KeelGovernance backbone · by Amplinth

Agentic AI you can put in front of a regulator.

Keel is the governance backbone beneath every Amplinth agent. The model proposes and narrates; a deterministic engine decides; a named person signs anything consequential — and every step is recorded, attributed and replayable. Governance you can prove, not just promise.

Propose · decide · sign — the model never has the last wordEvery action on a tamper-evident audit ledgerBuilt into every Amplinth solution — not an add-on
The problem

Most agentic AI asks you to trust the model. Boards can’t.

Put an agent near a real decision — a price, a credit line, an award — and “the AI is usually right” stops being good enough. Your risk team needs a line of accountability, not a leap of faith.

01
Ungoverned fleetsAgents reach production that no one registered, owns, or can audit. When something goes wrong, there’s no one to point to and nothing to inspect.
02
Silent changeA prompt is swapped, a model updates, a memory drifts — and nothing records it. The system you signed off is not the system running today.
03
No line of accountabilityWhen an agent moves a number, can you say who decided, on what evidence, and replay exactly how? For most stacks, the honest answer is no.
04
Regulation is arrivingEU AI Act, ISO 42001, SM&CR: your risk team is increasingly asked to evidence controls, not intentions. Screenshots won’t clear an audit.
The one rule

The model proposes.
The engine decides.
The human signs.

Keel makes one rule structural, not optional: language models propose and explain, deterministic logic decides, and a person signs anything consequential — nothing ungoverned ships, and no change is silent. A figure a model emits stays proposed — it can’t become the basis of a decision until deterministic logic or an accountable person confirms it. So a wrong output, or a poisoned memory, can’t quietly move a number, a decision, or money. That’s the difference between governance you hope holds and governance that’s built in.

PROPOSEDDECIDEDSIGNED01Modelproposes · narrates02Enginedeterministic decision03Humansigns anything consequential04Actionmoves a number · moneyAUDIT LEDGER — every step recorded, attributed, replayable

A figure a model emits stays a proposal — it can’t move a number until deterministic logic decides and a named person signs. Every step lands on the ledger.

What Keel gives you

Every agent: registered, gated, metered, and provable.

Stated as outcomes — what your risk, security and finance teams actually get.

A complete, tamper-evident audit trail

Every agent action is recorded and replayable: what ran, which model actually served, what it was told, what it produced. When someone asks “why did the AI do that?”, you can show them — not reconstruct a guess.

Nothing ungoverned reaches production

Every agent is registered, versioned, owned by a named person, independently reviewed and signed off before it can act. An unregistered or unvalidated agent simply can’t be served. Governance isn’t a review meeting; it’s a precondition.

Full AI cost visibility

Every unit of spend is attributed by model, team and feature, against budgets — so AI cost is governed and predictable, not a surprise line item your CFO finds in arrears.

Agents that improve — safely

Agents can learn and get better over time, but every change is reviewed before it can affect a decision. No silent self-updates, no learning the wrong thing.

Tested, not trusted

Agents are stress-tested adversarially, and their safety guards are measured — not assumed to hold. “It’s safe” becomes a result you can point to.

Governed in proportion to consequence

The strictest controls apply where the stakes are highest — a human gate before an agent acts on anything material — and stay light-touch where they’re not.

Right to erasure, honoured

When a person’s data must be deleted, it’s removed across the record — with the audit trail’s integrity intact.

No change is ever silent

A prompt swap, a model update, a memory change: all recorded, always attributable to a person and a point in time.

It flags where your own governance is too weak

Keel surfaces governed recommendations where your controls look weaker than they should be. Every suggestion is advice only: acting on it is a separate, signed change. Governance that reviews itself, and still can’t act on its own.

The principle, in one line

A wrong answer can’t move a number, a decision, or money.

Say it to a board, a regulator, or your own risk committee: the governance is structural, and the record proves it.

The console

One pane of glass for every agent you put in production.

Keel isn’t only a backbone your engineers build on — it’s a console your risk, compliance and model-risk teams log into. Every governed agent, what it’s allowed to do, what it’s costing, where a human gate sits, and the evidence behind every decision — across the whole fleet, in one place. It runs inside your own environment: the view, the record and the controls never leave your tenant.

COMMAND CENTRE42Agents governed5Awaiting sign-off£18kAI spend · moGOVERNED AGENTSpricing-advisorHighSignedkyc-summaryMedClearedcredit-memoHighGatedsupplier-triageLowClearedforecast-noteLowSigned

Illustrative — an abstracted console view, not a product screenshot.

/ See

The whole fleet

Every governed agent, its owner, how tightly it’s governed and whether it’s cleared to act — one view instead of a spreadsheet no one trusts.

/ Prove

Evidence on demand

The record behind any decision, and control coverage mapped to the regimes you answer to — generated as agents run, ready when an auditor asks.

DETECTIVE · MONITORINGAgentActionMonitorwatches · alerts afterPREVENTIVE · KEELAgentKeelGATEAction✕ Refused

Detective tools watch from the side and alert after the fact. Keel sits in the path — it can refuse an ungoverned action before it happens.

Adoption

Governed without rewriting your agents.

You don’t re-platform to get governed. Keel wraps around the agents you already run, so you can start by recording everything with nothing blocked — a shadow run that proves the controls fit your reality — then switch enforcement on, one agent at a time, at the pace your risk team is comfortable with. Evidence first, enforcement when you’re ready.

Regulation

Built for the frameworks your risk team already answers to.

Keel enforces the load-bearing technical controls behind modern AI-governance regimes — and captures the record to prove they held — so the evidence your risk team needs is produced as agents run, then mapped to the regimes you answer to and exported when an auditor asks. One governed spine, mapped per jurisdiction.

AI governance & assurance

EU AI ActArticle-level obligations for high-risk (Annex III) systems.
ISO/IEC 42001The AI management-system standard.
NIST AI RMFThe AI risk-management framework.

Financial services & model risk

UK SM&CRSenior-manager accountability for decisions.
PRA SS1/23Model risk management expectations.
US SR 11-7Fed/OCC model-risk management guidance.
FCA · Consumer DutyFair-outcome and conduct obligations.
Swiss FINMASupervisory expectations for AI in finance.

Data & automated decisions

GDPR Article 22Rights around solely-automated decisions.
UK DUAAData Use and Access Act — UK data reform.
Right to erasureDeletion honoured across the record.
Audit & explainabilityEvery decision replayable on demand.
Decision recordIllustrative
Agentpricing-advisor · v4
Model servedverified at run time
Proposed → decidedengine-confirmed
Signed offnamed approver
Consequence tierhuman gate · reversible
EU AI Act ✓ISO 42001 ✓SS1/23 ✓GDPR Art 22 ✓
append-only · tamper-evidentreplayable on demand

Keel enforces and evidences the technical control subset — the auditable, provable part. It isn’t a certification and doesn’t make you “compliant” on its own; your organisation’s process and documentation complete the picture. What Keel removes is the hardest part: proving the technical controls actually held, every time.

Proof

Not a whitepaper — built into our own products.

Keel is the backbone Amplinth’s agents are built on: the governance primitives are proven and deployed, and we’re activating them across our solutions — MetalAiQ, our lending platform, and ProcurAiQ — so every Amplinth solution runs on the same governed spine. It isn’t a stance about how agentic AI should work; it’s the backbone ours is built on.

/ Innovate

MetalAiQ

Closed-loop metals R&D — every approved trial captured and governed on the same provable backbone.

/ Operate

Lending & ProcurAiQ

Regulated lending decisions and enterprise procurement — built as governed agents: registered, audited, with a human on every consequential call.

/ Everywhere

Every Amplinth solution

Not an add-on, not a premium tier — the same backbone ships under all our solutions by default. Explore the solutions →

Keel is the governance backbone Amplinth’s agents are built on — its primitives proven and deployed, unified into one framework.

Deployment & trust

Your data. Your cloud. Owned end-to-end.

The audit and decision path is Amplinth-owned and self-hostable — never a third-party AI service sitting inside the path your decisions run through. It runs in your cloud, on your data — so the record that proves your governance held stays entirely under your control.

/ SaaS

Amplinth SaaS

Fully managed — the fastest path to a governed agent fleet.

/ VPC

Your own cloud

Your data, decisions and audit record never leave your control.

/ Any region

Any cloud

Azure, GCP or AWS — data residency by default.

Owned audit pathAppend-only provenanceHuman-in-the-loopEvidence for SOC 2Evidence for ISO 27001Evidence for ISO/IEC 42001

Keel produces the evidence these frameworks ask for — it supports your certification and compliance work; it doesn’t, by itself, make you certified or compliant.

Why we call it Keel

The backbone, below the waterline.

A ship’s keel is the single beam laid down first — the backbone the whole hull is built on. It keeps the vessel stable so a rogue wave can’t capsize it, and holds a heading so it can’t drift off course. That’s the role here: the load-bearing structure beneath your agents. A wrong output can’t capsize a decision; the system can’t quietly drift. Agentic AI, kept on an even keel.

KEELSERVE LINEAUDIT LINE

Above the serve line — what the user sees. Below the audit line — what’s recorded and provable.

Get started

See Keel govern your hardest agentic use case.

Bring the decision you’d never let an unsupervised model near. We’ll show you the propose-decide-sign boundary, the replayable audit trail, and the cost and control your risk team can actually sign off — on the same backbone that already runs in production.