GembaOS
How GembaOS works: routes, changes and decisions
Start with the business question, then inspect the mechanism.
What may AI do?
The playbook declares the decision route. Export a business-readable view of the same configuration the runtime uses, showing where the model, the host and a person each act. The review view retains identifiers and exact expressions.
Read the generated route and diff example
What changed in the rules?
Compare two playbook versions to inspect thresholds, signature requirements and data visibility. Changes can be labeled looser, stricter or data-related where the tool can determine direction. Unrecognized expressions remain for a person to assess.
Read the complete change explanation
What am I approving?
The approval view starts with the relevant summary, checked facts, changes since the last visit and the action being authorized. Source material remains available. Routine requests may follow automatic policy, and notification recipients need not become additional approvers.
Read what happens when approved facts change
Choose deployment and data paths with your technical team.
GembaOS is a zero-trust governance runtime for enterprise AI agents: policy-decided access, single-use clearance tokens bound to exact parameters, and accountable human approvals, running on-premises or in a private VPC.
Unified governance standards across business units.
GembaOS includes four representative enterprise workflows. Ownership and approval criteria are generated directly from playbook configurations, adapting automatically when rules evolve.
AI agentNever holds the directory’s admin password; it can only ask
- IT staffAsks for a leaver’s account to be disabled, one or a list
- AI agentLooks the account up
- AI agentPrepares the stop, with the reason
- GembaOSYour rule decides who must sign
- Your peopleIT administrator signs, and may be the one who asked
- Your peopleEmergency: IT administrator signs with a hardware key and a reason
- Your peopleSecurity is told, by name, every time
- IT staffThe account is disabled
- GembaOSDoes exactly what was signed, onceDisable account
Generated from the account-disable playbook that ships with GembaOS. Thresholds are its samples; you set your own.
AI agentPays nothing itself
- CustomerAsks for a refund
- AI agentReads the order and the payment, and checks the refund terms
- AI agentProposes the refund and drafts the reply
- GembaOSYour rule decides who must sign
- AI agentReviews the refund and recommends, with its reasons: advice, never a decision
- GembaOSUp to 3,000, low risk: no signature, the rule approves
- Your peopleUp to 50,000: CS supervisor decides
- Your peopleAbove 50,000, or a file raised a concern: CS supervisor decides, after a second AI review
- CustomerGets the answer
- GembaOSRe-checks the record, then does exactly what was signedExecute refund
Generated from the refund playbook that ships with GembaOS. Thresholds are its samples; you set your own.
AI agentMerges nothing itself
- EngineerAsks for a failing test to be fixed
- AI agentReproduces the failure in a sandbox and writes the fixRun CI · Clone repository · Commit · Show diff · Apply patch · Read file
- GembaOSRefuses any change to protected files, such as deploy settings and secrets
- GembaOSYour rule decides who must sign
- AI agentAI review (Team Lead role) reads the change first; an objection brings in the code owner
- GembaOSUp to 5 files, tests passing, no owned code: the rule your company confirmed merges it
- Your peopleTests passing, bigger or touching owned code: Code owner
- Your peopleTests failing or missing: Code owner → Platform lead
- EngineerGets the merged fix
- GembaOSRe-checks the record, then does exactly what was signedMerge branch
Generated from the bug-fix playbook that ships with GembaOS. Thresholds are its samples; you set your own.
AI agentProposes no amount, makes no decision, signs nothing, executes nothing
- CustomerFiles a claim with documents
- CustomerSends what was asked for
- AI agentReads the claim, the policy and every documentRead claim · Read policy · Read the policy wording
- AI agentSorts the documents and writes a brief: amount, date, what happened, each with its source
- AI agentPoints out what does not add up, quoting the policy clause, and names who to consult
- GembaOSHolds the brief’s amount and date to the claim record
- GembaOSCounts the required documents and asks for any that are missing
- GembaOSYour rule decides who must sign
- GembaOSUp to 20,000, documents in order: no signature, the rule approves
- Your peopleUp to 50,000: Claims adjuster
- Your peopleAbove 50,000: Claims adjuster → Claims manager
- Your peopleMedical review can be called in
- CustomerGets the answer
- GembaOSRe-checks the record, then does exactly what was signedReserve payment on the claim
Generated from the claims playbook that ships with GembaOS. Thresholds are its samples; you set your own.
Last updated: · GembaOS v0.2