiscovery.
We chose the term deliberately. 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 works as one system. Donna plays the same role between frontier AI models and a matter. It gives the agents the evidence, the tools and the method, and it governs what they can see, what they can do and when they must stop for a person.
In practice, that means the model is one part of the system rather than the whole of it. This article sets out the seven layers of that system and what is live in Donna today.
The platform, layer by layer
The evidence, made usable. Mailboxes, chats, file shares, phone extractions, disk images, ledgers and bank statements come in. Each file is fingerprinted with SHA-256, scanned for malware, de-duplicated, threaded and made searchable.
1. Processing
Everything starts with the evidence. Donna takes in mailboxes, chat exports from Slack, Teams and WhatsApp, file shares, Cellebrite phone extractions, disk images, scanned records, ledgers and bank statements. It handles anything from one document to a whole company’s data.
Each file is given a SHA-256 fingerprint the moment it arrives, so you can show it has not changed since. It is scanned for malware before anyone opens it. Then Donna extracts the text and metadata, removes duplicates, threads conversations and makes every page searchable. From that point the evidence is working material, with its chain of custody recorded from intake.
Processing, de-duplication, threading, legal hold and chain of custody are live today.
2. Text and Numbers, Together
Most investigations split in two. The eDiscovery team reads the emails. The forensic accountants work the ledgers. The two halves meet in a meeting, if they meet at all.
Donna reads both. It turns scanned statements into tables, cleans and joins records from different systems, and puts each payment beside the messages written around it. A supplier invoice sits next to the email that approved it. A transfer sits next to the chat that arranged it. That is where most findings live.
Table extraction, data cleaning, Sheets, the evidence graph and dashboards are live today.
3. Agents
This is the layer that does the analysis. You brief Donna as you would a senior analyst: trace the funds out of this account, find who knew about the side letter, test these payments for duplicates and round amounts. Donna’s agents plan the work, split it across a team when the data is large, run the tests, check their own output and try again when it does not hold up.
One rule is built into the system: the loop always ends at a person. Anything that changes the matter in a way others can see stops for approval. Donna shows what it found, how it found it and the evidence behind each step. The investigator approves, changes or rejects it.
Agents, forensic tests, fund tracing, chronologies and Workflow Agents are live today. Every finding carries a citation that opens the exact page, row or message behind it.
4. Review
Review is still the slowest part of most matters. Donna’s agents take the first pass across the whole population, not a sample. They code for relevance and issues, flag privilege with a reason you can read and run QC across the calls.
Reviewers spend their time on the documents that need judgement. Donna measures recall as the review runs, with statistics you can put in front of a court or a regulator, so the decision to stop is a number rather than a feeling.
Agent review, privilege, QC and review statistics are live today.
5. Redaction and Production
Whatever leaves the matter has to be right. Donna proposes redactions for personal information and privilege, a person checks them and signs off the set, and an audit copy is kept. Productions go out with Bates numbers and load files, to the protocol the other side or the regulator expects.
After a breach, the same layer builds the notification list. Agents read the stolen files, find each person and the data exposed, de-duplicate them into one entry each and track the regulator’s clock. Every entry links back to the page it came from.
Redaction, productions, breach notification lists and QC are live today.
6. Working with Clients
All of that work has to live somewhere the team and the client can use it. In Donna, that place is the Space: one secure workspace per matter, where your team, your client and Donna work on the same evidence.
The Space holds the evidence, the review, the chronology, the dashboards and the draft report. The client sees live findings and comments in the same place. Nobody passes exports around, and nobody works from last week’s version.
Each person sees only what they are permitted to see. A client can read the findings dashboard without seeing the team’s working notes.
7. Security and Governance
We draw security and governance at the top of the stack, but it wraps 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 that data never reaches the model.
Roles are set per matter. Every action by a person or an agent is logged, and every download is recorded with its hash. Data is stored and processed in Australia, including AI processing and backups. Donna never trains on your matters.
Governance also controls what agents may do. Reading a document, proposing a redaction and releasing a production are different actions and do not carry the same permission. This layer decides where the investigator’s approval is required.
We built this layer first, even though it is drawn last, because every other layer depends on it.
Why the Whole Stack Matters
With all seven layers in place, a matter can move from the first mailbox to the cited report in one place. Agents carry the volume. Your team makes the calls. The record of who did what, and why, builds itself as the work happens.
For an investigation team, that is the difference between adding an AI feature and changing how a matter runs. The time comes back in the handoffs that no longer happen.
Where This Leaves Us
Each layer is there because an investigation cannot be trusted to a model alone. The findings have to be provable, the review defensible and the evidence untouched. We built Donna so all three hold at once, and we keep deepening every layer with the teams who use it.