AI Disclosure

CogniWork is built on large language models. This page says plainly what that means for your data and your decisions. It is written to be read before you trust the thing, not after.

What the models see

To answer a request, a model receives:

• your message,
• the relevant part of your workspace — the records, schema, files or view definitions needed for that request,
• the workspace memory relevant to the task,
• results returned by tools it calls, including from integrations you have connected.

It does not receive other customers’ data, or data from workspaces you are not working in.

Which models

CogniWork uses third-party foundation models, and may use models we host ourselves. Providers are named in Subprocessors, which is the authoritative, dated list.

Model choice varies by task. Where it matters to you, it is visible and configurable: each AI worker shows the model it runs on and lets you change it.

If we add a model provider that processes customer content, we update the Subprocessors page and notify workspace owners 30 days before the change takes effect, so you can object.

Training

Your content is not used to train foundation models — ours or our providers’.

Our agreements with model providers prohibit training on data submitted through their APIs. Providers may retain inputs briefly for abuse monitoring under their own published terms; we use enterprise API terms rather than consumer products for exactly this reason.

We use aggregated, de-identified telemetry — feature counts, error rates, latency — to improve the product. That never includes the content of your records, files or conversations.

What the system can do on its own

This is the part worth understanding properly.

CogniWork is agentic. Configured to, it will take actions without a person present:

• create, update and soft-delete records in your workspace;
• create, change and delete views and workflows;
• run code against files in your workspace’s own isolated environment;
• call tools on systems you have connected, using credentials you supplied;
• send messages through channels you have connected;
• run on a schedule, unattended.

The controls on that:

• Nothing is scheduled unless someone schedules it. Everything scheduled is listed, with its history, under Scheduled jobs.
• An AI worker in a group chat only acts when directly addressed. Un-addressed, the most it does is offer.
• An integration can only reach what its credential allows, and you choose the scope when you connect it.
• An external agent connecting to your workspace gets the workspaces you granted it and no others — a grant is a ceiling, and your live membership is the floor.
• You can interrupt any run mid-flight.
• Record deletion is soft by default and recoverable.

Where it gets things wrong

Being specific is more useful than a disclaimer.

• Confident invention. A model may state a figure or a relationship that does not exist in your data. Views that display numbers are backed by a query you can read — do read it when a number matters.
• Wrong interpretation of your intent. “Last quarter” can be read three ways. Ask what it computed if the answer surprises you.
• Stale assumptions. Memory can hold a fact you have since changed. Correct it in a sentence and it stops applying.
• Partial completion on long jobs. A worker that hits a limit reports what it finished rather than silently claiming success. Read the run summary.
• Non-determinism. The same request twice may produce different phrasing, different structure, or a different reasonable choice.

Do not use it for

Without competent human review, do not rely on CogniWork outputs for: medical, legal or financial advice; safety-critical decisions; employment, credit, housing or insurance decisions about individuals; or anything where being wrong is not recoverable.

Automated decision-making that produces legal or similarly significant effects on a person requires meaningful human review — this is a legal obligation in several jurisdictions and a condition of our Acceptable Use Policy.

Human review, on our side

We do not read your workspace content routinely. Access by our staff is limited to:

• what you explicitly share with us in a support request,
• an investigation into a specific abuse report or security incident,
• where legally compelled.

Such access requires authorisation, is logged, and is limited to what the task needs.

Questions

[email protected] for data questions, [email protected] for anything else.