AI agents are new enough that the words for governing them are still settling. Here is how we use them, in plain language, with a link to where each one shows up in the product or the regulation.
Core concepts
Making autonomous AI systems accountable: knowing which agents run, recording what they decide, bounding what they may do on their own, and being able to prove all of it after the fact. It is the layer beneath the agents, not another agent.
A model is a scoring function: inputs in, output out. An agent is a system built around one or more models that retrieves context, calls tools, chains reasoning, and acts. Two firms can run the same model with very different risk, because risk lives in what the agent is allowed to do.
A live inventory of every agent and model in production: its owner, autonomy tier, provenance, and a verifiable identity. The agent-era equivalent of KYC. Nothing runs unregistered.
See the platformCapturing each decision at the moment it happens: the model identity, the tokenized inputs, the reasoning steps, and the outcome, written to a record that cannot be quietly edited later. Post-hoc reconstruction of an AI decision is a fiction; the record has to be contemporaneous.
How it worksA sequence of decision records where each one is hashed together with the hash of the one before it. Recomputing the chain re-derives every hash, so any change is not only detectable, the chain shows exactly where it happened. Nothing to take on faith.
The policy that decides, per action, whether an agent may act on its own, must route to a human, or is blocked, based on stakes and confidence. A formally checkable set of rules, not a vibe.
Human-in-the-loop only means something if the loop is measured. A near-zero override rate on a high-volume queue is not a good model; it is a reviewer who has stopped reviewing. Risk-tiered deep review beats shallow ratification of everything.
Why regulators careThe control that halts an agent or narrows it to safe mode. An untested kill switch is a compliance claim, not a control, so it should be drilled on a schedule with signed attestations of propagation time and scope.
Telling a customer when a decision involved AI, and giving a clean path to a human. Increasingly an obligation, not a courtesy. Recorded per decision like everything else.
A signed bundle that maps your live decision records to a specific regulator's obligations, clause by clause, ready to hand to an examiner. Generated on demand, verifiable by the recipient.
Replacing personal identifiers (Aadhaar, PAN, card, account numbers) with reference tokens before anything is stored, so raw sensitive data never enters the evidence. Where the raw value must live somewhere, it goes in a compliant vault and Sakshi only ever holds the reference.
Checking decisions for disparate impact across declared and proxy attributes, gated on statistical significance so the screen does not cry wolf, and run on real production decisions, not just a pre-deployment benchmark set.
India's AI rulebook
The reason a governance layer for India looks different from a Western one. Each is encoded as data and scored against your live evidence.
The Reserve Bank's draft guidance on managing model risk across banks, NBFCs, and more. Its definition of "model" is broad enough to cover essentially every AI system a financial institution runs, including agentic ones. Names inventory, validation, human oversight, and a kill switch as expectations.
RBI mappingThe Digital Personal Data Protection Act (2023) and its draft Rules. For a Significant Data Fiduciary, it implies algorithmic due diligence: a DPIA from live evidence, data minimization and localization, and an auditor bundle.
DPDP mappingBinding since 2025: the regulated entity carries sole responsibility for AI and ML outputs, including those from third-party models. Which makes verifiable vendor model identity, on every decision, not optional.
Anticipated insurance-sector guidance, read across from the FREE-AI committee's direction. We keep a skeleton mapping ready to cut the day the working-group report lands.
Open a synthetic-data sandbox and watch an agent get registered, a decision get witnessed, and the chain get recomputed.
