Privacy Considerations for the Troubleshooting Agent

Logging and Storage

The Troubleshooting Agent logs and stores AI prompts and completions within LangSmith (see below for more information) for the sake of performance evaluation, which includes measuring accuracy, monitoring for misuse, and improving reliability of responses.

Monte Carlo also uses internal tooling to observe the Troubleshooting Agent and retains prompt data and aggregated results. This is a replication of data already stored within the Monte Carlo platform and does not include row level data.

Information Shared with Subprocessors

Amazon Bedrock

The Troubleshooting Agent sends prompt data, which includes user-entered prompts + relevant data (which may include row level data if data sampling is turned on), necessary to generate an answer (completions) to be processed by Amazon Bedrock within the Monte Carlo AWS environment.

Amazon Bedrock is not allowed to store or use any data submitted through Monte Carlo to train its models. The Bedrock models do not retain any data.

LangChain (LangSmith)

Monte Carlo uses an observability and evaluation application, LangSmith, to monitor the Troubleshooting Agent, as noted above.

Due to this monitoring, if data sampling is enabled, log output (traces) sent to LangSmith may contain aggregated row level data (sampled customer data).

Data Protection Practices

Because the Troubleshooting Agent requires the sharing of data with a third party, Monte Carlo implements the following protections:

• The Troubleshooting Agent only shares the minimum data necessary to enlist the help of the AI subprocessor in producing troubleshooting insights.
• The Troubleshooting Agent only accesses data on a just-in-time basis dictated by user interaction.
• Data is strongly isolated and segmented between customers. Data from one customer is never co-mingled or shared with another customer, in line with Monte Carlo’s core security practices.