EcoKure Enterprise Gateway Pilot stage

Enterprise AI control
without rebuilding your platform.

EcoKure is designed to integrate with the systems organisations already run. It adds an independent layer for verification, authority decisions and auditable evidence at the action boundaries that matter.

Discuss an enterprise pilotCompare integration paths
Gateway · request flowIllustrative
How the Enterprise Gateway sits between AI and your systems An AI agent, a copilot and a workflow send requests to the EcoKure Gateway. It checks identity, scope, policy and evidence. Two requests pass to the payments API and the tools; one, aimed at the records system, is stopped at the gateway. AI agent Copilot Workflow Payments API Records Tools EcoKure Gateway Identity Scope Policy Evidence every call, recorded
Allowed after checksStopped at the gateway

Integration paths

Three ways to connect.
Pick the lightest one that works.

Path 1

API / SDK integration

Developers call EcoKure verification and authorisation checks at chosen points inside existing software.

  • You keep your application, model and workflow
  • Checks go only where they are needed
  • Enforcement is as strong as the call sites you add

Integration kit available · pilot stage

Path 2

Enterprise Gateway

EcoKure sits between AI-driven workflows and sensitive execution interfaces such as tools, APIs and consequential operations, and mediates every call through it.

  • One place to enforce policy across agents
  • Denials bind every route through the gateway
  • Only enforces if direct bypass paths are removed

Mediation tested in the lab · enterprise gateway in development

Path 3

Observation & evaluation mode

EcoKure evaluates selected workflows alongside production and records what it would have decided, with no change to execution.

  • The lowest-risk way to start
  • Compare EcoKure decisions with what actually happened
  • Does not prevent anything. Observation only.

Usable for pilots

Architecture

Where the boundary sits,
and what enforcement needs.

Where EcoKure sits Existing AI models, agents and enterprise applications send proposed actions to the EcoKure verification and authority boundary. It checks policy, identity, evidence and approval, then allows, denies or escalates the action before it reaches existing APIs, databases and tools. Every decision is recorded. YOU KEEP AI models Agents & copilots Business applications Workflows & scripts proposed action EcoKure boundary verification & authority Policy Identity & scope Evidence Approval Allow · Deny · Escalate Decision record · replayable only permitted actions pass YOU KEEP Enterprise APIs Databases Tools & services Payments & records
Needed for enforcement: the decision and approval steps sit in the execution path, and direct routes that bypass EcoKure are removed. Works in evaluation mode: policy, identity and evidence checks can run alongside production and record what EcoKure would have decided.

Capabilities

What the Gateway provides,
and how mature each part is.

How to read our labels

  • PublishedEvidence you can open, and often run, on this site.
  • Recorded exampleA real run, captured. Not a live query.
  • Internal lab testTested by us. The public write-up isn't out yet.
  • Synthetic scenarioA worked example of the mechanism on made-up data.
  • ConceptA part that still has to be built.
CapabilityWhat it meansStatus
Independent verificationDeterministic checks that don't depend on the model being checked, and that abstain rather than guess.Published evidence
Scoped authority & permissionsEach agent or workflow gets an explicit, narrow capability. Anything outside it is refused.Lab-tested
Approval & escalationHigh-consequence or uncertain actions wait for a named, independent approver.Lab-tested · live approver enrolment pending
Evidence & replayEach decision is recorded with its inputs, policy version and outcome, and can be replayed offline.Published evidence
Agent action containmentAgent tool calls run through a broker and an isolated worker, where supported.Lab-tested on supported platforms
Integration with existing systemsAPI/SDK, gateway and evaluation-mode paths, described in the integration kit.Pilot stage
Customer-controlled deploymentBuilt to run in the customer's own cloud or on-premises environment.Design target · assessed per pilot
Model-agnosticIt checks the action and the evidence, not the model's internals.By design

Not claimed: universal compatibility, completed enterprise certifications, or production deployments. Deployment options → · Integration kit →

Integration boundaries by industry

The same boundary,
in your setting.

Each flow shows the existing system, where EcoKure sits, and the system it protects.

How to read our labels

  • PublishedEvidence you can open, and often run, on this site.
  • Recorded exampleA real run, captured. Not a live query.
  • Internal lab testTested by us. The public write-up isn't out yet.
  • Synthetic scenarioA worked example of the mechanism on made-up data.
  • ConceptA part that still has to be built.
Illustration

Public knowledge

A researched answer, held until a person signs it

Recorded example

HOLDMachine checks passed. Held for human review.
Case detail
The problem
An AI answer combines a sourced fact with a calculation. A confident model is not the same as a checked answer.
Proposed action
Publish "The Eiffel Tower is 330 m, about 1,083 ft" as a reusable answer.
What can be checked independently
The arithmetic is recomputed exactly in your browser; the source sentence is re-read on recompute.
Illustration

Financial services

A payment run that tries to pay an invoice twice

Synthetic scenario

DENYDeterministic match against the paid ledger. Payment withheld.
Case detail
The problem
An automated accounts-payable agent queues an invoice that the ledger already shows as paid.
Proposed action
Pay invoice INV-2207 (synthetic) a second time.
What can be checked independently
Matching rule, ledger snapshot hash and decision record.
Illustration

Healthcare

A clinical assistant reaching outside its patient scope

Synthetic scenario

ESCALATEOut of scope. Request blocked and routed to a named clinician.
Case detail
The problem
An assistant scoped to one patient episode requests a record belonging to a different patient.
Proposed action
Read record P-0093 (synthetic) for a session scoped to P-0041.
What can be checked independently
Scope grant, request and escalation record.
Evidence on this site
None yet. This is a worked example on synthetic data, not a deployment.
Illustration

Energy & infrastructure

A cheaper grid dispatch that fails a contingency check

Synthetic scenario

ABSTAINConstraint could not be shown to hold. No automatic approval.
Case detail
The problem
An optimiser proposes a lower-cost dispatch. Under one line outage the plan overloads a transformer.
Proposed action
Apply dispatch plan D-17 (synthetic) to the next interval.
What can be checked independently
Constraint set, the failing contingency and the abstention reason.
Evidence on this site
None yet. This is a worked example on synthetic data, not a deployment.
Illustration

AI agents

An isolated worker that can propose but cannot commit

Internal lab test

DENYThe worker holds no signing authority. Proposal queued for an independent approver.
Case detail
The problem
A coding agent drafts a change. It must not sign, merge or execute the change itself.
Proposed action
Commit and sign a generated change from inside the sandbox.
What can be checked independently
Capability grant, denied call and signed approval receipt.
Evidence on this site
Tested internally. The public write-up is not published yet.
Illustration

Government

A retrieval agent retrying a denied resource another way

Concept ? not built

DENYThe earlier denial binds every route to the resource, provided no bypass path exists.
Case detail
The problem
After one denial, an information-retrieval agent tries a different tool to reach the same restricted file.
Proposed action
Fetch restricted document R-12 (synthetic) via an alternative connector.
What can be checked independently
Denial record and the second, refused route.
Evidence on this site
None. The gateway route this case needs is not built yet.

Every case uses synthetic data and none is a customer deployment. "Synthetic scenario" means a worked example of the mechanism; "Concept" means a component it needs is still to be built.

Next step

Prove it on
one real workflow.

A pilot starts in evaluation mode on synthetic or sanitised data, and moves to enforcement only once integration and security review allow it.

See how a pilot worksRead the evidence →
Next · See it work Eco Control Tower