Example Governed Decision Record, AI Oversight

Companion artefact, EU AI Act Article 14

Warren Smith, WRKS Holdings Ltd  ·  May 2026  ·  omegaprotocol.org/record/example/ai-oversight/

This example shows how an OMEGA record provides direct, contemporaneous evidence that human oversight occurred before a high impact action was taken by an autonomous system. The setting is a customer service agent that escalated a refund decision sitting above its authorised threshold.

Under EU AI Act Article 14, providers and deployers of high risk AI systems must ensure that natural persons can effectively oversee the system during its period of use. A record of the oversight event, not just a log line, is what proves the obligation was met.

This is a stylised example. It is structurally accurate but not authoritative legal or regulatory guidance, and does not represent any specific firm, customer, or transaction.

GOVERNED DECISION RECORD

Record ID: omega-record/2026-05-14/cs-agent-refund-1187442
Schema Version: omega/1.0
Domain: ai-oversight.customer-service.refund-escalation

═══════════════════════════════════════════════════

CONTEXT

A customer (account #C-884213, retail tier, 4 years
active) contacted the autonomous customer service
agent at 09:42 requesting a full refund of order
#O-2026-04471. The order value is £4,820.00 and
covered a bespoke configuration of professional
audio equipment delivered 11 days earlier.

The customer states the equipment does not match
the specification quoted at point of sale and has
provided a written description plus three product
photographs. The order is outside the standard
14 day return window for non-bespoke goods but
sits inside the firm's published 30 day quality
assurance window for bespoke configurations.

The autonomous agent operates under a published
refund authority schedule:
  - Tier A (up to £500): agent decides
  - Tier B (£500 to £2,500): agent decides with
    second-pass policy check
  - Tier C (above £2,500): agent must escalate;
    the agent may not commit a refund decision
    without a documented human approval

This request is Tier C. The agent's task is
therefore not to decide the refund but to halt,
gather the context that a human supervisor needs
to decide, and record the handover.

═══════════════════════════════════════════════════

DECISION

Approve refund of £4,820.00 to the customer's
original payment method. Initiate return collection
of the bespoke configuration. Flag the account for
a single goodwill credit of £120 covering the
customer's documented courier insurance.

Decision time: 09:58
Decision authority: Sarah Okafor, Senior Customer
Operations Manager, badge #ops-2218.
Agent role: prepared the case, did not commit
the decision.

═══════════════════════════════════════════════════

REASONING

The agent's role was bounded by policy. Its
reasoning is therefore separated from the human
supervisor's reasoning, and both are recorded.

Agent reasoning (advisory only):
  1. Order value £4,820 exceeds the agent's
     Tier B ceiling. Authority schedule requires
     escalation.
  2. Customer's stated complaint type (specification
     mismatch on bespoke goods) falls inside the
     30 day quality assurance window, so the
     timing eligibility test is met.
  3. Photographs submitted by the customer were
     compared against the order specification using
     the agent's vision tool. Two of the three
     parameters listed in the order (channel count,
     output gain configuration) appear inconsistent
     with the delivered item. Confidence on this
     comparison is recorded as moderate, not high.
  4. The agent's recommendation, if asked, would be
     to approve. The agent explicitly did not act
     on this recommendation.

Human supervisor reasoning (decisional):
  1. Reviewed agent's prepared case file at 09:53.
  2. Re-examined the photographs and the original
     specification independently.
  3. Confirmed the channel count mismatch by reading
     the rear panel labelling visible in photo 2.
     This is direct evidence rather than inferred.
  4. Treated the agent's vision comparison as
     supporting only and not as the basis for
     decision.
  5. Concluded that the firm is at fault under the
     published quality assurance terms and a full
     refund is the correct outcome.

═══════════════════════════════════════════════════

EVIDENCE AND ASSUMPTIONS

Evidence used:
  - Order record #O-2026-04471 (specification,
    payment, delivery confirmation)
  - Customer message thread (3 messages, 09:42
    to 09:51)
  - Three customer-supplied photographs, hashes
    recorded
  - Agent vision tool output (parameter comparison
    table, confidence scores)
  - Quality assurance policy in force on the order
    date (policy version 2025-11-03)
  - Authority schedule in force on the decision
    date (schedule version 2026-02-17)

Assumptions explicitly recorded:
  - Customer identity verified through account
    login plus second factor at 09:42
  - Photographs are assumed authentic; no forensic
    inspection was performed
  - Agent vision tool confidence is moderate, not
    high; the supervisor did not rely on it as
    the sole basis for decision
  - No fraud signal was raised by the firm's
    risk system on this account in the prior
    180 days

═══════════════════════════════════════════════════

AUTHORITY

Decision authorised by: Sarah Okafor, Senior
Customer Operations Manager, badge #ops-2218.

Authority basis: refund authority schedule
v2026-02-17, Tier C threshold (above £2,500),
which requires a named human decision before
the action is taken.

Agent system identification:
  - System name: CS-Agent
  - Version: 3.4.1
  - Provider: in-house
  - Risk classification under EU AI Act:
    limited risk, with operator imposed Article 14
    style human oversight controls on financially
    material actions
  - Authority schedule binding the system:
    schedule version 2026-02-17

Article 14 oversight mechanism in use:
  - Hard stop on Tier C value; the agent's tool
    layer cannot emit a refund commit for any
    amount above £2,500 without a signed human
    approval token
  - Human supervisor presented with full prepared
    case file, not a single recommendation button

═══════════════════════════════════════════════════

CONSTRAINTS CHECKED

The following constraints were evaluated before
the decision was committed:

  ✓ Order falls inside published 30 day
    quality assurance window for bespoke goods
  ✓ Customer identity verified at session start
  ✓ No active fraud signal on the account
  ✓ Authority schedule version cited matches
    the version in force on the decision date
  ✓ Human supervisor identity verified through
    SSO plus hardware key
  ✓ Refund amount within the supervisor's
    own delegated limit (£15,000 per case)

  ⚠ Agent vision tool confidence was moderate.
    Flagged for review but does not block the
    decision because the supervisor reached the
    conclusion independently from photo evidence.
  ⚠ Goodwill credit of £120 sits inside the
    supervisor's discretion limit. Recorded
    separately from the refund itself for
    accounting and audit clarity.

═══════════════════════════════════════════════════

DISPUTES

One disagreement was recorded during this decision.

The agent's recommendation (advisory only) was to
approve a partial refund of £3,800, retaining
£1,020 to cover the cost of the bespoke
fabrication that could not be recovered by return.
The supervisor considered this and disagreed: the
firm's published terms place the cost of a
specification error on the firm, not the customer,
and a partial refund would not satisfy that
obligation.

Resolution: full refund approved by the
supervisor. The agent's partial refund proposal is
preserved in the record as a non-binding signal
that the firm may wish to revisit its standing
guidance on bespoke specification errors at the
next policy review. No override of policy was
required because the agent's role here was
advisory.

═══════════════════════════════════════════════════

CONSENT

Customer consent recorded for:
  - Return collection at the customer's stated
    delivery address
  - Use of order photographs as evidence in this
    decision and any related downstream review
  - Communication of decision outcome by email and
    in-app message

Customer consent NOT obtained for (and not
required for current decision):
  - Inclusion of this case as a training example
    for the customer service agent system; the
    firm's policy requires a separate opt-in for
    training data
  - Publication of the case in any external
    reporting

══════════════════════════════════════════════════

HARM TRACE

Anticipated harm chains evaluated:

If full refund proceeds (chosen path):
  - Risk: firm incurs £4,820 cost plus return
    handling
  - Risk: precedent effect on similar future cases
  - Mitigation: published terms already place the
    firm at this risk for specification errors
  - Residual risk: financial, low impact at this
    transaction size

If agent had auto-approved without human review
(forbidden path):
  - Risk: firm violates its own published Tier C
    authority schedule
  - Risk: firm cannot demonstrate human oversight
    for an action above its agent's authorised
    ceiling, exposing it under EU AI Act Article
    14 expectations and consumer law fairness
    duties
  - Residual risk if pursued: regulatory and
    reputational, material

If agent had refused or stalled (rejected path):
  - Risk: customer harmed by delay on a valid
    complaint
  - Risk: firm fails its quality assurance
    commitment
  - Residual risk if pursued: customer detriment,
    moderate

══════════════════════════════════════════════════

OUTCOME

Recorded at 10:11 (13 minutes post-decision):

Refund of £4,820 initiated to the customer's
original payment method, expected settlement
within 3 to 5 business days. Return collection
booked for 2026-05-15. Goodwill credit of £120
applied to the account. Customer acknowledged the
outcome at 10:09.

Pre-action checks at the payment processor cleared
without exception. The supervisor's signed approval
token was consumed by the agent's tool layer,
which then permitted a single refund commit and
no further actions at Tier C without a new token.

The decision is closed with outcome COMMITTED.
No retrospective review flagged. Standard quality
assurance sampling will pick this case up for
calibration purposes alongside other Tier C
escalations in the same week.

══════════════════════════════════════════════════

PROVENANCE

Content hash:
sha256:9c2e8f1a3b5d7c4e6f2b9d1a8c3e5f7b4a6d2c1e9f8b3a5c7d4e2f1b6a9c8e3d5f7b2a4c1e9d6f8b3a

Previous record hash:
sha256:1f8c3a5b7d2e9f4c6a1b8d3e5f7c2a9b4d6e1f3a8c5b2d7e9f4a1c6b8d3e5f2a7c9b1d4e8f6a3c5b2d
(preceding agent oversight record in CS-Agent
escalation chain)

Schema validation: passed (omega-contracts v0.2.2)
Cryptographic seal: valid
Composition: agent_action_records →
             omega_records (canonical envelope)
Human approval token: redeemed once, sealed by
                      ops-2218 hardware key

The substrate paper describing the underlying architecture is available at /substrate/.