The following use cases are an illustrative example per sector. Each snapshot shows what the agent evaluates before acting, what it executes autonomously when it can do so with guarantees, and what the analyst receives when it cannot.

Prescriptive Decision AI · Security & Computer Vision
Retail · Security
Shoplifting alert triage — computer vision

Computer vision models detect potential shoplifting and generate alerts that a security operator must triage. The problem is not detection: it is alert volume without context and the asymmetric cost of being wrong. In retail security, the critical error is not letting a theft slip — it is intervening incorrectly on a legitimate customer. A missed theft costs €45–€120 in goods. A wrong intervention can cost €5,000–€50,000 in complaints and EU AI Act regulatory exposure, plus reputational damage no P&L captures. AyGLOO adds an agentic layer on top of that model with a precise decision function: optimise expected intervention cost at store and network level — not just a single-alert score. The agent explicitly chooses between three strategies based on expected cost in the specific segment: intervene (via escalation), verify first, or do not act. That choice is not a filter — it is a business decision repeated hundreds of times per day.

1
The agent escalates with high priority when the behavioural sequence is complete and ISA confirms reliability in the segment. It requests clip review when the signal is ambiguous. It archives without consuming operator time when the signal is weak. → Operator acts only when certainty justifies intervention risk
2
The visual attention map proves the alert was triggered by behaviour, not appearance — the key regulatory argument under the EU AI Act and in any customer complaint. → Regulatory defence automatically documented per alert
3
ISA detects segments with the highest false positives — low night lighting (31% FP), self-checkout zones — before the operator acts. In those segments the agent blocks direct escalation and requests visual verification. → FP in low light: from 31% without ISA to <8% with guided verification
4
Fairness automatically verifies the alert is not correlated with appearance attributes — and full EU AI Act traceability (Arts. 9, 13, 14, 15) is generated with no extra team effort. → Audit trail ready for regulatory inspection on every decision
Today: what a security operator typically receives
Alert #SL-20260218-1042 Confidence: 87% · HIGH Store: Madrid Centro · Camera: Aisle-7-East · No additional context
With AyGLOO. Same alert, fully enriched
Alert #SL-20260218-1042 · Aisle-7-East Confidence: 87% · Priority: HIGH The agent acts
XAI · Twin Behavioural sequence detected CV Visual attention map ISA Model reliability in this segment Fairness Demographic bias check CF What would change the decision Econ Expected cost per action Action Agent decision PDF Regulatory traceability
1. Which behavioural sequence triggered this alert. Twin model (depth 7 · fidelity 96.3%)
100%IF concealment_gesture = true AND item_transfer_inside_clothing = true AND shelf_dwell_time > 4× segment_mean → Complete sequence detected: agent escalates with high priority. Operator confirms before intervening.
79%IF anomalous_exit_trajectory = true AND concealment_gesture = true without confirmed transfer → Likely but incomplete signal. Agent requests review of the previous 30 seconds before escalating.
58%IF anomalous_exit_trajectory = true with no other factors → Weak signal in isolation. Agent archives automatically. Intervention cost exceeds the item value at risk.
The agent acts differently by pattern: escalate when the sequence is complete, request verification when incomplete, archive when isolated. The 58% rule is most valuable operationally: it makes explicit that intervening on a weak signal destroys more reputational/regulatory value than it protects.
XAITwin
2. Visual attention map. Which image regions activated the model
Activation concentrated in three spatial zones: hand-to-clothing transfer (38%) · shelf interaction boundary (24%) · exit-trajectory corridor (14%) · background customers suppressed (<4% each) · Spatial attention confirms the model anchored on behaviour sequence, not physical appearance.
This is the most differentiating module in CV: it explains not a numeric feature, but where the model attended in the scene. It demonstrates in an audit that the alert was generated by conduct, not appearance — a key defence for any complaint.
CV
3. Model double-check in this segment (ISA)
Segment: "interior aisle, normal lighting, daytime" · False positive rate: 14% · Accuracy: 86% · No drift detected in the last 8 weeks · Sufficient reliability to escalate: agent sets high priority.
Highest-uncertainty segment: low night lighting (18:00–22:00) · false positives rise to 31% due to camera gain/shadow interference · in that window the agent blocks direct escalation and requests visual clip verification before notifying the operator.
ISA is the double-check before escalation: it confirms the model behaves well for this specific lighting/angle/time. A red ISA blocks escalation even if Twin says 100% — because with 31% FP, intervening without verification is economically irrational.
ISA
4. Demographic bias verification (Fairness)
No demographic bias signals on this alert: invariance checks across skin tone, age, and gender completed · Alert generated exclusively by behavioural features, with no contribution from appearance attributes · Alert rate by demographic profile within EU AI Act Art. 10 thresholds · Verification record generated automatically for the audit file.
In behavioural surveillance, Fairness is not optional: the EU AI Act requires high-risk systems to demonstrate bias-free operation per decision. If Fairness detected correlation with appearance, the agent would not escalate — it would raise a model-review alert.
Fairness
5. CF. What would need to change in the scene for the alert to disappear
Remove concealment gesture: confidence drops from 87% to 41% (below alert threshold) · Exit trajectory alone without gesture: 38%, no alert · Decision driven by gesture + transfer sequence, not location or appearance.
CF confirms the decision can be reversed by removing behaviour, not changing who the person is — a direct regulatory argument. It also gives the operator the exact causal chain needed to document intervention or non-intervention.
CF
6. Economic decision function. Expected cost by action (segment-based)
Action
Cost if FP (14%)
Cost if FN
Net expected cost
Intervene without verifying
€800–€50,000
complaint + regulatory exposure
€0
High · irrational with 14% FP
Verify + intervene ✓
€0
human confirmation removes FP
€0
minimal verification delay
Minimum
Do not act (weak signal)
€0
€45–€120
item value
Low · rational for weak signals
→ Agent chooses verify + intervene: removes FP before acting · in night segment (31% FP) it blocks direct escalation — expected cost of intervening without verifying exceeds any item value at risk
Cost asymmetry is the key: a false positive can cost up to 400× the item value at risk. The agent does not optimise detection rate — it optimises expected action cost given segment reliability. That makes “do not escalate” economically explicit, not an omission.
Econ
7. Agent decision
The agent escalates with high priority — operator confirms before intervening
Complete sequence detected, ISA confirms 14% FP in this daytime segment, Fairness clear · Agent notifies operator with full file: attention map, CF, confidence level, regulatory traceability · Operator reviews the marked 30-second clip and confirms intervention · Human confirmation record generated automatically per EU AI Act Art. 14
or, if the signal is likely but incomplete (79%) or ISA flags night segment (31% FP)
→ Agent requests clip review: automatically extracts the 30 seconds prior to the alert · highlights relevant attention areas · operator verifies before deciding to escalate · no intervention without visual confirmation of transfer
or, if the signal is weak and isolated (58%)
→ Agent archives automatically: closes the alert without consuming operator time · logs as likely false positive for segment retraining queue · expected intervention cost exceeds item value under any 58% scenario
The agent does not filter alerts — it chooses among three strategies with distinct economic consequences. Escalating without verification at night can cost more than the item at risk in a single incident. Archiving a weak signal accumulates tolerable shrinkage. The correct choice depends on segment reliability, not isolated confidence. Multiplied by hundreds of alerts per day in a store network, this is a P&L decision. EU AI Act Art. 14 requires human confirmation before intervening — the agent does not replace it, but determines evidence, urgency, and expected cost delivered to the operator.
⚠ Human confirmation is required before any intervention on a customer. EU AI Act Art. 14 — mandatory human oversight for high-risk behavioural surveillance. AyGLOO generates the record of that confirmation automatically for each alert.
ActionPDF

AyGLOO integrates on top of any existing computer vision model without accessing underlying code. Video data stays within the client perimeter. Illustrative example — deployments adapt to each client.

Estimated impact · stores with active alerting
−78%
Wrong interventions in low-light segments
From 31% FP without ISA to <8% with agent-guided verification · complaints and regulatory exposure removed in that segment
0
Alerts processed without regulatory traceability
EU AI Act Arts. 9, 13, 14, 15 covered automatically per alert · audit trail ready with no extra effort
+40%
Operator triage capacity
Agent auto-archives weak signals · operator handles only alerts where judgement adds real value