Notes / Workers
How AI workers handle sensitive business data
What happens to your information when an AI Worker reads a document, prepares an update or uses a model—and what to agree before it starts.
By Edu Rigonato. Published Oct 25, 2026. 14 min read.
The document you want an AI Worker to process may contain a policyholder’s address, a patient’s history or details of a claim. Your team needs help with the work, but you’re responsible for what happens to that information.
So when an AI provider says, “We don’t train on your data,” you may still have questions. Does the company keep a copy? Could someone there read it? Can you use AI without sending those details to a model provider at all?
You can build a Worker around those concerns. Its access, the models it uses, the information it saves and the people who can see its work all need to be agreed for the job. That takes more than choosing a familiar AI brand.
Imagine a Worker helping your team maintain case records. It receives a document, finds the matching case, spots missing information and prepares an update for someone on your team to check.
What information does an AI Worker need?
Start with the responsibility you want to hand over. In this example, your Worker needs to read incoming documents and compare them with existing case details. It might need a case reference, contact information and a few fields that show what’s missing.
It doesn’t follow that it needs your whole file library. Payroll records, other departments’ folders and account administration have nothing to do with preparing that update. You may also want it to propose changes while leaving the final decision with your team.
Those choices should shape the account and permissions the Worker uses. Telling it to stay within a certain folder is useful, but the software connecting it to your records needs to enforce that limit too. OWASP identifies sensitive-information disclosure as an AI risk and explains why instructions alone aren’t enough protection.
Frequently asked questions
Can we start a pilot without using real customer information?
Often, yes. Invented examples can help test the workflow and reveal missing requirements, though they may not cover every situation your team encounters. Agree what the pilot will establish and what needs checking before real records are introduced. Moving personal information into a test environment doesn’t make it anonymous.
Does a separate Worker environment mean nobody else can access the data?
No. Ask both how your environment is separated from other businesses and which administrators, support staff or providers can still access it. Logical separation doesn’t necessarily mean dedicated physical hardware. The access arrangements need their own review.
Can all processing stay in a country or region we choose?
That needs checking across the full job, including model processing, storage, logs and relevant providers. The location of the Worker alone won’t answer it. Microsoft’s deployment options, for example, affect processing geography. Tell us your location requirements so we can assess the available arrangements against them.
Is “HIPAA certified” enough to approve an AI Worker?
No. HHS doesn’t recognize private Security Rule certifications as removing your organization’s legal obligations. Your review still needs to cover the work itself, the responsibilities of everyone involved, the agreements and safeguards. HHS
Evidence note: Sources linked throughout include official standards, regulators and model providers. Provider terms were reviewed in October 2026 and can change. The records job and data-flow diagram are illustrative; they don’t establish that a particular deployment meets legal or security requirements.
Explore AI for your business