Trust

Trust

Get demo Security questionnaire

At a glance

The short version: the tool works where your people work, under the controls IT already enforces. No new perimeter to draw.

What stays on your machine

Screenshots, recordings, run evidence, vault credentials, and the documents the agent reads all live on the workstation's own disk. Because that workstation already sits inside your network, the agent inherits your firewall, endpoint protection, access controls, and monitoring.

What leaves the machine

Three flows carry work. A fourth, telemetry, is optional and carries product health only.

Your apps and portals Workflow actions Your model account Model calls Your portal Portal sync LaunchAI (health only) Telemetry (optional) Workstation
FlowWhat travelsWhere it goes
Workflow actionsThe submission, email, or upload a person would make anywayThe same third parties, from the same session and IP
Model callsOnly the screen region being read, plus one narrow questionThe model account you own (OpenAI, Anthropic, Azure, Bedrock) or a model you host
Portal syncRun status, review outcomes, rules and tables, device heartbeatYour organization's LaunchAI portal
Telemetry (optional)Product health and versionLaunchAI

A workflow decided entirely by exact rules can finish with zero model traffic. Point the setting at a self-hosted model and nothing crosses your perimeter.

Signing in without knowing passwords

  1. Step 1

    Stay logged in. The agent works inside sessions you already opened. It never meets a login screen, so it never types or stores a password.

  2. Step 2

    The vault, when you can't avoid it. Credentials go into an encrypted vault backed by the operating system's keychain. Workflows name an entry and never hold the value. Values are masked in every log, screenshot, and recording.

  3. Step 3

    MFA stays on. When a portal asks for a code, the run waits, a named person types it, and the run continues. Codes are never stored.

Least privilege, enforced

Each workflow is fenced. A workflow declares the apps, folders, and databases it may touch. Anything else is refused before it reaches the screen. An AP workflow can't open payroll, even if an invoice tells it to.

Each person holds separate rights. Build, Approve for production, Edit rules and tables, Review, Run, View history, and Manage credentials are seven separate permissions. The backend enforces them, and every change is logged.

Tenants, devices, and step-up. Each company or client gets its own tenant. Only enrolled machines run agents, and revoking one stops them. Sensitive actions can demand a fresh sign-in.

Try it: who holds what?

Seven permissions by sample role. Toggle a cell to grant or remove a right.
PermissionAP clerkControllerIT lead
Build
Approve for production
Edit rules and tables
Review
Run
View history
Manage credentials

Threat model

Threat: A doctored PDF or web page tries to redirect the agent Our control: Screen text is treated as data; money decisions run on exact rules; the workflow fence refuses other apps Your part: Place a review step before posting and payment
Threat: A fake supplier asks to change bank details Our control: Review step with a call-back to a number on file Your part: Write the call-back policy
Threat: An insider edits a payment workflow Our control: Build and Approve are separate; approved versions are read-only and versioned Your part: Keep the two rights with different people
Threat: A laptop is stolen Our control: Revoke the device and its agents stop; vault entries are encrypted Your part: Full-disk encryption and screen lock
Threat: Someone tries to cover their tracks Our control: Records are append-only and hash-chained per run Your part: Set retention so evidence outlives questions
Threat: An impatient double-click on Run Our control: Each run request carries its own idempotency key Your part: Give writing steps a business key the ERP checks

Joiners, movers, leavers

Workflows, tables, and evidence belong to the tenant, not the person.

  • grant only the rights the role needs
  • add them to review assignments
  • name vault entries for the system, not the person

Once a quarter, confirm who holds Approve, Edit rules and tables, and Manage credentials.

If something goes wrong

Stopping isn't rolling back. A posting that landed stays posted, and reversing it is a person's call in the ERP.

  1. Stop

    Abort the run, pause the schedule, revoke the device, or remove a person's rights. Pick the smallest stop that contains it.

  2. Investigate

    Open the run, trace the bad value to its source, check which version ran and who approved it.

  3. Rotate

    Change affected vault entries, account passwords, and your model key.

  4. Restore

    Roll back to the last good version, get it approved, and supervise the next run.

Your model provider, your settings

The model account is yours, so its retention and training terms are yours to set.

  • OpenAI
  • Anthropic
  • Azure
  • Bedrock
  • self-hosted
  • Create a dedicated project or workspace for LaunchAI
  • Request zero retention where the provider offers it
  • Pin the region your data-residency policy requires
  • Keep regulated workflows on a self-hosted model or exact rules only

What we don't claim

  • No certification claims on this site. Ask for current attestation status during your review.
  • No accuracy or uptime percentage.
  • No guarantee a portal will never show a captcha.

Security questionnaire

Does LaunchAI receive our business data?
No. Screen content, documents, credentials, and extracted values never reach LaunchAI.
Is our data used to train models?
Not by LaunchAI. Training and retention at the model provider follow your account's terms.
Could a malicious document redirect the agent?
Not past the fence. Text read from a page is data, never an instruction, and money decisions still have to clear an exact rule or a reviewer.
Who can see screenshots of our screens?
Only people with View history, and only for as long as each workflow's retention allows.
How do we tell the agent's ERP entries from a person's?
The run record shows every entry the agent made. If auditors need the split inside the ERP, run under a dedicated service account.
Does it install a browser extension or inject scripts?
No. It uses OS-level input in your standard Chrome or Edge profile.
What does support see?
Only what you put in the thread. See the Support page.

For the architecture behind these controls, see Engineering.

Inside Trust

No prompts. No babysitting. No Integrations.

Get a demo with our specialist team.

Get demo