AI Native · Public Model Cards · No Facial Recognition
AI guard tour software, documented end to end
Production-grade AI for guard tour: anomaly detection (74-89% precision), predictive routing, incident classification. Per-site model isolation. Every model published with training data, evaluation methodology and limits.
Built on the PatrolTech platform · GDPR Article 22 by design · Customer data never leaves your tenant
Trust signals procurement teams cite
- Model cards public
- GDPR Article 22 compliant
- No facial recognition
- Per-site model isolation
- Customer data never leaves your tenant
AI for guard tour without the snake oil
Most security software vendors market 'AI' as a marketing layer. guardtourai.com is built differently: every AI capability ships with a public model card, documented training data, evaluation methodology and known limits. Anomaly detection, predictive routing and incident classification run on production-grade models with per-site isolation — your patrol patterns never train another customer's model. The platform is GDPR Article 22 compliant by design: every AI alert requires a human review step, and we explicitly do not build facial recognition or behavior scoring on individuals.
Production AI capabilities
Three production models, two in active development. All shipped with public model cards.
Anomaly Detection v2.1
Isolation Forest + temporal LSTM. Flags patrols deviating from learned site patterns. Per-site model isolation. 4-week minimum training data. Production precision 74-89% across 12 deployments.
Incident Classification v1.3
Fine-tuned RoBERTa. Auto-categorizes incident reports into 47 standard categories. Production precision 91%. Supports English, Spanish, French, Portuguese, German, Italian.
Predictive Routing v1.0
Constrained optimization with learned weights. Suggests checkpoint order optimizing time, risk coverage and patrol fatigue. Production beta. Reduces average patrol time 8-14%.
Photo Analysis v0.8 (development)
Anomaly flagging on patrol photos (fire, water intrusion, intruders). Public beta target Q3 2026. Will ship with the same model card discipline as production models.
Speech-to-Incident v0.5 (development)
Voice-to-text incident logging. Multilingual ASR with on-device inference for offline operation. Target Q4 2026.
Public model cards
Every shipped model gets a public model card: training data sources, evaluation methodology, known precision, recall, false positive rate, demographic disparities, and limits.
What we explicitly do not build
The procurement teams that buy AI guard tour cite this list more than the feature list.
No facial recognition
Privacy and bias risks make facial recognition incompatible with the trust contract we offer customers. Not on the roadmap.
No behavior prediction on individuals
Models work on patterns, not people. We do not score guards or contractors as individuals; we surface patterns at the patrol-route or site level.
No automated decisions with operational impact
GDPR Article 22 by design. Every alert requires a human review step before it triggers an operational action. AI augments operators; it never replaces them.
No black-box claims
Every customer can request the model card, the precision metrics and the training data sources. If we cannot explain it, we will not ship it.
No cross-customer training
Per-site, per-tenant model isolation. Your patrol patterns and incident reports never train a model used by another customer.
Compliance and trust contract
AI guard tour is held to a higher bar than the rest of the stack. guardtourai.com aligns with the most cited frameworks for AI procurement.
- GDPR Article 22 — automated decision-making, human review by design
- EU AI Act — risk classification, transparency obligations, conformance documentation
- NIST AI Risk Management Framework — model governance and risk register
- ISO/IEC 42001 — AI management system
- Model cards published per release — evaluation, limits, demographic disparities
Where AI guard tour produces measurable results
Three production deployments where AI features moved an operational metric beyond rule-based automation.
Data centers
Anomaly detection on cage rounds catches deviations from learned 4-times-a-day patterns.
- 74-89% precision in production
- Per-site training, no cross-customer leakage
- Integrates with SOC 2 evidence exports
Critical infrastructure
Predictive routing optimizes coverage of high-risk perimeter segments under changing threat profiles.
- 8-14% reduction in patrol time
- Constraint-based, fully auditable
- Operator can override every suggestion
Financial services
Incident classification routes branch incident reports to the right second-line team in 47 categories.
- 91% production precision
- Multilingual: EN, ES, FR, PT, DE, IT
- Reduces triage time by 60%
Frequently asked questions about AI guard tour software
How is guardtourai.com different from a guard tour platform that markets 'AI features'?▼
We publish a model card per shipped model, with training data, evaluation methodology, precision/recall metrics and known limits. Most vendors marketing AI features do not publish any of this. If a procurement team asks for the technical documentation behind an AI claim, we hand it over the same day.
Does the AI use my patrol data to train models for other customers?▼
No. Per-site, per-tenant model isolation is foundational. Your anomaly detection model is trained exclusively on your site's data. Our incident classification model is fine-tuned on a public corpus plus consented anonymized data; we never use customer data without an explicit opt-in.
Is the platform GDPR-compliant for AI features?▼
Yes. GDPR Article 22 (automated decision-making) is enforced by design: every AI alert requires a human review step before any operational action is taken. We provide a Data Protection Impact Assessment template and a Record of Processing Activities entry covering AI features.
Do you offer facial recognition or biometric guard verification?▼
No. We have decided not to build facial recognition for privacy and bias reasons. Guard verification uses individual login credentials, hardware tokens or biometric sensors on the device (where local law permits) — never a facial recognition database.
How do you measure AI model precision in production?▼
Each model has a customer-visible dashboard showing rolling 30-day precision, recall and false positive rate. We publish ranges across deployments (e.g. 74-89%) instead of single benchmark numbers because real-world precision varies by site complexity. Customers can drill into per-incident outcomes.
What happens when an AI model produces a false positive?▼
Every false positive is loggable as feedback by the operator. The feedback loop retrains the per-site model on a quarterly cadence with explicit approval. We never silently retrain models — a customer always sees what training data is used.
Can I see the training data sources before adopting?▼
Yes. Each model card lists training data sources at the source-and-license level. For the per-site anomaly detection model, the source is exclusively your own site's patrol history, with the schema documented. For the incident classification model, the public corpus is named and the consented private corpus is described in aggregate.
AI guard tour without the marketing fog
Read the model card, talk to the team, run a 30-day pilot. No facial recognition, no black-box claims.
Public model cards per release