Olivier Dreyer

Strategy · Operations · Transformation

How I bring AI into an operation I am accountable for

I designed and delivered Compass as a working case study of how AI can move operational work from evidence to action — without taking authority over business facts or decisions.

15+ years across FinTech & Payments, SaaS and HealthTech.

The simple operating model

  1. Evidence

    Operational data, documents and events

  2. AI proposes

    Interprets evidence and suggests action

  3. Human decides

    Judgement, ownership and approval remain human

  4. Governed action

    Only approved actions reach existing systems

Business rules, auditability and decision authority remain outside the model.

A governance decision I had to make

When Compass could not truthfully record the origin of an AI-proposed action, it blocked the action rather than weaken the audit trail.

Governance should be capable of saying no even when the technology can say yes.

What this demonstrates

See the case study ↓

Case study

Compass is a private working case study. The examples below use synthetic data, but the operating patterns and control mechanisms shown are implemented.

From operational exception to accountable action

A detected exception does not become an AI decision. Compass separates evidence, proposal, approval and execution so accountability remains explicit.

  1. Exception identified

    Material adverse variance

    A deterministic rule identifies an exception from authoritative evidence.

  2. Response proposed

    Create a management action

    Policy validation: valid

    AI proposes a bounded response using the available evidence.

  3. Human decision

    Owner: Operations Lead

    Priority: High

    Due date: 07 Oct 2026

    Ownership, priority and timing remain explicit human decisions.

  4. Controlled execution

    1 step approved

    Approval does not change the authoritative record. Execution is a separate action through an existing governed service.

The model proposes what could be done. Policy determines what is permitted. A person determines what will be done.

Illustrative representation based on implemented Compass capabilities. Synthetic data.

Seeing how work actually flows

Recorded operational events can be reconstructed into observed workflow behaviour — showing where work moves, where people intervene and where escalation occurs.

9 cases observed · 38 projected events · 8 observed activities · 9 observed transitions

Observed, not assumed

The workflow is reconstructed from recorded evidence rather than a predefined process diagram.

Human decisions change the path

Acceptance, editing and rejection remain visible as distinct outcomes.

Escalation remains traceable

Local and escalated AI review appear as explicit transitions rather than hidden model behaviour.

Recorded timing shows sequence, not necessarily processing duration.

Illustrative representation based on implemented Compass capabilities. Synthetic evidence.