Independent evidence for enterprise AI

Prove what your AI agents did.

Every interaction, decision and action your agents take, signed and sealed to an independent record as it happens. Evidence your auditors, regulators and insurers can verify at any point, without trusting you or us.

Request a demoRead the docs

For banks, insurers and other regulated industries. Your sensitive data never leaves your systems.

Credit underwriting agentRetail lending · production
7sealedRecording
  1. Application intakeAPP-4471 · personal loan · €48,000 · 60 months
  2. KYC and customer data retrievedIncome verification · 36 months of account historycontent stays in your systems
  3. AML and sanctions screening delegatedscreening agentNo sanctions or PEP match
  4. Credit bureau report retrievedBureau score 612 · 2 delinquencies in 24 months
  5. Credit risk scoredPD model v4.2.1 · probability of default 7.9%
  6. Application declined · reason code 04Debt-to-income 48%, above the 40% policy limit
Evidence trail recorded7 records · 2 agents · witnessed on an independent public recordVerified

A small price to pay for your customers’ trust.

As AI makes production nearly free, trust gets expensive.

Agents now decideThey approve credit, price risk and move money. Every action creates a question someone will later ask: can you prove what the system did?
The judgment is where the exposure sitsInputs can be deterministic. What the agent does with them is judgment, and judgment is what gets challenged.
The cost climbs with every systemBuilding AI is no longer the bottleneck. Being permitted to deploy it, and standing behind it, is.

One line of code puts any AI agent under evidence.

Verify the agents you already run, scale seamlessly. No new platform and no change to your models. Your raw data stays with you; only a signed commitment leaves.

In-house agents and AI servicesWhatever your teams built, on whatever models you choose.
Framework adaptersLangChain · Vercel AI SDK
SDKsJavaScript and TypeScript · Python
Log storageAmazon S3, or any S3-compatible bucket in your own estate.
Read the quickstart
import { createAnchorer } from "@sanning/anchor"

const anchorer = createAnchorer({
  apiKey, subject, signer, logStore,
})

// hashed and signed here; content stays in your systems
await anchorer.anchor({ content })
from sanning_anchor import Anchorer

anchorer = Anchorer(api_key=key, subject=agent,
                    signer=signer, log_store=store)

# hashed and signed here; content stays in your systems
anchorer.anchor(content=content)

When regulators or customers question your AI, what will you show them?

For most enterprises, the answer is self-attested logs their own teams can change. Hard to trust, easy to question, even when nothing went wrong.

01

Could you reconstruct every step behind a credit decision your agent made eighteen months ago?

02

Would your auditor accept a record your own engineers can edit?

03

What does one decision you cannot prove cost you: a lapse in customer trust, a blocked deployment, escalating audit costs?

Defending a decision without Sanning

Reconstruct, then argue

Log pulls, screenshots, interviews and sign-offs for every audit, model review, dispute and insurer questionnaire. And at the end, evidence the reviewer still discounts.

vs

With Sanning

Hand over the trail

The evidence already exists, complete and independently witnessed. Review starts from proof, not from reconstruction, and nobody has to take your word for it.

An external witness for every step your agents take.

Sanning captures every step your agents take. Each one is signed with the agent’s own key and anchored to an independent public record. Neither you nor Sanning can alter it afterwards without verification failing, and anyone can check.

01

CaptureHooks into the frameworks and tools your agents already use, so every step is recorded.

02

SignEach record is signed by the agent at source. Raw data stays in your environment; only its hash leaves.

03

AnchorSigned records are sealed on the public record at regular intervals, so any change to a record is evident.

04

VerifyEvidence packs anyone can verify independently, offline, with no call to Sanning.
Trail APP-4471credit-underwriting/4.2.1
  1. Captured
  2. Signed
  3. Anchored
  4. Verified
  1. session.openApplication intakeVerified
  2. data.readKYC and customer data retrievedVerified
  3. hand-offAML and sanctions screening · 2 signaturesVerified
  4. model.callCredit risk scoredVerified
  5. decision.emitApplication declined · reason code 04Verified
  • passed:Digest recomputed
  • passed:Signature
  • passed:Anchor reference
  • passed:Sequence

All 5 records verified · offline

$ npx @sanning/proof verify bundle.json --logs logs/

Hand your auditor a complete evidence pack.

For any agent, case or period, one command builds the evidence pack on your own systems: the records, their signatures and their seals on the public record. Your auditor verifies it on their own machine.

MinutesFrom request to packInstead of weeks of log pulls, screenshots and sign-offs.
YearsDefensible after the factThe record holds long after the model version is retired and the team has moved on.
ZeroTrust in enterprise logs requiredVerification runs offline, with no call to us.
OneNamed record on failureAny alteration breaks the proof and points to the exact step.

Evidence for the record-keeping these frameworks require

  • EU AI Act Art. 12
  • DORA RTS Art. 12
  • ISO/IEC 42001 A.6.2.8
  • ISO 27001 A.8.15
  • AIUC-1 E015
  • FDA 21 CFR Part 11
  • HIPAA §164.312(b)

Sanning proves the records were not altered. Your auditor and your supervisor decide what satisfies each requirement.

A record of your agents’ behaviors available for your customers at any point.

When every agent that touches a customer decision is under evidence, you can say so publicly. Customers, regulators, auditors and counterparties can inspect, analyse and challenge the record themselves, and none of it is self-attested. Sanning is the witness that removes the need to take your word for it.

Self-attested logs compared with a record under evidence with Sanning
Self-attested logsUnder evidence with Sanning
Who wrote the record
Your own systemsYour agents, signed with their own keys
Who can alter it
Anyone with database accessNo one, undetected. Not even Sanning
Who fixes the time
Your own serversAn independent public record
Who can check it
Whoever you choose to showAnyone you hand the trail to, offline
What a reviewer assumes
That you are telling the truthNothing
For your customersA decision they can question, and a record that answers.
For your supervisorsRecords you could not have altered, available on request.
For auditors, insurers and legalOne trail every reader can use, without collecting evidence by hand.

One evidence layer across the whole AI estate.

Every event across every agent joins the same trail, and the trail grows with your AI footprint.

  1. 01Agent actions
  2. 02Model decisions
  3. 03Tool calls
  4. 04Model inputs
  5. 05Cross-system hand-offs

One evidence layercaptured · signed · anchored · verifiable by anyone

Ready for your security review.

Sanning is built on open cryptographic standards, keeps your data inside your estate and produces evidence anyone can verify, with or without us.

Standards

Open, standard cryptography

Records are hashed and signed with published standards, then sealed on the public record. Nothing proprietary to take on trust.

RFC 8785 · SHA-256 · Ed25519 · RFC 9162

Data control

Your data never leaves

Content-blind by design. Prompts, outputs and customer data stay in your estate, logs sit in your own S3-compatible bucket, and you hold the signing keys.

hashes and signatures only

Retention

Tamper-evident retention

Each record is signed at source and sealed on the public record, so any change shows up on verification.

signed · sealed on the public record

Independence

No lock-in

Evidence packs verify offline, including in the browser. Your evidence holds even if Sanning doesn’t.

offline · no account

Assurance

Security and assurance

SOC 2 Type 2 is in progress.

SOC 2 Type 2 in progress

Integration

Fits your stack

One line of code and no change to your models. SDKs and adapters for the frameworks your teams already use.

SDKs · LangChain · Vercel AI SDK

Your data stays in your estate. Only signed proof leaves.

Sanning never needs to see what your agents saw. It witnesses hashes and signatures, not content.

Stays in your systems

  • Prompts and model outputs
  • Customer and case data
  • Tool inputs and results
  • Your signing keys

Leaves, per step

  • event hash
  • agent signature
  • time signed

Don’t ask your customers to trust you. Show them.

Reputation alone is no longer enough. See a live agent recorded, one record altered, and verification catch it.

Request a demo