ARBOR GOVERN

Every AI system you run, on one register.

Teams ask to use a model, a vendor tool or an agent, and the request lands in one place. It gets a risk tier under the rules that apply to you, the assessment that tier calls for, and a person who signs it off. Every step is written to a ledger that an auditor can check without taking our word for anything.

  • EU, US and India rules side by side
  • Signed, hash-chained evidence
  • Seats unlimited

Governance, not runtime security: blocking data on its way to a model is Arbor Guard. Built to run in a region you choose, India, the EU or the US, or inside your own Kubernetes cluster.

In plain terms

When a regulator, an auditor or a customer asks which AI you use and how you checked it, the answer is a file you export, not a fortnight of spreadsheets.

What happens

From a request to use AI to a record you can defend

Numbered because the order is real. A system has no tier until it is registered, no assessment until it has a tier, and no assessment counts until a person has signed it off.

  1. 01 · REGISTER

    Somebody asks to use AI

    An intake form goes to triage, approval or rejection, with a clock on it. Systems, models, datasets, agents and the vendors behind them all sit on one register, each with an owner, a lifecycle stage and a model card you can export as a PDF.

  2. 02 · TIER

    The tier is worked out, not guessed

    Risk tiers are computed from the intake answers under the EU AI Act, RBI materiality, SEBI and the MeitY guidelines. The rules are content rather than code, so a rule change is a data update and every tier shows the answers that produced it.

  3. 03 · ASSESS

    The assessment the tier calls for

    An AI impact assessment, a fundamental rights impact assessment, a DPDP Rule 13 assessment, an ISO/IEC 42001 gap by clause. AI drafts answers with citations and a provenance flag, and a person accepts every one of them.

  4. 04 · EVIDENCE

    Controls, policies and proof in one place

    Results map to controls through one crosswalk. Evidence files are stored by their hash, linked to the controls they support, given an expiry and a reviewer, and labelled by how they arrived: collected, captured, self-reported or attested.

  5. 05 · PROVE

    An export that stands on its own

    Auditor packs, board reports and a Statement of Applicability, each drawn from the same ledger. The ledger verifies offline with a public key, so the person checking it needs neither our database nor our help.

The rules it knows

One product, a pack for each region

22 regulatory packs ship today, and India sits beside the EU and the US as a peer rather than an afterthought. Packs and tier rules are content, not code, so a new instrument is a data update. Where an instrument is a draft, or is read by clause number only, it says so.

European Union and UK

  • EU AI Act
  • GDPR
  • UK Data (Use and Access) Act

United States

  • NIST AI RMF
  • NIST AI 600-1
  • California CPPA rules on automated decisions
  • Texas TRAIGA
  • Illinois HB 3773
  • Connecticut SB 5
  • NYC Local Law 144
  • Utah AI Policy Act
  • Washington AI chatbot law
  • OMB M-25-21

India

  • DPDP Act and Rules
  • MeitY AI governance guidelines voluntary
  • RBI FREE-AI
  • RBI model risk management draft
  • SEBI AI and ML
  • IT Rules 2026

Standards

  • ISO/IEC 42001 by clause number
  • ISO/IEC 27001
  • SOC 2

Assessments that ship as templates

  • AI impact assessment
  • Fundamental rights impact assessment (EU AI Act, Article 27)
  • EU deployer duties (Article 26)
  • GDPR Article 22
  • ISO/IEC 42001 gap, by clause
  • NIST AI RMF
  • California risk assessment and automated decisions
  • Employment AI (NYC, Illinois, Connecticut)
  • US state and federal
  • DPDP Rule 13(3)
  • MeitY self-assessment
  • RBI model validation
  • SEBI AI and ML checklist
Around the register

The rest of the programme, in the same place

  • Vendor AI

    A questionnaire your vendors answer in a portal with no login, attestations they upload, a map of which language models sit behind each sub-processor, and an assessment for the foundation model itself.

  • Policies and people

    A policy register with versions, hashes and staff acknowledgement, short AI literacy modules, an incident register, and an appeal route for a person affected by an automated decision.

  • Testing and oversight

    A test harness that writes one signed record per run, kill-switch drills, a CI/CD approval gate written in OPA Rego, and tool permissions and run audits for agents.

  • Transparency

    A chatbot disclosure snippet, a labelling kit for synthetic content declarations under the IT Rules 2026, an output register and a public transparency page.

  • The tools you already run

    Two-way Jira and ServiceNow, a sync to the GRC system you already have, single sign-on, and an API with webhooks and SIEM export.

  • Consent, if you run Sammati

    When an AI system relies on personal data, the register reads consent coverage from the Sammati consent manager, and a withdrawal or an erasure reaches the systems that used it. Everything else works with this switched off.

The evidence ledger

Evidence your auditor checks without trusting us

The same discipline as the Sammati consent ledger, because it answers the same question: not whether a record exists, but whether anyone could have changed it since.

  • Each entry carries the one before it

    Every entry stores a SHA-256 hash of its own content joined to the previous hash, and is signed with ECDSA P-256. Change one entry and every hash after it stops matching.

  • A signed root every day

    A cumulative Merkle root is computed and signed daily, so a single entry can be proved present with an inclusion proof rather than by handing over the whole chain.

  • A copy nobody can quietly rewrite

    Entries and roots are mirrored to write-once object storage in a separate bucket, with a separate key and chain in each region.

  • A break is reported, never repaired

    A broken chain is quarantined and raises an alert. Nothing smooths it over, because a ledger that repairs itself is a ledger nobody can rely on.

What it does not do yet

Stated plainly, so nobody demos something that does not exist

A governance buyer reads a product page the way an auditor reads evidence. These are the gaps you would otherwise find in week three of an evaluation.

Regions are code, not yet a customer region

India, EU and US stacks are infrastructure code that has been validated, and has not yet carried customer data. A workspace is pinned to its region when it is created, and the first customer in a region is the first deployment there.

Cloud AI connectors have met published APIs only

The connectors for Bedrock, Azure AI Foundry, Vertex, SageMaker, Databricks and the rest are written against each provider's published API and have not yet been run against a live account.

Test-tool adapters read documented output

The harness adapters for garak, PyRIT, promptfoo, Ragas, DeepEval and Evidently are tested on each tool's documented output, not on live runs. Harness signing keys are files today, not keys held in a KMS.

Two Indian instruments are not final

The RBI model risk management framework is a draft, and the SEBI pack is still being checked against the primary text of the circular. Both are marked as such wherever they appear.

Consent linkage is partly built

Coverage, withdrawal and erasure propagation work. Purpose tags and AI notice clauses have been tested against the Sammati consent manager's contract, not yet against the live service.

We hold no certification, and grant none

Arborworld holds no ISO/IEC 42001, ISO/IEC 27001 or SOC 2 certificate today. The product prepares the gap assessment and the evidence. Certification is a decision for an accredited body.

How it is bought

Per AI system, not per seat

Priced per AI system under governance, with seats unlimited, and quoted on request. Self-hosting inside your own cluster is available on the larger licences.

Questions we get asked
Is Arbor Govern an AI security tool?

Arbor Govern is a governance product, not a runtime security product. It records which AI you use, what risk it carries, how it was assessed and who approved it. Stopping personal data on its way into a model is the job of Arbor Guard, a separate product managed from the same console.

Which laws and frameworks does it cover?

Arbor Govern ships packs for the EU AI Act, GDPR and the UK Data (Use and Access) Act; NIST AI RMF and the US state laws on automated decisions, hiring and chatbots; and India's DPDP Act, the MeitY guidelines, RBI and SEBI. The RBI model risk framework is a draft and is marked as one.

Can an auditor check the evidence without trusting Arborworld?

An auditor can verify the evidence ledger offline with a public key and no access to our database. Every entry is hash-chained and signed with ECDSA P-256, a signed Merkle root is produced daily, and an inclusion proof shows a single entry is present without exposing the rest.

Does Arbor Govern need the Sammati consent manager?

Arbor Govern works fully on its own. Customers who also run the Sammati consent manager can link the two, so the register shows consent coverage for an AI system and a withdrawal reaches the systems that relied on it.

Does it reproduce ISO/IEC 42001?

The ISO/IEC 42001 gap assessment works by clause number and Annex A identifier with our own short paraphrase. The standard itself is copyright, so you need your own licensed copy to read the text.

Does AI write the assessments?

AI drafts answers in an assessment, each with its citations and a flag showing it was drafted. A person accepts or rewrites every answer, and a drafted answer does not count until a person has accepted it.

Last reviewed . General information, not legal advice. Arbor Govern is a separate product from Arbor Guard and from the Sammati consent manager, and every product is compared here.

Where to start

Bring us the AI system nobody has written down.

A walkthrough starts with one system you already run: register it, tier it and draft its first assessment, so you see the evidence it produces before you decide anything.