AI agents act in seconds. Trust can’t wait for a quarterly review. We build your rules into the work itself, so every person and every agent stays inside them, and every decision leaves proof anyone can check.
Input fingerprint ffc2…04bb matches the recorded input
Clock pinned to 2026-10-25 14:02:11Z
Same answer: held for human approval.
Illustrative example with synthetic data. Replay recomputes this record’s SHA-256 in your browser, the same check an auditor makes.
GRC was built as a gate
AI doesn’t wait at gates
AI agents now move money, change records and talk to customers in seconds. A review queue can’t keep up, and a rule written in a document can’t stop an agent. Guardrails can.
So trust has to be built into each decision as it happens, not assembled for an audit months later. The decision in the record above shows how, one step at a time.
2 December 2026EU AI Act marking of AI-generated content for systems already on the market.
2 December 2027Annex III high-risk obligations, including the fundamental rights impact assessment.
2, 10 or 15 daysThe deadlines for reporting a serious incident under Art. 73.
As a gateWork stops and waits in review queues.
Work keeps moving, inside clear lines.
Rules live in documents, and people read them differently.
Rules run as code, the same for every person and agent.
Evidence is chased by email before every audit.
Evidence writes itself as the work happens.
A sample gets checked, months after the fact.
Every action is checked at the moment it happens.
Agents get blocked outright, or slip through unseen.
Agents get exactly the access they need, and a human steps in when it counts.
Registered
Before it acts, the agent is known: what it is, who owns it and how risky it is
The payments-assistant is in the register with a named owner, a risk class and its Annex III screening. You can’t trust what you can’t see, so trust starts with knowing every asset, every owner and every risk, and where to focus first.
The same framework covers every asset type you run, and a new one is added as content, without rebuilding anything.
AI agents
Copilots
Models
MCP servers and tools
Cloud
Data
Applications
Vendors
Business processes
Your next one
Built withRiskLens IQ
You keepYour AI register, use-case heatmap, Annex III screening, and provider and deployer role mapping.
Ruled
One rule, the same for every person and every agent
Your people and your AI agents should play by the same rules. We ground both in one shared source of truth, so a policy means the same thing to an analyst reading it and to an agent’s tool call being checked against it.
Person · payments team
Sees the rule in their workflow and knows exactly when an approval is needed.
One shared rule
Payments above your threshold need a named human approver.
Cites EU AI Act · NIST · your standard
AI agent · payments-assistant
The same rule checks its tool call and holds it for a human approver.
Clear for people.Rules in plain language, shown where the work happens, so nobody has to guess.
Enforced for agents.The same rules run as code where agents call tools and touch data.
Risk appetite as reviewed parameters.The threshold changes through review, not through code, and every change is recorded.
A clean set to start from.Your current controls are assessed first, and overlaps and duplicates are cleaned up before anything is automated.
Built withPolicyForge and controls rationalization
You keepYour Responsible AI policy suite, lifecycle gates, oversight rules, exception workflow and a rationalized control set.
Checked
At the moment of action, guardrails know the difference
Go
Most actions should just go.
A service agent looks up a customer’s account balance.
Ask
Some need a human first.
This decision: a 48,200 EUR transfer, held for approval.
Stop
A few should never happen.
A research agent tries to connect to an unverified MCP server.
Guardrails at work
AI agent · service-assistantLooks up a customer’s account balanceAllowed
Person · finance analystShares a labelled report with an advisorAllowed, labels kept
AI agent · research-botConnects to an unverified MCP serverStopped
Model · claims-triage v3Promoted to productionAllowed, review on file
Every decision is written down as evidence, the moment it happens.
Illustrative example: the same guardrails deciding actions by people, AI agents, tools and models.
AI security protects AI from misuse and attack
Agents use only the tools they’re approved for. Everything else is denied by default.
Models and MCP servers are checked before anyone trusts them.
Every action is tied to the agent and the person it acts for.
A kill switch stops an agent’s actions where you choose.
AI safety makes sure AI behaves as intended
A human approves high-impact decisions before they happen.
Every use case gets a risk class, an owner and the right lifecycle gates.
Controls fail closed: an error or missing evidence means deny, never pass.
Readiness for the EU AI Act, built in rather than bolted on.
Microsoft 365 CopilotBeside Purview and DLP
Copilot StudioBeside Purview and DLP
AWS BedrockBeside native AWS controls
AgentCoreBeside AgentCore Policy
Built withControl automation
You keepYour controls as versioned policy code with tests, running on your platforms beside the native controls you already pay for.
Approved
A named person decides. The accelerators never do
The transfer waits until someone with the authority says yes or no. Accelerators draft and assemble; your risk, compliance and legal teams decide, and the decision is recorded with who made it.
First lineBusiness and technology
Owns the evidence and the fixes, and approves within its authority.
Second lineRisk, compliance, legal
Owns policy and risk appetite, and challenges the first line.
Third lineInternal audit
Reads and re-checks, but never edits the evidence the other lines rely on.
Built withPolicyForge
You keepYour oversight rules, role charters and exception workflow.
Recorded
Evidence that writes itself, the moment the decision is made
The rule version, the inputs and the time go into a hash-chained record. Nothing is chased by email before an audit, and missing evidence is never a pass: if a control can’t decide, you see a gap, never a green light.
Test once. Satisfy many.
Every framework you answer to maps onto one common control library, so one control, tested once, counts toward all of them.
EU AI Act
Built and tested
A full AI governance control library, written as code and tested.
NIST and CSA cloud controls
In use today
Security and cloud controls that run continuously against your environment.
Your internal standards
Added as content, like any other catalog: your own policies, held to the same bar and proven in the same record.
SOC 2
Coming next
Mapped onto the same library, so what you’ve already tested counts.
Mortgage compliance
Coming next
TRID, ECOA and Reg B, HMDA and fair lending, tested on every loan instead of a sample.
SOX ICFR
Coming next
Business controls evidenced as the work happens, not assembled at year end.
Ask us on a call where each “coming next” framework stands for your organization.
Built withEvidenceOps
You keepYour evidence repository, testing packs, attestations, issue logs and assurance dashboards.
Replayed
Proof? Check it yourself
Trust shouldn’t rest on a slide. Years from now, your auditor can take this decision, re-run it offline and get the same answer, because the record keeps the rule version, inputs and time it used.
Same answer, every time.
The policy pack, the input fingerprint and the clock are pinned, so a replay can’t drift.
Tamper-evident.
Each record carries the hash of the one before it. Change one, and the chain shows it.
Independent by design.
Internal audit reads and replays with its own access, and never edits.
Built for everyone accountable for trust
Each leader’s question is answered somewhere in the life of one decision
CIOSpeed with guardrails
Can we scale AI agents without a governance bottleneck?Every agent and owner in one inventory, and self-service gates where the rules are clear, so routine agents never queue behind risky ones.See it at Registered
CROProvable compliance
Can we show a regulator what we did, and when?Coverage by obligation from live data, tamper-evident evidence, and risk appetite set as reviewed parameters.See it at Recorded
CSOControl at the point of action
Can an agent be tricked into doing damage?Tools denied by default, human approval for high-impact actions, checks on models and MCP servers, and a kill switch.See it at Checked
Compliance and GRCCoverage you can trace
Which obligations are covered, by what, with what result?Answered from live data instead of a spreadsheet, because every control cites the provisions it implements. Your lawyers review the reading first.See it at Ruled
Internal auditAssurance without the fire drill
Can we rely on this evidence without collecting it again?Recorded decisions you can replay, evidence collected once, and read-only access that keeps you independent.See it at Replayed
Business ownersRun the business, prove the controls
How do we evidence controls without stopping the business?Evidence collected where controls run, not chased by email, with exceptions and issues in one workflow.See it at Approved
Trust, engineered with you
Not sold to you
No licence to pay.
We’re a consulting firm, not a software vendor. The accelerators come with the engagement, with no subscription, seat count or renewal waiting at the end.
Yours to keep.
The code, the configuration and the know-how stay with you. Your IP grows with every engagement, and your team knows how to extend it long after we’ve gone.
Practitioners on your side of the exam.
Our founders spent decades running GRC inside regulated industries. We know the request lists, attestations and committee packs, and now we lead with AI.
Transformation at speed.
Your register, policies, controls and evidence start from working accelerators, not from a template and a spreadsheet.
What we don’t do.
Sell you a licence
You pay for the engagement and keep what it builds. Open foundations such as Open Policy Agent and OpenTelemetry mean nothing depends on us.
Give legal advice
Our controls encode a technical reading of regulation. Your legal and compliance teams review it before anyone relies on it.
Replace your native controls
Purview, DLP and AgentCore Policy stay in place. Our controls sit beside them and record what they decided.
Pretend every duty is automated
Some obligations, such as impact assessments and incident reports, need people. We make sure the record behind them is complete.
The art of the possible
Questions worth a conversation
What if an examiner’s request list answered itself from evidence you already hold?Each request mapped to controls and evidence, with the rule version in force on each date.
What if internal audit re-performed a control across the whole population, not a sample?Recorded decisions re-run on demand, with the workpapers generated.
What if your self-assessment arrived pre-filled from recorded evidence?People answer only for the controls that are still manual.
What if every loan file were tested, not a sample?Mortgage compliance checks such as disclosure timing, run on every loan.
What if you knew which controls a new regulation touches the day it is published?Affected requirements, controls and business areas, with drafted updates for your reviewers.
These describe what’s possible, not what’s on a shelf today. On a call we’ll walk you through where each one stands and what it would take for your organization.
Engineer trust. Unlock speed.
A first call is a working conversation about your frameworks, your AI platforms and where your evidence lives today. We’ll tell you plainly what we can bring now and what’s possible next.
