Donna’s agents do the analysis in an investigation. By agent we mean a system that takes a whole piece of work, plans it, carries it out on the matter’s real evidence, checks its own output and hands the result to the investigator who signs it.
The word agent is used widely right now, and often loosely. So before we describe what Donna’s agents do, we want to be precise about what we mean. The difference between a model, a model system and an agent decides how much of an investigation can be handed over, so it is worth setting out.
How We Define AI Systems at Donna
The building block is a model: a frontier large language model. We use a combination of models from Anthropic and OpenAI, chosen per task. Given a single prompt, a model produces a single response. It is highly capable, but it is stateless: it knows nothing about your matter, keeps no memory between requests and cannot take action in any system.
- QueryA document and a question.
- ModelOne prompt in, one response out.
- Output
Connect a model to task-specific tools and knowledge, and you get a model system. When you ask Donna who approved a payment, a model system retrieves the right emails and ledger rows, reads them and answers with citations. A model system runs a fixed pipeline of model calls and handoffs, the same way every time.
- QueryA document and a question.
- Model + toolsA general purpose model with retrieval, search and matter data, in a fixed pipeline.
- Output
An agent works differently. You give it an objective rather than an instruction: investigate the supplier payments, review this mailbox for the regulator’s notice, build the breach notification list or trace the funds across these accounts. The agent decides what steps that takes, performs them with tools, inspects its own output and loops back when the work does not hold up. In Donna, one more thing is part of the definition itself: the loop ends at a person.
- ObjectiveThe whole job.
- PlanThe agent decides what steps the job takes.
- ActWith tools on the Space.
- VerifyCite or redo. Work that does not hold up goes back to Plan.
- Your approvalThe loop exits only here.
- Then the work is done.
Why Investigations Are Hard for Agents
General-purpose agents can perform well in a demonstration and still fail on a real matter. Investigations are hard for them for three reasons. First, the work is grounded: a finding that is not anchored in this matter’s evidence is a liability. Second, the work is accountable: someone signs the report, and it may be tested in court or by a regulator. Third, the work is large: the answer sits somewhere in millions of emails, messages and transactions, and missing it is the failure that matters.
We built Donna Agents around those three constraints.
Grounded in the Evidence
Every agent run happens inside a Space, the one place a matter lives. That gives the agent the actual evidence: the mailboxes, the chats, the ledgers, the bank statements, the chronology so far and what the team has already found. Agents work only from the parts of the matter the person asking is allowed to see. Permissions are enforced beneath the model, so an agent cannot read anything the person it works for cannot.
Verified Against the Source
The verify step in the loop is a real check. Every claim an agent makes must resolve to a citation, and a citation does more than name a file. It opens the exact page, ledger row or message it came from, so an investigator can check the work with one click. When output fails verification, the agent does not soften the wording. It sends the work back around the loop.
Approval Before Anything Changes
Donna acts for the investigator, and the investigator decides. Anything that changes the matter in a way others can see stops at an approval gate first. That includes releasing a production, sending a finding to the client and signing off a redaction set. The agent shows what it did, why and the evidence. You approve, adjust or reject it, and every run leaves a full audit trail of each step, each tool call and each source consulted.
What Agents Do in Donna Today
- Supplier-payments fraud. Agents test every payment for duplicates, round amounts, new vendors and split invoices, then link each flagged payment to the emails and approvals around it.
- A mailbox review for a regulator’s notice. Agents map documents to the notice categories, code the whole population, flag privilege with reasons and prepare the production with Bates numbers.
- A breach notification list. Agents read the stolen files, find every person and the data exposed, de-duplicate them into one entry each and link every entry to its source page.
- Tracing funds across accounts. Agents turn bank statements into tables, follow the money from account to account and put each transfer beside the messages that arranged it.
- Work of any size. A large job is shared across a team of agents working at once, then brought back together for your review. Workflow Agents run your team’s method the same way on every matter.
Where This Leaves Us
Agents carry the volume, and your team carries the judgement. As the work gets larger, the approval gates stay where they are, and the investigator still reviews every finding before it leaves the matter.
To try one, open a Space and ask for an outcome instead of an answer.
