The 7 sutras of AI governance are seven guiding principles in India’s AI policy framework: trust is the foundation, people first, innovation over restraint, fairness and equity, accountability, understandable by design, and safety, resilience and sustainability. For enterprises, the risk is that principles stay on paper. Runtime controls turn the sutras into enforceable decisions, while policy, oversight, and assurance govern the broader program. Aurascape maps each sutra to an inline control, so the work itself produces the evidence.

Last updated: September 2026.

Where Does the Seven-Sutra Framing Come From?

The seven-sutra framing comes out of India’s AI policy process. India’s AI Governance Guidelines describe a principle-based national approach organized around seven sutras, and recommend three new institutions to carry it forward: an AI Governance Group, a Technology and Policy Expert Committee, and an AI Safety Institute (Press Information Bureau, 2025).

The Reserve Bank of India’s FREE-AI committee report and India’s national AI governance guidelines both use the seven-sutra structure, the first scoped to financial services and the second written to apply across sectors.

Each sutra is a short directional statement, not a detailed rule. None of them names an enforcement point, a policy condition, or an evidence record.

The sutras overlap with recurring governance themes in the EU AI Act and the NIST AI Risk Management Framework: oversight, accountability, transparency, fairness, and safety. Shared themes do not make one control set sufficient for every framework, so map each obligation to its own control and its own evidence requirement. Treat the seven as a direction-setting layer and expect the enforcement detail to come from somewhere else.

The Seven Sutras, Defined

The 7 sutras of AI governance are seven guiding principles describing how AI should be built, adopted, and supervised: trust, human primacy, enablement, fairness, accountability, understandability, and safety. Each is stated as a value rather than a requirement, which is why translating them into controls is the hard part. The labels below are working summaries; confirm the verbatim official names against the primary documents before treating them as the framework’s exact list.

1. Trust is the foundation. Adoption depends on the confidence of the people affected by it, so trust is the precondition for everything else, not a downstream outcome. Inside an enterprise, trust is not a sentiment. It is the property that approved actions complete and unapproved ones do not. A program that flags an unauthorized tool call after the fact has already spent the trust it was built to protect.
2. People first. AI serves human interests and stays under human oversight, with a person able to intervene in consequential decisions. The tenet has a direct regulatory analogue: the EU AI Act requires high-risk systems to be designed so natural persons can effectively oversee them (EU AI Act, 2024). For delegated work, oversight means an approval step on a consequential action, such as a credit decision or a customer refund, not a dashboard reviewed weekly.
3. Innovation over restraint. The default posture is enablement with proportionate safeguards rather than prohibition. Blanket blocking is the failure mode this sutra targets, and it backfires measurably: 43% of workers admit sharing sensitive workplace information with AI tools without their employer’s knowledge (National Cybersecurity Alliance, 2025). Without a sanctioned path, employees may shift work to personal accounts with weaker enterprise oversight.
4. Fairness and equity. AI outcomes should not create or amplify discrimination, and the benefits should reach broadly. Data controls can keep protected attributes and sensitive records out of a decision. Outcome testing must then identify disparate effects across affected groups, with defined review thresholds, an escalation route, and a remediation path when a threshold is crossed. Input controls and outcome testing are two separate obligations, and a program that runs only the first has covered half of this sutra.
5. Accountability. Responsibility for AI outcomes is allocated to identifiable people and functions across development, deployment, and use. Allocation only holds if the record supports it. Naming an accountable executive in a policy document does little when nobody can reconstruct which account, which application, which dataset, and which decision were involved in a specific action last Thursday.
6. Understandable by design. Systems should be transparent and explainable in proportion to their impact, with disclosure to the people affected. Tool-call records explain what an agent did and which data moved. Decision-level explainability remains a separate requirement for consequential outcomes, along with telling an affected person that AI was involved and on what basis.
7. Safety, resilience, and sustainability. Systems should withstand misuse, failure, and adversarial pressure across their life cycle, and be sustainable to run. OWASP ranks prompt injection (LLM01), sensitive information disclosure (LLM02), and excessive agency (LLM06) among the top risks in its Top 10 for LLM Applications (OWASP, 2025). These risks appear during use, so design review has to be paired with runtime monitoring and enforcement. Sustainability adds the life-cycle dimension most programs skip: knowing which AI systems are genuinely in use, retiring the ones that are not, and keeping licensing and compute tied to real workloads rather than abandoned pilots.

The Pillars and Institutions Around the Sutras

Neither framework leaves the principles standing alone. FREE-AI pairs the seven sutras with six pillars, split between enablement (infrastructure, policy, capacity) and risk mitigation (governance, protection, assurance), and calls for board-approved AI policies that define risk appetite. The split is the instructive part. Three pillars exist to make adoption possible and three exist to keep it controlled, which rejects the idea that governance is only a brake.

India’s national guidelines add an institutional layer on top, with a governance group, an expert committee, and a safety institute. Read their official mandates from the primary document rather than inferring them. Enterprises can borrow the shape at their own scale without borrowing the mandates: a governance body that owns risk appetite and sanctioning decisions, a technical review group that assesses systems before production, and an assurance function that tests behavior after deployment, when the traffic is real.

The pillars also expose a sequencing problem worth planning around. A team can stand up capacity and policy in a quarter. Protection and assurance depend on telemetry that many organizations do not yet collect, because AI interactions run over streaming and conversational protocols, and controls built for web and SaaS traffic commonly evaluate them at the destination level rather than decoding the exchange.

Why Principles Stall Before They Reach Production

The gap between a stated principle and an enforced control is measurable. 90% of organizations say employees use AI tools, while only 38% have a formal, comprehensive AI policy and 25% have none at all (ISACA, 2026). Writing a policy is the easy half, and most organizations have not finished it.

Where controls are thin, projects tend to die quietly rather than fail loudly. Gartner predicts over 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, or inadequate risk controls (Gartner, 2025). Governance failure often shows up on the adoption side of the ledger, as a program that never left pilot.

Discovery is the prerequisite no principle statement mentions. 82% of organizations have unknown AI agents in their environments (Cloud Security Alliance, 2026). A reliable inventory is what lets a governance team assign owners, controls, and evidence requirements to each AI system in use, including the unsanctioned coding assistant and the locally launched agent.

The World Economic Forum, 2026 describes what agents do when governance is weak:

“[Agents] can accumulate excessive privileges, be manipulated through design flaws or prompt injections, or inadvertently propagate errors and vulnerabilities at scale.” World Economic Forum, Global Cybersecurity Outlook 2026

The report’s prescription is continuous verification, audit trails, and accountability structures grounded in zero-trust principles that treat every interaction as untrusted by default. Those are runtime properties, and a signed policy produces none of them on its own.

How Can Enterprises Enforce the Seven Sutras?

Every sutra can be restated as a question about a specific moment: an employee pastes a contract into a chat app, a coding assistant streams a private repository, an agent calls a ticketing tool and writes the result somewhere else. Answer the question at that moment and the principle is enforced. Answer it only in a quarterly review and it is merely documented.

The control set also expands as the work changes. Human-to-AI use, where employees prompt tools directly, needs discovery, account control, and data protection. Human-to-agent delegation adds scoped permissions and approval on consequential actions. Agent-to-agent execution adds governance of the tool calls themselves. The table below maps each sutra to the enforcement point that makes it real.

Discovery comes before all of it. Aurascape discovers AI applications, accounts, and agents across the network, endpoint, and API planes, detects AI running locally on laptops and servers, and detects an agent launch and its configuration, including MCP server connections, before the agent takes its first action (Aurascape, 2026). That matters because the population will not hold still: Aurascape has detected that 50+ new AI apps release every day, and coverage reaches production within 48 hours for the vast majority of newly observed applications.

On the agent side, Aurascape discovers and secures local AI agents and their interactions, and adds a Zero-Bypass MCP Gateway that marks every tool call it approves and blocks unmarked calls, governing the agent-to-tool execution path inline rather than observing it (Aurascape, 2026). Model Context Protocol (MCP) is one common tool-execution pattern, not the whole agent access-control problem, so governance also depends on discovery, interaction context, and data policy across the wider workflow. On the paths the Gateway governs, the Gateway marks the tool execution channel and the AI Proxy watches the model channel and checks for that mark, so a call that skipped policy surfaces there and is stopped before the model acts on the result. High-risk write and execute calls can be held for human confirmation, tool definitions are pinned at sanction so a server-side change requires explicit review, and tool results are inspected before the agent acts on them.

Data enforcement runs on the same interaction. Aurascape applies 600+ real-time data classifiers, extended by customer-defined patterns and fingerprinting for data only the organization can name (Aurascape, 2026). Policy actions include allow, coach, notify, redact, redirect, and block. A governance team can then address the risky interaction without switching off the application the business asked for.

Attestation Documents Versus Interaction Evidence

Most AI governance programs produce attestation: a policy, a register of approved systems, a signed acknowledgment, a risk assessment refreshed on a cycle. Attestation answers what the organization intended. An auditor testing the accountability sutra asks a different question, and it usually starts with “show me.” Only a record created at the moment of the interaction answers that one.

A management system gives the principles structure. ISO/IEC 42001 specifies requirements for establishing and running an AI management system (ISO, 2023), and the NIST AI Risk Management Framework organizes the work into Govern, Map, Measure, and Manage (NIST, 2023). Each supplies its own management and risk structure alongside the sutras, rather than one framework sitting above another.

A workable readiness sequence for a governance team:

1. Discover before you draft. Enumerate the AI applications, account types, coding assistants, copilots, and agents actually in use, including the ones nobody registered.
2. Assign each sutra an owner and one metric a system can produce without a survey.
3. Convert each principle into a policy statement that names an enforcement action, not a sentiment.
4. Instrument evidence at the interaction, so the audit artifact falls out of the control instead of becoming a separate project.
5. Set a cadence for the long tail. The application population changes weekly, and a static allowlist goes stale fast.

Sector obligations sit on top of the same control set. Aurascape publishes mappings for enterprise AI compliance frameworks, for banks and investment firms, for financial services, and for healthcare and pharmaceutical organizations. Each sector maps the principles to its own obligations, workflows, controls, and evidence requirements.

Frequently Asked Questions

What are the 7 sutras of AI governance?

They are the seven principles defined above: trust, human oversight, balanced innovation, fairness, accountability, transparency, and safety. They anchor India’s AI Governance Guidelines, a principle-based national framework that also recommends new institutions to put the tenets into practice.

Who created the seven sutras?

The framing comes from India’s AI policy process. India’s AI Governance Guidelines present seven sutras as a national, principle-based framework, and the Reserve Bank of India’s FREE-AI committee report uses the same seven-sutra structure for financial services.

Are the seven sutras legally binding?

They are published as guidance principles, not as statute. Separate binding obligations, such as those in the EU AI Act and applicable sector rules, can require similar controls in practice, and that is a distinct legal basis. Confirm the status that applies to your organization with counsel.

How do the sutras relate to ISO/IEC 42001 and the NIST AI Risk Management Framework?

The sutras provide policy principles, while ISO/IEC 42001 and the NIST AI Risk Management Framework provide separate management and risk structures. Use the sutras to set direction, a management standard to structure the program, and runtime controls to produce evidence.

What is the difference between a sutra and a control?

A sutra states an intent; a control changes what happens. “Trust is the foundation” is a sutra. Stopping an unapproved tool call at the point it is attempted, rather than logging it afterward, is a control.

Which sutra is hardest to operationalize?

Accountability gets difficult when records cannot connect an action to the user, application, tool, data, and policy decision involved. Fairness is a close second, because it needs both input controls and outcome testing.

How does Aurascape support the seven sutras?

By enforcing them where AI is used and recording the result. Policy applies during the interaction, and the same interaction produces the record an auditor asks for.

Aurascape enforces governance policy where AI is used and produces the evidence from the same interaction. Discovery establishes scope, data classification identifies risk, and inline controls apply the policy decision. That turns each sutra into something a governance team can test, enforce, and evidence. Bring your own obligations and see them applied to live AI traffic.

See how Aurascape enforces AI governance principles at the point of use →

GARTNER® is a registered trademark and service mark of Gartner, Inc. and/or its affiliates and is used herein with permission. All rights reserved. Gartner does not endorse any vendor, product or service depicted in its research publications and does not advise technology users to select only those vendors with the highest ratings or other designation. Gartner research publications consist of the opinions of Gartner’s research organization and should not be construed as statements of fact.