FairMind
Request access
Menu

0items approved by you24,000 checks in the run · 0 changed without you

Agentic AI governance

Governance that runs itself.
Decisions that stay human.

FairMind runs governance as agent workflows: collect, check, suggest, act, verify, file. Every workflow stops at a gate, and a person decides. Right now, that person is you.

Request access How a run works

What you just did

Every tile is a check in a run. It stayed grey until you looked at it.

Deny by default. Nothing is recorded or changed until an approver decides, and every decision is kept with who made it and why.

Behind you

Your decisions become the evidence trail.

A hash-chained trace of inputs, model calls, tool calls, and approvals, mapped to the requirements it satisfies. Auditors read the trail; they don't take our word for it.

51Frameworks in the registry, EU AI Act to OWASP Agentic
90Governance workflows agents can run
109Evaluations mapped to requirement IDs
21Packs across Govern, Evaluate, Secure, Comply
0External changes made without a human approval

How a run works

The agent stops at you. Nothing moves until you decide.

  1. 01

    CollectPolicies, tickets, model cards, logs. Read-only scopes.Agent

  2. 02

    CheckRules first, model second. Every finding cites its source.Agent

  3. 03

    SuggestThe record, mapping, or fix, with its confidence stated.Agent

  4. 04 · Gate

    YouApprove, edit, or reject. Deny by default.Waiting on a person

  5. 05

    ActDry run, blast radius, rollback plan. Then the change.Agent

  6. 06

    VerifyRe-read the target. A mismatch rolls back and escalates.Agent

  7. 07

    File evidenceHash-chained trace, mapped to every requirement it satisfies.Agent

The agent suggests mapping this fairness run to EU AI Act Article 10(2)(f).

Autonomy

You set how far the agent goes. Per workflow.

Internal audit stays at Assist. Inventory upkeep can run at Act. External systems are never written without a person.

A0ManualThe agent does nothing unless asked for a draft.
A1AssistCollects and checks. Humans write every conclusion.
A2ProposeProposes; nothing is recorded until a human approves. The default.
A3Act on internal recordsFiles records inside FairMind; humans review by sampling. Never external writes.

Four pillars

Install the packs your systems need.

Govern

Run the management system

Policies, accountability, inventory, risk and impact, data, suppliers.

Evaluate

Test before and after release

Quality, fairness and bias, agent behaviour, production drift.

Secure

Protect LLM apps and agents

OWASP LLM and Agentic Top 10, MITRE ATLAS, supply chain.

Comply

Show it, requirement by requirement

EU AI Act, ISO/IEC 42001, AIUC-1. Gaps shown, not hidden.

109 evaluations

Tests for the systems you actually run. Thresholds declared first.

Each evaluation names its metric, its threshold and where that threshold comes from. Each evaluator says whether it is validated, external, metadata only, or not yet available.

Agents

Can it only do what it was allowed to?

Tool least privilege, out-of-scope actions blocked, MCP server permission audit, excessive agency, kill-switch time to stop, runaway containment.

LLM apps and RAG

Does it hold under attack?

Direct and indirect prompt injection, jailbreak resistance, system prompt leakage, groundedness, harmful content, over-refusal, knowledge-base poisoning.

Models and data

Is it still the model you tested?

Task accuracy, calibration, robustness, data drift, out-of-distribution detection, membership inference, backdoor detection, label quality.

Decision systems

Is it fair, with the statistics shown?

Selection-rate parity at the four-fifths rule, error-rate parity, intersectional slices, fairness drift. Suites for hiring, credit, insurance, benefits and education.

The agent is governed too

Twenty controls on our own agent.

Per-step scoped identity. A deterministic tool allowlist prompts cannot widen. Collected content treated as data, never instructions. A hash-chained trace of every call. A kill switch per run. You can see our agent's metrics and incidents; we cannot hide them.

run 7f3a…c21e workflow deployment_release_gate@v4
collect 14 items · scopes: read
check 11 pass · 2 fail · 1 cannot conclude
suggest finding RSK-0412, cites item 9 §3
ask approver: priya.n · decision: approve with edits
act dry-run ok · blast radius 1 record
verify target re-read · match
file evidence → ISO42001 6.1.2, EUAIA Art.9, NISTRMF GOVERN 1.7
trace sha256 e1b0…9a44 · chained to previous run

Questions teams ask

What is agentic AI governance?

Agentic AI governance uses AI agents to run the repeatable parts of governing AI systems: collecting evidence, checking it against framework requirements, proposing records and fixes, and filing the trace. The decisions stay with people. In FairMind, every workflow stops at an approval gate before anything is recorded or changed.

What does a FairMind agent actually do?

It runs a workflow made of seven step types: collect, check, suggest, ask a person, act, verify, and file evidence. Today the trigger, escalate and file steps run themselves, the gate steps open an approval request, and a person completes the collect, check, suggest, act and verify steps, which the ledger records as manual. Agent-executed steps arrive workflow by workflow, behind the same gates, with scoped read-only credentials, rules before model judgement, and a citation for every finding.

What does FairMind evaluate?

The catalogue defines 109 evaluations: 53 for security, 46 for behaviour and fairness, and 10 for governance and compliance. They cover agents (tool least privilege, out-of-scope actions, MCP permissions, kill switches), LLM apps and RAG (prompt injection, jailbreaks, prompt leakage, groundedness, harmful content), models and data (accuracy, calibration, drift, membership inference, backdoors) and decision systems (selection-rate parity, error-rate parity, intersectional slices). Each one names its metric, its threshold and where that threshold comes from, and every evaluator declares whether it is validated, runs through an external provider, reads metadata only, or is not yet available.

Can the agent change things on its own?

Only inside FairMind, and only if you set that workflow to autonomy level A3. The default is A2: the agent proposes and a person approves. External systems are never written without an approval. If an approver is unreachable or no policy covers the action, nothing happens.

Which frameworks does FairMind cover?

The registry holds 51 frameworks, including the EU AI Act, ISO/IEC 42001, NIST AI RMF and AI 600-1, the OWASP Top 10 for LLM and Agentic applications, AIUC-1, MITRE ATLAS, the OWASP MCP Top 10, CSA MAESTRO, Singapore’s Model AI Governance Framework, and national rules from India, Japan, Korea, China and US states. One workflow run counts toward every requirement it maps to.

How is FairMind’s own agent governed?

Twenty runtime controls apply to it: per-step scoped identity, a deterministic tool allowlist that prompts cannot widen, collected content treated as data rather than instructions, mandatory citations, pinned model and prompt versions with golden-set evaluations, a hash-chained immutable trace, cost and rate caps, and a kill switch per run. Customers can see the agent’s metrics and incidents.

Is FairMind a certification?

No. FairMind produces and organises evidence and records human decisions. Certification and conformity decisions stay with your reviewers, auditors, and, where relevant, accredited bodies.

Let the agents do the paperwork. Keep the decisions.

Request access