Human approval where it matters
Every workflow has a defined point where output waits for a person. You will know where that point is, because it is written into the workflow's specification before it is built.
Trust & safety
How access actually works.
Automation is only worth having if you can describe, out loud, what it is permitted to do. This page is that description. It is written to be checked, not to be reassuring.
The envelope
Read the inputs. Interpret them. Prepare the work. Leave it as a draft or a file. A person decides what happens next.
Nothing we build sends an email, publishes a post, moves money, or deletes a record on its own. If a workflow would need to, that is a separate conversation with its own design and your explicit written approval — never a setting we quietly enable.
Least access that works
Each connection gets the narrowest permission that does the job. If a workflow only needs to read your inbox and create drafts, that is all it is authorized to do — not full account access because it was simpler to set up.
Isolated deployments
Managed clients each get their own database, their own credentials, their own encryption key, their own backups, and their own hostname. Nothing is shared between clients, so nothing can leak between them.
Credentials stay out of source control
No client credential is ever committed to a repository. Secrets live in encrypted storage with restricted access, and rotation is documented rather than improvised.
Monitored, with alerts pointed at us
Health checks, failure alerts, and a dead-man's switch on the backup job. The alerts reach us first. You should hear about a problem from a person, not discover it from silence.
Backups you can actually restore
Daily, kept off the machine they back up, and — the part most people skip — restored on a test system to prove the procedure works. A backup nobody has restored is a guess.
Revocation is yours, not ours
Access is granted from your own accounts, which means you can withdraw it from your own accounts, immediately, without asking us and without waiting for a business day.
Export and offboarding on request
Ask, and you get your workflow definitions and your data in a portable form that imports elsewhere. Leaving is a procedure we have written down, not a negotiation.
The part most vendors leave out
What we do not claim.
We hold credentials for managed clients
Managed automation runs without you, which means the connection to your accounts is stored on infrastructure we administer. There is no version of always-on automation where this is not true, and any vendor telling you otherwise is describing something else. We isolate it per client, we say so in writing in the service agreement, and we make sure revocation is one click you control.
If you are not comfortable with that — and it is reasonable not to be — the Personal AI Assistant runs on your own machine and never has this property.
We hold no compliance certifications
We are not SOC 2 certified, not HIPAA compliant, and not ISO 27001 audited. We are a small operation that follows sound practice, and we would rather say that plainly than imply a certification we do not have. If your work requires one, you need a vendor who has been audited for it.
AI gets things wrong
The models we configure will occasionally misread context, state something confidently that is incorrect, or produce a draft that misses the point. This is exactly why the approval step exists and why nothing sends itself. Part of what we train you on is recognizing when output is wrong — including how to check a claim before you act on it.
Sensitive and regulated work is excluded by default
Workflows touching regulated data or decisions with legal consequence are out of scope unless separately assessed and designed. We would rather turn down that work than build it casually.
Questions about any of this?
Ask them before you buy something, not after. If something here does not fit how you need to work, that is worth finding out in the first conversation.