AI Governance | Audit, Compliance & Model-Risk for Production AI

The Stakes

What ungoverned AI actually costs.

An AI model is an engine. Governance is the rest of the car — the brakes, the dashboard, the service log and a named driver. Without them you don't have a vehicle, you have a fast object with no way to stop it. These are the failures that happen, and none of them are hypothetical.

The runaway agent

An agent with refund access wrongly issues 500 refunds overnight. The code compiled fine — the unchecked authority was the problem.

Where it's caughtPhase 3 · Design & architecture — least-privilege tool permissions and a human interrupt on irreversible actions, decided on paper before build.

The controlA tool-permission matrix scoping exactly what the agent may do, plus a hard human-approval step on any irreversible action.

The evidenceThe Architecture Decision Record (ADR) and the tool-permission matrix, in the Phase 3 evidence pack.

Open Phase 3 · Design →

Prompt injection & leakage

Untrusted input hijacks the model or exfiltrates data — the OWASP LLM Top 10 threats a functional test never sees.

Where it's caughtPhase 5 · Evaluate & test — adversarial and red-team evaluation for exactly the OWASP LLM Top 10 a functional test can't reach.

The controlInput/output guardrails set in design, then proven under attack with tools like garak and PyRIT before the release gate.

The evidenceThe red-team report and adversarial eval results, attached to the Phase 5 evidence pack the release gate relies on.

Open Phase 5 · Evaluate →

Silent quality drift

A system accurate at launch quietly degrades in production. Without monitoring, no one notices until a customer does.

Where it's caughtPhase 8 · Monitor & observe — continuous, sampled online evals plus drift and abuse monitoring, with a human-feedback loop.

The controlLive evals against the launch bar, drift alerts, and traces on every interaction — because traces, not code, are the record.

The evidenceMonitoring dashboards, drift reports and the next-review dates carried in the Living Register.

Open Phase 8 · Monitor →

Unaccounted data

PII indexed into a RAG corpus with no lawful basis and no provenance — ungoverned data makes everything downstream ungovernable.

Where it's caughtPhase 2 · Data & knowledge — every dataset and RAG source accounted for, with provenance, PII handling and lawful basis established.

The controlA data inventory with recorded provenance, PII detection and redaction, and a documented lawful basis before anything is indexed.

The evidenceThe data-lineage record and, where personal data is involved, the DPIA — both in the Phase 2 evidence pack.

Open Phase 2 · Data →

The Focaloid Framework

Nine phases across Map, Measure, Manage & Govern.

Governance is a thread, not a final gate. It starts before anyone writes code and doesn't stop until the system is decommissioned. Govern runs across all nine phases; Map, Measure and Manage own their stretch of the build. And it's a loop — Phase 9 feeds its findings straight back into Phase 1.

Map · context & riskMeasure · prove itManage · ship, watch, respondGovern · across every phase

Map · context & risk

Measure · prove it

Manage · ship, watch, respond

GOVERNAI inventory · RACI ownership · go/no-go gates · evidence packs — across every phase.

↻ It's a loop — Phase 9 feeds its findings back into Phase 1.

Eight gates — a documented go / no-go between every phase. Gates are cheap. Drift is expensive. A gate is a five-minute documented decision — not a committee.

All nine phases at a glance ▾

One Method · Three Frameworks

Evidence produced once counts everywhere.

The lifecycle is mapped to each framework, so a single body of evidence satisfies all three. No parallel compliance projects, no second team producing the same artifacts in a different format for a different auditor.

Function	EU AI Act	NIST AI RMF	ISO/IEC 42001	Phases
Map	Risk classification · data governance (Art 10) · DPIA / fundamental-rights assessment	Map — context, categorisation, risk identification	Cl. 4–6 — context, AI risk & impact assessment	Phases 1–3
Measure	Accuracy, robustness & cybersecurity (Art 15) · technical documentation (Art 11)	Measure — test, evaluate, benchmark, track	Cl. 8 — operational controls; Annex A controls	Phases 4–5
Manage	Conformity assessment · human oversight (Art 14) · post-market monitoring (Art 72)	Manage — respond, monitor, recover	Cl. 9–10 — performance evaluation & improvement	Phases 6–9
Govern	Quality-management system (Art 17) · accountability, roles & registration	Govern — culture, policy, accountability	Cl. 5 — leadership & AI policy (the AIMS itself)	All phases

Which applies to me?Where you operate What the AI decides

The Composed Stack

No platform closes the seam. Our phase packs do.

The tooling market is split. GRC platforms govern from the policy side. MLOps platforms govern from the model side. They meet — badly — at the Design → Build → Deploy seam, and nobody has closed it. That gap is where ungoverned AI actually ships.

GRC / Compliance-led

OneTrustCredo AIHolistic AISecuritiServiceNowVerifyWise

Strong on policy — stops at the build.

The seam

design →
build →
deploy

Model-lifecycle / MLOps

watsonx.governanceModelOpMLflowSageMakerVertex AIAzure ML

Strong on models — thin on policy.

Runtime / GatewayLiteLLMDatabricks Unity AI GatewayBifrostNeMo GuardrailsOpenFGA

Observability / EvalsLangfuseArize PhoenixTruLensLangSmithEvidentlyRagasdeepevalThese run the system — but none of them render a verdict.

Focaloid · Nine artifact packs — the connective tissue

A per-phase .docx Evidence Pack + .xlsx Living Register, joined by one AI System ID.

Where tools report numbers, we render a signed governance verdict — spanning the Design → Build → Deploy seam that GRC and MLOps platforms each stop short of.

Brand examples are illustrative, not endorsements — we compose to whatever stack you already run.

Our governance framework

The Governed-AI Lifecycle — our proprietary nine-phase framework.

Every AI system we govern moves through nine phases, each with a governance gate at the decision points and each mapped to the NIST AI RMF. Filter by NIST function or by phase to see how oversight is applied end to end.

Function

Phase

Behind every phase sits a curated toolchain — 40+ open-source and commercial controls for privacy, data lineage, guardrails, evaluation, red-teaming and production observability. We map the specific tools and gate criteria to your stack in a governance review.

On the NIST mapping: the framework maps phases 1–3 to Map, phase 5 to Measure, phases 7–9 to Manage, and Govern runs across all nine — Build (4) and Review (6) are the build step and the go-live gate. Filter above to see each function in isolation.

The Deliverable

A signed verdict — not a dashboard of numbers.

Observability tools tell you what happened. Neither a trace nor a drift chart is an answer to “show me this system was approved, by whom, against what bar.” That answer is an artifact, and it’s what we hand over.

.docx

The Evidence Pack

One per phase, signable. Model card, eval and red-team results, data lineage, threat model, the DPIA where personal data is involved, the go/no-go decision and any conditions attached to it. The document a client’s risk team or a regulator can read end to end.

.xlsx

The Living Register

The AI inventory that stays current. Every system, its risk tier, its owner, its gate history, its residual risk, its review dates. The thing an auditor asks for first and most organisations cannot produce.

AI System ID · FCL-AI-0142

Everything joins on one identifier.

Each system gets a single AI System ID, and every artifact, gate decision, eval run, trace and register row hangs off it. That’s what turns nine separate phases into one auditable thread — and it’s the specific thing no GRC platform and no MLOps platform gives you, because each of them only sees half the lifecycle.

Evidence Pack — Phase 6 · Review & sign-off .docx

SAMPLE

Evidence Pack · Review & sign-off

AI System ID: FCL-AI-0142 · Claims Triage Copilot · rev 3

Model card

Purpose	Triage inbound claims & draft a first response
Owner	Governance, Risk & Legal — J. Meyer
Risk tier	High (EU AI Act, Annex III)
Model	Claude Sonnet 5 + retrieval, guardrailed

Evaluation summary

Metric	Result	Bar	Pass
Groundedness	94%	≥ 90%	✓
Prompt-injection resistance	98%	≥ 95%	✓
Toxicity rate	0.08%	< 0.5%	✓

Decision: GO — with conditions

Reviewer: A. Okafor (independent) · Risk & Compliance · 2026-06-28 · Condition: human approval on payouts > €5,000.

Living Register — AI inventory .xlsx

SAMPLE

AI System ID	System	Owner	Risk tier	Phase	Last gate	Residual	Next review
FCL-AI-0142	Claims Triage Copilot	GRL · Meyer	High	8 · Monitor	go-live ✓	Low	2026-09-01
FCL-AI-0138	Support Copilot	PO · Naidu	Limited	9 · Operate	go-live ✓	Low	2026-08-15
FCL-AI-0151	Underwriting Assistant	GRL · Meyer	High	5 · Evaluate	meets-the-bar ⧗	Med	2026-07-22
FCL-AI-0155	Marketing Copy Gen	PO · Silva	Minimal	7 · Deploy	go-live ✓	Low	2026-10-01
FCL-AI-0160	KYC Doc Extractor	AR · Chen	High	3 · Design	design-review ⧗	—	2026-07-18

Get a sample Evidence Pack template

The redacted .docx we hand over — so you have something to put in front of a client who’s asking. Score first, gate second.

Sample templates are being finalised with our delivery team — enter your email and we’ll send the current redacted pack, or reach us at connect@focaloid.com.

Accountability

Every phase has a named owner. That’s the point of a RACI.

“The team is responsible” is how governance fails. Each phase names who is Accountable (one person, the decision), who is Responsible (does the work), and who is Consulted and Informed. Eight roles carry the lifecycle between them.

§The independence rule. Phase 6 review is done by people who didn’t build it. Sign-off by the builder isn’t sign-off.

AAccountableRResponsibleCConsultedIInformed

Click a phase row to see who owns it end to end.

Two Ways In

Building new AI, or fixing the AI you already shipped.

Governed by construction

We’re building it — the nine phases are the build.

Governance costs almost nothing when it’s threaded through from Phase 1 — and the evidence pack assembles itself as you go, because every gate produced one. You ship with the proof already in hand.

Cheapest when it’s designed in from the start.

Review-ready, retrofitted

You already shipped — and someone’s now asking.

You have AI in production and a customer or regulator is asking questions. We run the lifecycle backwards over what exists — inventory it, risk-tier it, evaluate and red-team it, document it, and close the gaps that matter. You end up with the evidence you should have had.

The fastest way out of a stalled review.

Start Here

A Governance Health Check — turn “trust us” into evidence you can show a client.

A structured, eight-dimension diagnostic of your AI estate. Scored, with a profile of your strengths and gaps and a practical 90-day roadmap. Fixed scope, fast, and it ends with something you can act on rather than a proposal.

01 · Assess

Score eight dimensions

Inventory, risk, data, testing, security, guardrails, monitoring and policy — scored honestly.

02 · Prioritise

A risk-based gap list

Not everything at once — what matters most, first, ranked by risk.

03 · Roadmap

A practical 90-day plan

A plan to close the gaps, and a map to the right engagement level.

Governance maturity — 8-dimension self-check0 / 16 answered

Your score is computed in your browser and shown instantly — nothing is sent until you request the full report.

Engagement Levels

Four levels. Start where you are.

Level 1

Assessment & Roadmap

The Health Check, scored, with a prioritised 90-day plan. Where most engagements begin.

Start here

Level 2

Govern one system

Take a single AI system end to end through the nine phases — the full evidence pack and a signed verdict.

Indicative — being finalised

Level 3

Govern the estate

Inventory, policy and a Living Register across every AI system you run — governance as a standing capability.

Indicative — being finalised

Level 4

ISO/IEC 42001 readiness

The full AI management system, prepared for certification against ISO/IEC 42001.

42001-aligned

Who It’s For

Built for teams whose AI has to survive someone else’s scrutiny.

The EU AI Act applies to us and we don’t know what we’d have to show.

An enterprise customer’s risk review is holding up the deal.

We shipped AI fast and now we can’t evidence any of it.

We have agents taking real actions and no one can say what they did, or why.

Our board asked who is accountable for the AI, and the room went quiet.

We need ISO 42001 or our procurement stalls.

Usually a CTO, VP of Engineering, Head of Risk or Compliance, a DPO, a CISO, or a founder whose deal is stuck in a security questionnaire — in Europe, the UK and the US.

Why Focaloid for Governance

We govern AI because we build it.

A method, not an opinion

Nine phases, eight gates, a RACI and an artifact per phase — proprietary, documented, and mapped to three frameworks. Most firms bring principles. We bring a lifecycle.

Engineers, not auditors

We build agentic systems in production. We know what a guardrail costs, where a trace has to be wired, and why a red-team finding is real — because we’ve had to fix them.

We close the seam

Our artifact packs span the Design → Build → Deploy gap that GRC and MLOps platforms each stop short of. That’s the whole reason the method exists.

Composed, not sold

We don’t resell a platform. We compose to whatever stack you already run — and we’ll tell you when a tool you already own is enough.

Partners & Certifications

ISO/IEC 27001:2022 CertifiedClaude Partner NetworkDatabricks PartnerSnowflake Partner

Common questions

Before you book.

Does the EU AI Act actually apply to us?

If you place AI on the EU market or your output is used there, usually yes — regardless of where you're headquartered. What changes is your role (provider or deployer) and your risk tier, and that classification is exactly what Phase 1 settles.

We're US-based. Is this only a European problem?

No. The NIST AI RMF is becoming the de facto US bar, state laws are arriving, and your enterprise customers' security reviews already ask these questions. The same lifecycle answers all of it.

Will this slow our delivery down?

A gate is a five-minute documented decision, not a committee. Threaded in from Phase 1 the overhead is small — it's the retrofit, after someone asks for evidence you never captured, that's expensive.

Can you govern AI you didn't build?

Yes — that's the review-ready path. We run the lifecycle backwards over what's already in production, and close the gaps that matter.

Do we have to buy a governance platform?

No. We compose to the stack you already run, and we'll tell you when what you own is enough. Our artifact packs are the connective tissue — that's the part no platform sells.

What do we actually walk away with?

A signable Evidence Pack per phase, a Living Register of your AI estate, and a scored maturity profile with a 90-day roadmap. Documents you can hand to a client, an auditor or your board.

How fast is the Health Check?

Fixed scope and fast — you get a scored profile and a prioritised roadmap, not a proposal. It's the cheapest way to find out how exposed you actually are.

Don't promise your AI is safe. Prove it.