Trust / Security and control
What is already running, what we deliberately do not claim, and the principles behind what we build next.
In production today / OPENSIGHT Lens
These are properties of how Lens is built and run today, not promises. Your own IT team can verify each one.
Lens reads your ERP with a database account that has read permission and nothing else. Writing back is not something we promise to avoid; the database refuses it. Your IT team can confirm the grant with one query and revoke it with another.
The standard setup runs on your own server, Windows or Linux, next to your ERP. Nothing is sent to a third-party cloud by default, and there is no per-seat meter.
Nobody sees a page until it is granted, and every grant is per person and per company. Creating a powerful account such as Owner or CEO requires the administrator to re-enter their own password.
Activity is written to a log that can only be added to, protected in the application and by a database trigger. You can also see which reports were opened, by whom, and when.
Every figure is computed from the ledger by deterministic code. AI can help people ask questions, but it never produces a number. Same data, same result, every time.
During adoption your ERP is never modified. Revoke the read access and it is exactly as it was: no cutover to unwind and no data to bring back.
Plain speaking
We hold no ISO or SOC certification and will not imply otherwise. Instead, an architecture you can inspect: a read grant you control, on a server you own.
Phrases like that say nothing. We would rather state exactly what the system can and cannot do.
We do not say your data is safer with us. It stays on your premises. That is the design.
For the agents we are building next: Respond, Reach, Chat, Flow, and the others. These are design commitments, not claims of certification or completed deployment.
An agent should verify the person and the channel before sharing sensitive information or changing a record.
Each workflow should expose only the knowledge and actions that role requires.
Business facts should come from trusted systems. When the answer is unsupported, the agent should say so and escalate.
Judgment, exceptions, and sensitive outcomes need an explicit human path.
Teams need to understand what the agent said, which tools it used, and where the workflow failed.
New scenarios should be evaluated against normal requests, edge cases, and failure conditions before wider use.