drshta.ai

Compliance by Construction

Autonomy your auditor can verify.

Every consequential agent action leaves a signed receipt: what evidence it saw, which policy allowed it, who approved it and what changed. Agents gain autonomy one risk tier at a time as that evidence is accepted.

Receipts follow IETF SCITT, the open standard for signed, verifiable statements.

See Demo How it works
Action receiptIllustrative

Disable 14 dormant privileged accounts

Subject
Access-review agent · run R-0117
What evidence?
14 account records, inactive 90+ dayssha256 3f9a41e0…c07e
What policy?
Privileged access, v12 · NIST AC-2(3)decision: approval required
Who approved?
Reviewer RV-07, IT Risk · 14:31:40on evidence 3f9a41e0…
What changed?
14 accounts disabledrecord 9b2c07d1… → 51ef8a30…
Registration
Transparency log · entry 48,213statement sha256 a41c9e02…77d0signature ES256 30450221…f3b9

Four questions

The auditor sets the limit.

An agent's autonomy is as wide as the evidence behind it. Each receipt answers the four questions an auditor asks of any consequential change.

  • What evidence?

    A hash of every record the agent read, anchored in a tamper-evident ledger.

  • What policy?

    The policy version, the rule that decided and the framework controls it maps to.

  • Who approved?

    A named reviewer, their decision and the time, on the evidence they saw.

  • What changed?

    The state of the target record before and after the action.

How it works

Two paths: one for the action, one for the proof.

The action path decides whether an agent may act. The evidence path records every step as it happens and turns it into a receipt anyone can check.

Action path and evidence path Action path: an agent proposes, the policy gate checks it, actions above the risk tier wait for approval, then the system of record changes. Each step is written to the evidence ledger, which produces a receipt; the open verifier checks the receipt. ACTION PATH EVIDENCE PATH Agent proposes with its evidence Policy gate every action Approval above the risk tier System of record change applied Hash-chained evidence ledger · Merkle proofs Receipt signed · SCITT Open verifier checks offline Verified Action path and evidence path Action path: an agent proposes, the policy gate checks it, actions above the risk tier wait for approval, then the system of record changes. Each step is written to the evidence ledger, which produces a receipt; the open verifier checks the receipt. ACTION PATH Agent proposes with its evidence Policy gate every action Approval above the risk tier System of record change applied Evidence ledger EVIDENCE PATH Receipt signed · SCITT Open verifier checks offline Verified
Grey: the action path. Blue: the evidence path. Gold marks a receipt that verified.

Assured Autonomy Levels

Autonomy, earned one level at a time.

Five levels, from advice to continuous assurance. As an agent's freedom grows, so does the evidence it must carry.

  1. A0

    Advise

    AI recommends. A person decides.

    Evidence barEach recommendation cites its evidence

  2. A1

    Act, low risk

    AI acts on low-risk actions. People spot-check afterwards.

    Evidence barFull receipt for every action

  3. A2

    Act with approval

    AI acts after a named person approves, on the evidence.

    Evidence barApproval and effect recorded

  4. A3

    Act, audited

    AI acts on its own within a defined kind of action.

    Evidence barThat action's receipts passed audit for N cycles

  5. A4

    Continuous assurance

    Broader autonomy, because controls, evidence and monitoring justify it.

    Evidence barThe auditor accepts receipts as primary evidence

Levels apply per kind of action. One client can let agents close duplicate tickets on their own (A3) while agents only recommend account disablements (A0). A kind of action moves up one level only after the auditor accepts the evidence from the level below.

The receipt

An open receipt, built on IETF SCITT.

  • Open format

    Each receipt is a signed statement registered with a transparency service under IETF SCITT (RFC 9943). The profile is open and vendor-neutral.

  • Hashes only

    Receipts carry hashes and identifiers. Personal data stays in your source systems.

  • Complete or invalid

    Every field is required. A receipt with a missing field fails verification.

  • Verifiable offline

    The open verifier is designed to check the signature, the required fields and the inclusion proof on its own, without Drshta.

Receipt field groups
Field groupThe question it answers
SubjectWho acted?
ActionWhat was attempted?
EvidenceOn what basis?
Policy decisionWas it allowed?
ApprovalWho approved, on current evidence?
EffectWhat changed?
GroundingIs the explanation backed by evidence?
RegistrationCan it be verified independently of the vendor?

By industry

Same receipt, each regulator's question.

  • Banking & financial services

    Infrastructure and CMDB changes

    “Who approved this production change, independently of who made it, and on what evidence?”

    Suggested start: A2 on standard changes, then A3

    DORA RTS Art. 17DORA RTS Art. 12

  • Healthcare & life sciences

    Regulated record updates

    “Show the audit trail for every create, modify or delete on this record.”

    Suggested start: A1 on intake, A2 on record changes

    21 CFR 11.10(e)

  • Health insurers

    Prior-authorisation triage

    “What was the specific reason for this denial, and was it decided in time?”

    Suggested start: A0 on denials, A2 on approvals

    CMS-0057-F

  • Retail

    Agent-initiated purchases and refunds

    “Did the customer authorise this, and does the action match the mandate?”

    Suggested start: A1 on refunds under a threshold

    AP2 mandates

  • Media

    Release of AI-generated content

    “Which model and policy produced this asset, was it marked, and who approved its release?”

    Suggested start: A2 with human release approval

    EU AI Act Art. 50C2PA 2.4

  • Telecom

    Network remediation

    “At what autonomy level did the network change itself, and what gated it?”

    Suggested start: A1 on reroutes, A2 on core changes. A0–A4 maps onto the L0–L5 scale telecom teams already use.

    TM Forum Autonomous Networks L0–L5

Why it pays

Three returns from one receipt.

Automation gained per level

Each level a kind of action climbs moves more work from people to agents, with your risk function's sign-off.

Audit testing no longer re-performed

Once the auditor accepts receipts as primary evidence, they sample receipts in place of re-running control tests.

Outcome-billing disputes avoided

On outcome-priced work, a verifiable receipt settles what was done before it becomes a dispute.

Company

A public benefit corporation.

Drshta.ai PBC builds governance infrastructure for AI agents, wherever automated decisions have to stand up to an auditor.

To run Drshta on your own workflows, email hello@drshta.ai.

Drshta on LinkedIn