Training clauses in enterprise software
Do enterprise vendors train AI on customer data? Their own words
Updated
Before a company licenses its record to an AI buyer, it helps to know whether the software holding that record already uses it to train models.
Ask about my company's record 4 of the 8 vendors we read leave title to the data with the customer. Tell us which systems hold yours and what stage the company is at, and a person will reply about your own case.
The vendors that say no
Snowflake's trust page says it does not use usage or customer data, inputs and outputs included, to train any model made available to other customers. Databricks says data submitted to its assistive features does not train the foundation models it offers to third parties, and Palantir says no customer data is used to retrain the third-party models AIP calls.
Source: Enterprise software data terms
Where consent or a product line changes the answer
Microsoft's Copilot guidance for Dynamics 365 and Power Platform says customer data does not train Copilot unless the customer consents. Epicor's Prism page keeps customer data out of base model training unless fine-tuning is done for that customer. AWS's service terms let a few named services, CloudWatch's machine learning features among them, use content to develop and improve themselves.
Source: Enterprise software data terms
What a buyer will check
Of the 8 vendors on our board, 2 leave training unaddressed in what we read. Silence in a website summary is not a licence either way: the subscription agreement and any AI addendum are what a buyer's lawyers will ask to see.
Elsewhere on this site
Every vendor's four answers are on the home page board, and the organisations that take the record are on who buys enterprise data.