Precision AI OS is built for businesses that hold sensitive client information. Every business's data is kept apart, encrypted with keys unique to that business, and handled by rules written into the system, not into a policy binder. Read how we keep CPA client data private on the CPA firms page.
A named agent. A clear job. A person responsible for review.
Hi Alex,
I’ve prepared the proposal following our discovery call. It brings together the scope, deliverables and suggested next step.
Let me know what you think.
Jamie
The draft and the recipient stay with the decision.
How data is kept separate, how model access is constrained and what happens to the record.
Every table in the database carries a business identifier, and the database itself, not just our application, refuses to return another business's rows. Row-level security is enforced on every table, and the accounts our services use to reach the database can't switch it off. Only two narrowly scoped database functions can resolve a user's business, and our tests check that no others exist.
Documents and sensitive data are encrypted with AES-256-GCM using a key for your business alone, and that key is itself protected by Google Cloud Key Management Service. Encrypted files are sealed so any tampering is detected before they're read, and they're bound to your business, so a file can't be moved into another business's space and opened there.
Language models run only under rules we write into the code. Only models on a list we review can be called. Every request asks our gateway for zero data retention, refuses data collection and is limited to the provider endpoints named for that model, with the gateway's own fallback off. If a model is unavailable, the same request may go once to a second listed model, which can be a different provider, unless that is switched off for the agent.
We rely on the gateway's published zero-retention listings; we don't yet hold a signed zero-retention agreement with each provider. Providers may hold a prompt briefly, including in a short-lived cache, to process requests.
Every model call is recorded: which skill made it, which model and provider answered, and what it cost. Scanned documents and pictures are read by a language model under the same rules.
When your team approves work, the approval is written to a record that can't be edited or deleted through the Platform during its retention period, including by your administrators or ours, and it's tied to exactly what was approved.
Agents prepare the work and, by default, a person approves it. In a workflow, the approval goes to the person who runs it; for sends covered by a business rule, to the person the rule names. Only that person can approve, and the approval is recorded. The platform doesn't require the approver to be someone other than the person who ran the step.
A legal hold stops deletion of anything it covers. Deletion runs as a preview by default, and nothing is removed until the deletion is applied. When a business leaves, its encryption keys are not destroyed until its owner has confirmed receipt of a readable export of its data.
Its encryption keys are then destroyed, no sooner than a day after destruction is scheduled, and from then on any encrypted copy that remains, in live storage or in backups, can no longer be read. Database backups expire within 30 days. A legal hold blocks key destruction until it is released, and records kept for a required retention period stay protected until that period ends. The exit steps are agreed in your contract.
Email legal@getprecision.ai with legal or data-deletion requests.
Agents prepare work; your people make the calls. By default, anything going to your clients, and anything an agent isn't sure about, waits for a person on your team. An owner can allow routine replies; never money, new people or deletions. Every approval is recorded.
How human review and approval work
For operator access from an AI client, see the MCP access and tool controls.
Have your IT lead or security reviewer walk through it with us.
No. Every record carries your business's identifier, and the database enforces row-level security on every table, so queries can't return another business's data even if application code made a mistake.
We don't train models on your data. Every model request requires zero data retention and refuses data collection, so a provider that keeps or trains on prompts can't be routed to. We rely on our gateway's published zero-retention listings rather than a signed agreement with each provider.
Scanned documents are read by a language model under the same rules.
With AES-256-GCM using a key for your business alone, protected by Google Cloud Key Management Service. Encrypted files are tamper-evident and bound to your business, so they can't be opened in another business's space.
On Google Cloud in the United States. Database backups are kept daily for 30 days, with point-in-time recovery over the last seven days.
Not yet. Our controls and written policies are designed against SOC 1 and SOC 2 criteria, and we'll publish audit status when it's complete. We can share our security packet during a discovery call.
Offboarding respects any legal hold and follows the exit steps in your contract. Keys are not destroyed until your owner confirms receipt of a readable export, and no sooner than a day after destruction is scheduled. Encrypted copies then can't be read. Database backups expire within 30 days.