A forty-page supplier contract sits in a shared folder. Someone opens an AI assistant, attaches the file, and types one line: summarize the termination terms and flag anything unusual. Seconds later, a definitive answer comes back. From where that person sits, the whole exchange happened in a single window.
The contract traveled much further than that. Inside it were names, signatures, wire instructions with bank account numbers, and a home address in the notices clause. That is Personally Identifiable Information (PII), carried in a document nobody would think to call sensitive.
The risk is measurable. IBM’s 2025 Cost of a Data Breach report found that 13% of organizations had a breach involving their AI models or applications, and 97% of those lacked proper AI access controls. One in five organizations also reported a breach tied to shadow AI, meaning AI tools used without approval. Seeing the full route your data takes is where control begins.
Where your data goes after you click upload
The first stop is the upload itself. Sensitive details sit inside ordinary files: a pasted email thread, a spreadsheet exported for a quick question, a scanned form with a tax ID in the corner.
Next is the assistant or internal tool wrapping your request. It adds its own instructions, pulls in related documents for context, and saves the conversation so you can return to it. Each of those steps puts a copy of your data somewhere new.
If the assistant is linked to your email, file storage, or customer system, it can reach into those sources too, including material you never attached.
Then the AI reads your request and writes its answer. That work happens on computers owned by a provider, in a regional data center, or inside your own organization. Behind a provider there are usually other companies handling storage, monitoring, and support, each one named somewhere in a contract.
The copies that stay behind
Prompts and answers are often stored for troubleshooting, abuse monitoring, or billing. Each store has its own retention clock and its own readers. Chat history keeps the exchange around long after the person has moved on.
Search indexes go further. To find documents by meaning, a system breaks them into passages and stores a numerical summary of each one. Those summaries are built straight from your text, and they can reveal more of it than you would expect. Delete the chat and the index entry can remain, along with the log line that recorded the request.
Five questions to ask any AI vendor
- Where does the AI do its work?
- What is kept, and for how long?
- Who can read the logs?
- Can my data be used to improve the AI?
- Are sensitive values replaced with secure tokens before the AI sees them?
A vague answer to any of these means the vendor can’t show you where your data goes.
The route Pensato builds: short, and inside your walls
Here is the path that same contract takes through Pensato’s secure AI:
- You upload the contract.
- Pensato finds the PII (names, account numbers, addresses) and swaps each value for a secure token, a stand-in code that holds its place.
- Only what the task needs moves forward. For termination terms, that is a few clauses, not the whole file. A gateway checks the request on the way, much like a mailroom clerk who inspects every outgoing envelope and pulls anything marked confidential before it leaves the building.
- A local LLM reads the tokenized text and writes the answer. Local means the AI runs on hardware at your location. It is not sent to the cloud to do its work.
- The answer returns, and the tokens are swapped back for the real values inside your environment. The token vault, the lookup table that matches each token to its real value, stays with you.
- Every step is recorded in an audit trail, and people make the decisions.
That contract will be uploaded again tomorrow, by someone else in the company. When PII travels as secure tokens and the AI runs on a local LLM inside your environment, that upload is routine work.
We are building Pensato so your PII stays local from the start. Learn how we keep PII out of public LLMs with local AI and secure tokens.