Declare
Write the outcome, the autonomy budget, and the guardrails into a signed, versioned policy file. Diffable in a pull request.
Declare the policy, replay it on last quarter, ship it, reverse it if you must. Your risk team follows the thread from the rule to the cash in fewer clicks than approving an expense report.
Each step leaves a record the next reviewer can read. Nothing happens that a person could not have seen coming.
Write the outcome, the autonomy budget, and the guardrails into a signed, versioned policy file. Diffable in a pull request.
Replay the policy on last quarter's real transactions. See the cash and the risk before a single customer is touched.
Go live in suggestion mode, the agent drafts, a human signs. Autonomy widens only as the accept rate earns it.
Inside the window, one signed reversal unwinds the action and posts the credit. The original and the reversal both stay on the ledger.
A high-level view of the request path in production, authenticated, policy-checked, executed, logged.
Governance that is added after launch is theatre. These are load-bearing parts of the platform. You cannot ship an agent without them.
Every rule runs under a spend cap in euros or count. Exhaust it and the rule circuit-breaks to a human.
Revenue-material actions need two signatures before they post. Default on; configurable downward only.
Every action carries a redo path. Confidence comes from undo, not from a probability score.
Signed, versioned, PR-reviewable. Any change that widens blast radius needs two approvals.
Every action is logged and can't be edited after the fact, only reversed with a matching credit. Tenant-scoped, exportable to your GL.
Self-hosted in your environment, or EU-hosted by us in the region you choose, with your own encryption keys (BYOK). Model calls pinned in-region either way. Your data never trains a foundation model.
Kaiva is pre-launch and working with a small design-partner cohort. Rather than imply certifications we do not hold, here is the honest position, your procurement team can plan around it.
We would rather answer the hard questions on day one than paper over them in a pilot. Come with your worst case.