Werkborg
NederlandsBook a call

Human oversight

See exactly where AI stops and a person decides

This page describes the proposed control model from observation and drafting to possible controlled execution. Every external action remains blocked in the current local evidence.

Intended delivery

Implementation objective: a control model with a named owner, concrete approval rule and reviewable audit trail.

Evidence required before delivery

  • Current local evidence: Service Intake adapter 1.1.0 with a required human role.
  • Frozen tests: 150 calibration and 120 holdout records.
  • Zero external actions; L2/L3 execution, response times and customer roles still require implementation and acceptance evidence.

Authority boundary

What this solution explicitly does not promise

  1. 01

    An approval button is not sufficient without an authorised role, context and response time.

  2. 02

    High-impact decisions always remain outside autonomous execution.

Authority is technical policy

Intended authority levels; owner per implementation

A0

Observe

Reads only approved sample or source data; it does not yet create a draft action.

A1

Prepare

Structures and checks information and creates a draft; a person decides.

A2

Limited execution

Design — not executed locally. Only reversible, pre-authorised actions within agreed limits.

A3

Material action

Design — not executed locally. Only after human approval for that specific transaction; never as blanket permission.

This is a design model, not evidence of production execution. Each implementation records the owner, primary approver, backup, maximum wait, escalation, stop threshold and mandatory exception reason for every action.

Next step

Start with one provable workflow.

No generic AI presentation. Bring volume, systems, exceptions and the human decision-maker.

Watch an approval flow

Logical next step

SolutionTurn a service request into a controlled work-order draftThe tested local adapter structures entered report text, identifies the customer, location, asset and issue, flags missing information and creates a reply and work-order draft only. Email, form and target-system connectors are not yet proven.TrustClaim security controls only with implementation evidenceThis page defines a proposed review structure for identity and access, encryption, secrets, development, backup and incident response. No production environment has yet been verified as evidence that these controls operate for Werkborg.MethodSee what the local Service Intake test does and does not proveThe Proof Lab shows three frozen scenarios with an evidence ID, version, date, input class, expected outcome, local outcome and explicit evidence boundary, alongside the interactive local control flow.