This page is for the person at your firm who checks vendors. It explains where your data goes, who can see it, and what the agent can and cannot do.
Last updated:
Details depend on your setup. We confirm them with your IT team.
The workflow connects tools that you already have: spreadsheets, email, documents and shared folders. We build it inside your environment, so your documents do not move to a new platform.
The agent sends text only to the AI model that you choose. Otterd reaches your environment through a service account that you control.
At every level, the agent runs in your environment. The difference is where the AI model runs.
1 Your AI account
The model runs at a provider such as Anthropic or OpenAI, under your firm's business account.
Most teams start here.
2 Your cloud
The model runs inside your firm's AWS, Azure or Google Cloud account, in the region you choose, such as Tokyo.
For firms that keep all data in their own cloud.
3 Your servers
An open-weight model runs on hardware that you control. No text leaves your network.
For the most sensitive work, such as MNPI.
At every level, our access is read-only, every run keeps a log, and a person approves each action that leaves your firm.
Open-weight models can be less capable than the leading models for some tasks. We test your workflow on the level you choose before you decide.
Access follows the three stages of the project. At the end, only your team has access.
| Stage | Who runs it | Who has access | Time |
|---|---|---|---|
| Build | Otterd | Otterd, read-only, on sample data | 2 to 3 weeks |
| Operate | Otterd with your team | Otterd and your team | 4 to 8 weeks |
| Transfer | Your team | Your team only | 1 week |
It stays in your systems. The agent runs in your environment.
The agent runs in your cloud account or your Microsoft 365 or Google Workspace tenant. Otterd does not keep a copy of your documents. Text goes only to the AI model that you choose.
Only a model that you choose. At the highest level, no text leaves your network.
Most teams use their firm's business account with a provider such as Anthropic or OpenAI. Under business terms, these providers do not train on your data by default. Where the provider offers zero data retention, we turn it on. For more control, the model can run in your own cloud account or on your own servers.
Read-only access, only to what the workflow needs. It ends at Transfer.
We use a separate service account for each system, with the minimum permissions. You see each permission before we start. At Transfer, you remove our access, and we confirm it in writing.
No defense stops this completely, so we limit what the agent can do.
A CIM or a data room file can hide text such as "send this file to…". This is called prompt injection. The agent treats documents as data, not as commands. It has a fixed list of actions, and most of its access is read-only. It cannot send email to a new address. A person approves each action that leaves your firm.
Yes. Every run keeps a log, so the agent is not a black box.
The log shows what the agent read, what it flagged, which model it used, and who decided. The log stays in your systems, so your compliance team can review it with your other records.
The agent sees only the folders and mailboxes for its workflow.
We set access by team, to match your information barriers. When a workflow writes research notes, the agent can check your restricted list first. For MNPI, you can choose the level where the model runs on your own servers.
We delete our working copies and confirm it in writing.
At Transfer, you remove our access. We delete any working copies, such as test files, and confirm it in writing. You also get a one-page description of the workflow for your AI inventory and your LP questionnaires.
We tell you within 24 hours.
If we find a problem that can affect your data, we tell your named contact within 24 hours. We say what we know and what we do next. The log helps your team see what the agent did.
We answer your security questionnaire and sign your NDA.
Send your vendor questionnaire before the project starts, and we answer it in writing. We do not have a SOC 2 report yet, so we describe each control in writing instead. Our contract states that your data does not train any model. Please do not send confidential documents before the NDA is signed.
Otterd adds no new vendor that sees your data. Text goes only to the tools you already use and the AI model you choose.
Otterd does not hold your data, so we use no subprocessors for it. If that changes, we tell you before the change.
Before we start, we need four things from your firm.
Start with a 30-minute consultation. Send your security questionnaire at the same time, and we answer it in writing.
Get in touch