We govern your AI in nine phases, with a clear yes or no between each one. Sentinel, our runtime layer, checks every request and records the proof. One set of evidence covers the EU AI Act, NIST and ISO 42001. You get a signed decision you can show a client, an auditor or your board.
Governance on paper is not enough, and guardrails without a method are just filters. We give you both, built by one team, so the rules you set at the design table are the rules the system follows in production.
How we decide whether an AI system is safe, fair and accountable.
The runtime that holds the system to those decisions on every request.
You use them together. The method sets the bar. Sentinel holds the system to it. Both come from the same engineers, and both run in your own cloud.
The model works. Then it meets security, risk and compliance, and the questions start: what stops a bad input, who approved this, where is the evidence. Most systems cannot answer. So they wait.
A prototype that impresses in a sprint demo still has to clear security, model-risk and compliance before it reaches a customer. Without guardrails, tests and an audit trail, it cannot answer the questions those teams are paid to ask. So it waits.
Input and output checks, test gates and logging get put off until compliance asks. Then they are added under pressure, to a system that was never designed for them. Governance added at the end is expensive, fragile and never quite complete.
“Trust us, it is safe” is not a document. Under the EU AI Act and the NIST AI RMF you have to show what was tested, what was decided, and what actually happened in production. Most AI systems never record any of it.
An AI model is an engine. Governance is the rest of the car: the brakes, the dashboard, the service book and a named driver. Without them you do not have a vehicle. You have a fast object with no way to stop it. These are four failures we see in real systems, and where our method catches each one.
An agent with access to a refund tool issued 500 refunds overnight. The code ran fine. The problem was authority nobody had limited.
A message hides an instruction that takes over the model or pulls data out. These are the OWASP LLM Top 10 attacks (the ten most common attacks on AI language systems). A normal functional test never sees them.
A system that was accurate at launch slowly gets worse in production. Without monitoring, nobody notices until a customer does.
Personal data gets loaded into a search index with no legal basis and no record of where it came from. Data with no owner makes everything built on it impossible to govern.
Each failure has a phase where it is cheap to catch. That is what the nine phases are for.
It starts before anyone writes code and does not stop until the system is switched off. Three groups of work own their stretch of the build. Map works out the context and the risk. Measure proves the system does what it should. Manage ships it, watches it and responds. Govern runs across all nine phases. And it is a loop: what Phase 9 learns goes straight back into Phase 1.
Eight gates. A gate is a written yes-or-no decision between two phases. It takes five minutes and a named person, not a committee. Gates are cheap. Drift is expensive.
Sentinel wraps any model or agent with the controls an enterprise review expects. They are enforced at runtime, on every request, not written in a policy document nobody reads. It works around the AI we build with you, and around the AI you already run.
Every request is screened before it reaches the model: personal and sensitive data is detected, prompt-injection and jailbreak attempts are caught, and topic and policy filters apply. A harmful or off-limits request is stopped at the door, not after the damage.
Every answer is screened before it reaches the user: personal data is masked, toxic or unsafe content is filtered, and the policies your business runs on are enforced. What the model says is checked as carefully as what it was asked.
Tests for groundedness (does the answer rest on real sources?), accuracy and safety run before anything ships, and keep running in production. A change passes the gate only when it meets the bar you set. Quality becomes a release requirement, not a hope.
Tracing, response time, cost, quality and drift on every call, with an alert when something moves. You can see what your AI is doing in production, and you are told before a small drift becomes an incident.
Every prompt, decision, tool call and action is written to a trail that cannot be edited, with the lineage that links an answer back to its inputs and its version. When someone asks what happened, and why, the answer is already recorded.
Who, and what, may call which model, tool or dataset is defined once and enforced centrally, with role-based access and a policy engine. Authority is limited on purpose, not discovered after an agent does something it should not have.
Five requests, checked one after another. Sentinel sends each down the same stack: the checks on the way in, the model, the checks on the way out, and the evaluation gate. It returns a decision and writes an audit entry to match. Pick any request to jump to it. This shows how Sentinel behaves. It is not a live model call.
The same stack on every request. The safe ones pass. The rest are masked, refused or held for a person. All of it is logged, and the log is what the evidence pack is built from.
Sentinel sits between the request and the response. It screens what goes in and what comes out, gates what ships, and records all of it, around a model or agent it does not need to own.
Sentinel follows the same path whether it is governing a system we built or wrapping one you already run.
Classify the AI system and its risk: what it does, what data it touches, and how the EU AI Act and the NIST AI RMF apply. Governance is sized to the risk, not one size for all.
Wrap inputs and outputs with the guardrails the system needs: personal data, injection, topic, safety and policy, enforced on every call from day one.
Set up evaluation gates for groundedness, accuracy and safety, and wire them into the release path so nothing ships below the bar you set.
Add tracing, metrics, cost and drift detection, with alerts that fire before a drift becomes an incident.
Capture every prompt, decision and action in an audit trail that cannot be edited, with the lineage that ties each answer to its version and its inputs.
Produce the evidence pack that risk, security and regulators actually ask for: the tests, the decisions and the production record, ready to hand over.
The lifecycle is mapped to each framework, so one set of evidence satisfies all three. No parallel compliance projects. No second team producing the same documents in a different format for a different auditor.
The European Union's law on artificial intelligence. In force since 2024, with its duties phased in over the following years. It sorts AI systems by risk level and sets duties for providers and deployers of high-risk systems: risk management, data governance, technical documentation, human oversight, accuracy and robustness, and monitoring after the system is on the market.
Applies if you place AI on the EU market, or its output is used in the EU, wherever you are based.The AI Risk Management Framework from NIST, the US standards body. Voluntary, but it has become the common reference for AI risk in the US and in enterprise security reviews everywhere. It has four functions: Govern, Map, Measure and Manage. Our nine phases use the same names on purpose.
The bar your enterprise customers' security reviews already ask about.The international standard for an AI management system, published in 2023. It is the one you can be certified against, in the way ISO/IEC 27001 certifies information security. It asks for an AI policy, leadership, risk and impact assessment, a set of controls, and continual improvement.
The certificate that unblocks procurement when a buyer asks for it.Pick one framework to see what we produce against it, in which phases, and the clause it answers.
If you place an AI system on the EU market, or its output is used in the EU, then usually yes, wherever your company is based. What changes is your role (provider or deployer) and the risk level of the system. Phase 1 settles both, and the answer goes in the register.
Monitoring tools tell you what happened. A trace or a drift chart is not an answer to “show me this system was approved, by whom, against what bar.” That answer is a document, and it is what we hand over.
One per phase, ready to sign. The model card, the test and red-team results, the data lineage, the threat model, the DPIA where personal data is involved, the yes-or-no decision and any conditions attached to it. A document a client's risk team or a regulator can read from start to finish.
The AI inventory that stays current. Every system, its risk level, its owner, its gate history, the risk that remains, and its review dates. It is the thing an auditor asks for first, and the thing most organisations cannot produce.
Everything joins on one identifier. Each system gets a single ID, and every document, gate decision, test run, trace and register row hangs off it. That is what turns nine separate phases into one thread an auditor can follow, and it is the thing no GRC platform and no MLOps platform gives you, because each of them only sees half the lifecycle.
| AI System ID | System | Owner | Risk | Phase | Last gate | Left | 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 ⧗ | Medium | 2026-07-22 |
| FCL-AI-0155 | Marketing Copy Generator | PO · Silva | Minimal | 7 · Deploy | go-live ✓ | Low | 2026-10-01 |
| FCL-AI-0160 | KYC Document Extractor | AR · Chen | High | 3 · Design | design approved ⧗ | – | 2026-07-18 |
Illustrative entries. Every row points at a system's evidence packs through its AI System ID.
The tooling market is split. GRC platforms (governance, risk and compliance) govern from the policy side. MLOps platforms govern from the model side. They meet, badly, at the seam between design, build and deploy, and nobody has closed it. That gap is where ungoverned AI actually ships.
Strong on policy. Stops at the build.
Strong on models. Thin on policy.
Where the tools report numbers, we produce a signed decision. The packs span the design → build → deploy seam that GRC and MLOps platforms each stop short of.
Sentinel composes these. We do not resell a platform. We work with whatever stack you already run, and we will tell you when a tool you already own is enough.
“The team is responsible” is how governance fails. For each phase we name who is Accountable (one person, who makes the decision), who is Responsible (does the work), and who is Consulted and Informed. Eight roles carry the lifecycle between them.
Accountable for the decision to build (Phase 1) and for authorising the release (Phase 7).
Accountable for the design (Phase 3) and for the quality of the build (Phase 4).
Accountable for intake, data sign-off, whether the testing was enough, ongoing oversight and retirement.
Responsible for prototyping and building the system.
Responsible for sourcing, cleaning, indexing and versioning every dataset.
Responsible for the tests: correctness, groundedness, and the path an agent takes.
Responsible for attack testing, and leads when there is an incident.
Responsible for deployment, observability and running the system.
The independence rule. The Phase 6 review is done by people who did not build the system. Sign-off by the builder is not sign-off.
A typical split. It is confirmed against your own organisation at the start of an engagement. Hover a phase to see its name in full.
Both end in the same place: a system with the proof in hand. The difference is where you start.
Governance costs almost nothing when it is threaded in from Phase 1. The evidence pack builds itself as you go, because every gate produced one. You ship with the proof already in hand.
Cheapest when it is designed in from the startYou have AI in production and a customer or a regulator is asking questions. We run the lifecycle backwards over what exists: list it, give it a risk level, test and attack it, document it, and close the gaps that matter. Sentinel wraps the system, so it can pass the review on its next attempt, not its fifth.
The fastest way out of a stalled review · weeks, not quartersA structured check of your AI estate across eight areas. Scored, with a profile of your strengths and gaps, and a practical 90-day plan. Fixed scope, fast, and it ends with something you can act on rather than a proposal.
We score eight areas: inventory, risk, data, testing, security, guardrails, monitoring and policy. Honestly, not generously.
A gap list ranked by risk. Not everything at once. What matters most, first.
A practical 90-day plan to close the gaps, and a map to the right level of engagement.
Two minutes. Pick the answer that is true today, not the one you would like to be true. Nothing is sent anywhere. This runs in your browser only.
The real one is scored by us, against evidence, with a 90-day plan out the other side.
Most engagements begin with the Health Check. Where they go next depends on what it finds and on what someone is asking you to show.
The Health Check, scored, with a prioritised 90-day plan. Where most engagements begin.
Where most startTake a single AI system end to end through the nine phases: the full set of evidence packs and a signed decision.
Scope being finalisedInventory, policy and a Living Register across every AI system you run. Governance as a standing capability, not a project.
Scope being finalisedThe full AI management system, prepared for certification against ISO/IEC 42001.
Certification-readyIf an AI system touches regulated data, faces a review, or makes a decision someone will later question, it needs both the method and the mechanism. These are the sentences we hear on the first call.
“The EU AI Act applies to us and we do not know what we would have to show.”
“An enterprise customer’s risk review is holding up the deal.”
“We shipped AI fast and now we cannot show evidence for any of it.”
“We have agents taking real actions and nobody can say what they did, or why.”
“Our board asked who is accountable for the AI, and nobody had an answer.”
“We need ISO 42001 or our procurement stalls.”
Wherever an AI decision has to be defensible.
Copilots and agents over payments, portfolios and advice, where a wrong or unsupported answer is a compliance event, not just a bug.
AI over clinical and patient data, where controls on personal data, groundedness and a complete audit trail are not optional extras.
Underwriting, claims and servicing AI that has to show fair, consistent and explainable decisions on demand.
Any team that has to map, measure and manage AI risk and produce evidence for it. Sentinel produces that evidence as the system runs.
The people who have to sign off. Sentinel gives them enforced guardrails, limited access, and a record they can actually inspect.
A model or agent that is live but ungoverned. Sentinel wraps it and makes it review-ready, without a rebuild.
Who usually calls: a CTO, a VP of Engineering, a Head of Risk or Compliance, a DPO (data protection officer), a CISO, or a founder whose deal is stuck in a security questionnaire. In Europe, the UK and the US.
Plenty of firms bring principles. Plenty of tools can filter a prompt. Fewer are built so you would stake a regulated system on the result, and hand your reviewers the evidence to prove it.
Nine phases, eight gates, named owners and a document per phase. Written down, and mapped to three frameworks. Most firms bring principles. We bring a lifecycle.
We build agent systems that run 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 have had to fix them.
Our evidence packs span the design → build → deploy gap that GRC and MLOps platforms each stop short of. That is the whole reason the method exists.
We do not resell a platform. We work with whatever stack you already run, and we will tell you when a tool you already own is enough.
With Sentinel, guardrails, tests, the audit trail and access control are properties of the runtime from day one, not a retrofit for when compliance finally asks.
Sentinel runs inside your own environment and security perimeter. Your prompts, answers and audit trail never have to leave to be governed.
Straight answers to the questions that come up on the first call.
If you place an AI system on the EU market, or its output is used in the EU, then usually yes, wherever your company is based. What changes is your role (provider or deployer) and the risk level of the system. That classification is exactly what Phase 1 settles.
Yes. The NIST AI RMF is the common reference in the US, and US state laws are arriving. The UK regulates AI through its existing regulators. And your enterprise customers' security reviews, wherever they are, already ask the same questions. The same lifecycle answers all of it, because the evidence is captured once and mapped to each framework.
A gate is a five-minute written decision, not a committee. Threaded in from Phase 1, the extra work is small. What is expensive is the retrofit: adding evidence after someone has asked for it and you never captured it.
Yes. That is the review-ready path. We run the lifecycle backwards over what is already in production, close the gaps that matter, and Sentinel wraps the system with guardrails, tests, observability and an audit trail. It governs the AI without needing to own or replace it.
Both, in sequence. Sentinel is our governance framework and the runtime components that implement it: guardrails, evaluation gates, observability, audit and access control. We stand it up around your AI with your team. Then it runs in your cloud as part of the system, not as a service you send your data to.
Method versus mechanism. The method is how we decide whether an AI system is safe, fair and accountable: nine phases with a written gate between each. Sentinel is the machinery that enforces those decisions at runtime and produces the evidence each gate needs. You use them together. The method sets the bar. Sentinel holds the system to it.
Yes. Sentinel wraps any model or agent: Claude, open-weight, fine-tuned, or a third-party system, whether we built it or you already run it. That is exactly what makes an existing system review-ready.
No. We work with the stack you already run, and we will tell you when what you own is enough. The evidence packs and the register are the connective tissue. That is the part no platform sells.
No. Sentinel deploys inside your own cloud and security perimeter. Prompts, answers, test results and the audit trail stay where your data already lives. Nothing has to leave to be governed.
A signable Evidence Pack per phase, a Living Register of your AI estate, and a scored profile with a 90-day roadmap. Documents you can hand to a client, an auditor or your board.
Fixed scope and fast. You get a scored profile and a prioritised roadmap, not a proposal. It is the cheapest way to find out how exposed you actually are.
For a system that is already live, wrapping it with guardrails, tests, observability and an audit trail is usually a matter of weeks, not quarters, because Sentinel is added around what you run rather than rebuilt into it. The first milestone is usually the evidence pack your reviewers have been asking for.
The EU AI Act and the NIST AI RMF first, aligned to ISO/IEC 42001, delivered by an ISO/IEC 27001-certified team. Because the evidence is captured once and mapped to each framework, proving your controls in one place counts towards the others instead of starting over. Sentinel is model- and framework-agnostic by design, so a new control or a new regulation can be added the day it lands.
Start with a Governance Health Check: eight areas, scored, with a 90-day plan out the other side. It is how we turn “trust us” into evidence you can put in front of a client.
Fixed scope. Fast. We look at your AI estate across eight areas, score each one against evidence, and give you a profile of strengths and gaps with a practical 90-day plan. Bring one system you are building, or one that is already live.
Four fields. We reply within one working day with a time for a 30-minute call and a short list of what to bring.
Someone who has read your note will reply within one working day with a time and a short list of what to bring. If you already have an AI inventory, a policy, or a test report, keep them ready. That is where the Health Check starts.