What information does Uncountable send into an LLM model?
In accordance with the principle of least access, Uncountable’s Generative AI features expose the minimal amount of information necessary for an AI assistant to help a user with their tasks. In general, when a user is using an LLM assistant feature, Uncountable will expose roughly the information that is on the user’s screen, or easily accessible from it (e.g. in a dropdown menu), to the LLM. Different features will require different information to function.
For example, if a user is asking an assistant to help them enter values in an experiment via voice control, the LLM would see the structure of the specific experiment on the user’s screen so that it can record and structure the values read out by the user and record them on the recipe.
If the user is running a search over structured data, the LLM would see the filters and columns that are available on the user’s screen, a sample of the search results that the user is viewing, and the filters/columns that the user could add. This allows the assistant to help the user refine their search by adding new search filters and criteria.
Other features will expose different information to the LLM depending on the use case. For instance, an AI data scientist feature that automatically analyzes experiments and identifies trends would look at selected samples of experiment data, charts showing correlations in the data, etc.
How secure is our data when it is sent to third‑party LLM providers? Will it be used to train other models?
When Uncountable’s LLM features are enabled, a limited set of on‑screen context and the user’s prompt are sent to our LLM providers over their enterprise APIs, not consumer endpoints. These services are governed by enterprise terms that state prompts are not used to train public models. Uncountable does not use customer data to train or fine‑tune models that are shared across customers, and enabling LLM features does not change the confidentiality or security guarantees in our services contract.
For customers who want more detail, we can share links to our vendors’ enterprise data‑handling terms (AWS Bedrock, OpenAI, Azure OpenAI, Anthropic, Google Gemini).
Does Uncountable use Customer data to fine-tune or train global models shared across customers?
No, Uncountable does not use Customer data to train or fine tune generative AI models that are shared between customers, and does not integrate customer data into shared prompts. The use of LLM features does not change the guarantees Uncountable makes under its services contract regarding the confidentiality or security of your data.
What LLM models does Uncountable use?
LLMs are a new technology, and Uncountable wants to provide the best possible experience. Therefore we will use a mix of vendors and models for different use cases. The vendors we use are all best-in-class, industry-leaders in the GenAI space: Amazon Web Services (AWS), Microsoft (Azure), OpenAI, Google, and Anthropic. The models Uncountable uses today include GPT-4o, GPT-4.1 mini, Claude Sonnet, and others.
Uncountable pursues a policy of continuous improvement and integration for Generative AI technology. As LLMs continue to advance and improve, Uncountable will also continue to assess, benchmark, and adopt new and better models. Therefore, the suite of models used in 6 months may be different from the blend used today by Uncountable.
How does Uncountable handle changes and upgrades to LLM models?
LLM vendors continuously replace and remove older models, and newer models are often faster and more effective. Therefore, we aim to make sure that your users benefit from any upgrades to models available on the market, and are robust to any changes in availability of existing models and vendors.
Uncountable maintains a regression testing (eval) suite to benchmark performance of new models against old ones on standard tasks and examples. Before introducing a new model to existing features, we run this regression test suite to evaluate if the new model can do a consistently better job for users than an old one. We perform other testing and evaluation as well. Uncountable designs our AI features to be generic as possible across models so that new ones can be introduced with minimal friction, and rolled back if needed.
Is Generative AI data retained by Uncountable, and for how long?
Similar to other features in Uncountable, portions of user queries and responses are logged according to Uncountable’s standard data retention policies (Until customer requests deletion of data or end of contract. At any time the customer can request deletion of data, and Uncountable will produce a certificate of destruction signed by an officer of the company). Uncountable stores long-term log data in AWS Cloud.
How Uncountable’s LLM inference vendors handle data isolation and retention?
We have included below for your reference details about our current LLM inference vendors and their terms of service:
- Amazon Web Services: Uncountable would use AWS Bedrock, Amazon’s AI platform, and other AWS services. AWS’s service terms can be seen here: https://aws.amazon.com/service-terms/.
- OpenAI: Uncountable would use the OpenAI “API Platform” and other enterprise services. OpenAI’s policies around enterprise data handling are currently available here: https://openai.com/enterprise-privacy/ and https://openai.com/policies/business-terms/.
- Microsoft Azure: Uncountable would use Azure OpenAI, Microsoft’s AI platform, and other Azure services. Azure’s terms can be seen here: https://www.microsoft.com/licensing/terms/productoffering/MicrosoftAzure/MCA#ServiceSpecificTerms and a FAQ regarding Azure OpenAI https://learn.microsoft.com/en-us/legal/cognitive-services/openai/data-privacy.
- Anthropic: Uncountable would use the Claude API and other enterprise services. Anthropic’s terms can be seen here: https://www.anthropic.com/legal/commercial-terms.
- Google: Uncountable would use Google Gemini and other Google AI services. Gemini’s terms can be seen here: https://ai.google.dev/gemini-api/terms#data-use-paid
Can Uncountable AI allow a user to see or modify data they don’t have access to?
No, we do not allow users to bypass or circumvent our normal permissions and access controls using AI or LLM tools. AI assistants in Uncountable are sandboxed by the permissions of the users who activate and direct them. For instance, if a given user only has access to a certain limited set of projects in Uncountable, the AI assistant they trigger will only be able to access that limited set of projects.