Posture

Which agents fall under the rules, what you can evidence for each of them, and how much reach they hold that they have never used.

Compliance score
100%
19 of 19 controls evidenced
Open control gaps
0
Nothing outstanding
High-risk agents
1
Mandatory oversight, logging, accuracy
Least-privilege excess
55%
6 of 11 grants never used

The compliance score is not a judgement. It is the share of controls we can put evidence behind, taken from the declared boundary, the tier each agent connects on, and what it has actually called. Agents are ordered worst first, which is the work order.

Engineer Copilot

T1 · InlineMinimal risk

Owner: S. Iyer

100%
Fully evidenced
Data it can reach
Internal
Risk classification

Minimal risk. Internal or public data only, with no path to personal records.

No specific obligations beyond the estate's own policy.

Controls
  • Named accountable ownerISO 42001 · A.3 – accountability

    S. Iyer, Network Engineering lead.

  • Declared boundaryISO 42001 · A.6 – intended use

    Documentation only. Cannot touch network elements, tickets or customer data.

  • Human oversight on consequential actionsEU AI Act · Art. 14 – human oversight

    1 action gated through the approval queue.

  • Automatic record-keepingEU AI Act · Art. 12 – logging

    Tier 1: every call and tool invocation is recorded.

  • Accuracy monitored against a thresholdEU AI Act · Art. 15 – accuracy

    98.4% success rate, above the estate's 96% bar.

  • Boundary reviewed this quarterISO 42001 · A.9 – periodic review

    Last reviewed Jun 2026.

Least privilege5 of 5 grants exercised
0% excess

Measured, not asserted: every grant is checked against the tool calls this agent has actually made. Reach it never exercises is blast radius carried for free.

Every grant on this agent is exercised. No reach carried for free.

Orchestrator

T1 · InlineHigh risk

Owner: R. Mehta

100%
Fully evidenced
Data it can reach
InternalPersonal
Risk classification

High risk. It can write to systems holding personal data. A wrong write is a privacy incident, not a bug.

Human oversight, logging, accuracy and record-keeping are mandatory, not optional.

Controls
  • Named accountable ownerISO 42001 · A.3 – accountability

    R. Mehta, NOC operations manager.

  • Declared boundaryISO 42001 · A.6 – intended use

    Change requests and incident state in ServiceNow. Cannot execute network changes itself.

  • Human oversight on consequential actionsEU AI Act · Art. 14 – human oversight

    2 actions gated through the approval queue.

  • Automatic record-keepingEU AI Act · Art. 12 – logging

    Tier 1: every call and tool invocation is recorded.

  • Accuracy monitored against a thresholdEU AI Act · Art. 15 – accuracy

    97.1% success rate, above the estate's 96% bar.

  • Boundary reviewed this quarterISO 42001 · A.9 – periodic review

    Last reviewed Jun 2026.

  • Disclosed as AI to the people it servesEU AI Act · Art. 50 – transparency

    The agent identifies itself as AI at the start of every interaction.

Least privilege0 of 3 grants exercised
100% excess

Measured, not asserted: every grant is checked against the tool calls this agent has actually made. Reach it never exercises is blast radius carried for free.

Granted, never used

  • Core / OSS-BSS / transport lookups (read)
  • ERP read (15-min cache)
  • Incident and change history read

Revoke these and the agent loses nothing it has ever done.

Governance

T3 · Admin APIMinimal risk

Owner: S. Iyer

100%
Fully evidenced
Data it can reach
InternalFinancial
Risk classification

Minimal risk. Internal or public data only, with no path to personal records.

No specific obligations beyond the estate's own policy.

Controls
  • Named accountable ownerISO 42001 · A.3 – accountability

    S. Iyer, Service delivery lead.

  • Declared boundaryISO 42001 · A.6 – intended use

    Customer-facing SLA communication and status page. Cannot alter contracts or credits without approval.

  • Human oversight on consequential actionsEU AI Act · Art. 14 – human oversight

    2 actions gated through the approval queue.

  • Automatic record-keepingEU AI Act · Art. 12 – logging

    Tier 3: every call and tool invocation is recorded.

  • Accuracy monitored against a thresholdEU AI Act · Art. 15 – accuracy

    99.2% success rate, above the estate's 96% bar.

  • Boundary reviewed this quarterISO 42001 · A.9 – periodic review

    Last reviewed May 2026.

Least privilege0 of 3 grants exercised
100% excess

Measured, not asserted: every grant is checked against the tool calls this agent has actually made. Reach it never exercises is blast radius carried for free.

Granted, never used

  • SLA and availability report reads
  • Contract terms read
  • Outage history read

Revoke these and the agent loses nothing it has ever done.