Trust

Ask Keel and your data

Ask Keel is the assistant inside the Keel app. You ask a question about your workspace, it reads records through a fixed set of read tools, and it answers with links to the records it used. To do that it sends your question and parts of those records to an AI model. This page lists what is sent, what is not, and what Keel stores afterwards.

How to use it is in the Ask Keel guide.

Who receives it

Ask Keel sends its requests to Anthropic, through Anthropic’s API at api.anthropic.com, and uses an Anthropic model. Anthropic is listed on our Subprocessors page. Customer data submitted to Anthropic is not used to train any model.

Anthropic is the default set in Keel’s code, and two settings can change it. Keel’s platform administrators can save a model override for a single AI feature, and a deployment setting can choose the model for every feature that has no override. Keel’s list of models also includes models from DeepSeek and OpenAI, and nothing in the code stops either setting from naming one of them. If that happened, Ask Keel’s requests would go to that provider and this page would be wrong until it was changed.

A check of production on 23 September 2026 found no override saved. Whether the deployment setting is in use in production has not been checked.

What one question sends

Each call to the model carries a fixed set of instructions and one message. The instructions describe the tools the person asking is allowed to use, and hold no workspace data. The message holds:

today’s date;
the app page the panel was open on, as a path such as /risks/<id>. On a control, vendor, risk, policy or audit page it also gives the record’s type and id;
your question as typed, with email addresses replaced by [email removed];
the most recent messages of this conversation in the panel, both yours and Ask Keel’s, each cut to a maximum length and with email addresses replaced;
the result of every read made so far for this question: which read ran, what it asked for, how many records it shows and how many matched, and the records themselves.

One question can make up to 4 calls to the model, and each call sends all of the above again, including every earlier read result. The last call may not read, so a question runs at most 3 reads.

Each read returns at most 25 records. A record is sent as an id reference, such as risk:<id>, plus the fields below and no others.

Fields each Ask Keel read sends to the model
Records	Fields sent
Controls	Key, name, implementation state, owner name, number of linked evidence items.
Evidence	Title, kind, date collected, expiry date, owner name.
Vendors	Name, tier, residual tier, status, date last reviewed, notes. When the read asks only for vendors overdue for review: name, tier, last review, next review due, days overdue.
Risks	Title, description, level, inherent score, residual score, treatment, status, owner, number of mitigating controls.
Policies	Title, status, version, approval date, review due date.
Tasks	Title, status, due date, assignee name.
Audit programmes	Name, period, status, date created.

Before a record reaches the model, every text field in it has email addresses replaced with [email removed] and is cut to a fixed maximum length.

Two of those fields are free text that a person typed: vendor notes and risk descriptions. They are sent, shortened. Owner and assignee names are sent because people ask who owns something.

What is never sent

Email addresses stored on records. The owner, assignee and vendor contact email columns are loaded and then left out of every read. The email replacement above also catches an address stored in a name field.
Evidence files. No file is opened, and a read of evidence sends none of the record’s file name, storage key, link, description, content type or size.
Other long text: control descriptions, task descriptions, policy text and vendor websites.
Another workspace’s data. Every read runs under the signed-in person’s workspace, inside Postgres row-level security, and each read also filters on that workspace itself. No read accepts a workspace from the model.

Nothing at all is sent when your workspace has switched AI off, or when Ask Keel refuses a request before it starts, for example because the daily limit is reached, the plan does not include AI, or the workspace is out of credits.

What Keel keeps

Keel stores one record for each question that reached the model. It holds:

the question as typed;
the answer or suggestion as shown, or nothing when Ask Keel could not answer;
each read, with what it asked for and how many records it showed and matched;
references to the records the model was shown and to the ones the answer cited;
the suggested task’s fields, if there was one;
who asked and their role, the model used, the credits charged, the number of model calls, and how many records it read contained text addressed to an AI.

The record holds references to the records that were read, not copies of them. It does not hold the earlier messages of the conversation.

Workspace owners and admins can read these records on the Ask Keel history page in the app. They sit under the same row-level security as the rest of the workspace, and a record cannot be edited once written.

A daily job deletes each record 12 months after it was created. Two kinds of workspace keep records past that. The job finds workspaces through their members, so a workspace with no members left is not swept. A suspended workspace is skipped, and its records wait until it is restored or deleted. Deleting a workspace deletes its records.

The panel says this under the question box: “Questions and answers are kept for 12 months. Workspace owners and admins can see them.”

Known limits

The earlier messages of a conversation come from your browser. Keel shortens them and removes email addresses, but it does not check that a message presented as Ask Keel’s reply is one Ask Keel actually wrote.
The email replacement matches the usual name@domain.tld shape. An address written another way, such as “jane at example dot com”, is sent as written.
Only email addresses are removed. Vendor notes, risk descriptions and your question can hold phone numbers, names or anything else someone typed, and that text is sent.
Record ids are sent, so that an answer can link the records it cites. The page hint also carries the id of the record the panel was open on.
Keel has not confirmed how long Anthropic keeps the requests it receives, so this page gives no retention period for data held at Anthropic.

See how the rest of your data is protected on the Trust page, and every service Keel relies on under Subprocessors.