AI Strategy
80% SQL. 20% AI.
Our operational AI strategy is built around cost discipline and explainability. Most questions are answered by SQL. LLMs handle genuinely novel scenarios. Every decision is logged, every cost is tracked, and every organisation controls their budget.
SQL-First Architecture
Classification, not just prompting.
Every question is classified into a predefined operational category. 80% are answered by SQL with zero LLM tokens. Only novel questions route to an LLM.
Classification first
Every question is classified into one of 30+ predefined categories (inventory, occupancy, staffing, compliance, etc.). The classifier is a lightweight ML model, not an LLM.
SQL-first execution
Roughly 80% of operational questions are answered by templated SQL queries with zero LLM tokens. No hallucination risk, deterministic output, millisecond latency.
LLM fallback
Only truly novel questions route to an LLM. We use Vertex AI, Gemini, or Azure OpenAI depending on latency, cost and privacy requirements for that specific query.
Reasoning trace
Every decision is logged: question, classification, routing decision (SQL vs LLM), result, latency and token usage. Full audit trail for compliance.
Cost Discipline
Token caps, not token sprawl.
Every organisation controls its AI spending. Monthly budgets, daily per-user limits, and real-time monitoring prevent runaway LLM bills.
Per-org monthly cap
Each organisation can set a monthly token budget. Once exceeded, LLM queries are rejected with a human-readable "cap reached" message.
Per-user daily limit
Individual staff members have daily query quotas to prevent abuse and encourage efficient question design.
Global rate limiting
System-wide token spend is monitored. If we're trending towards overage, organisations are proactively notified.
Usage transparency
Every query shows provider, model, latency, exact token count and cost. Monthly invoices break down spending by organisation and query category.
Multi-Provider
Choose your AI, per workload.
Vertex AI, Gemini, Azure OpenAI, or any OpenAI-compatible endpoint. Each organisation decides which provider handles which types of questions.
Vertex AI (GCP)
Primary provider. Metadata-based auth (no API keys needed on GCP), streaming support, and direct integration with GCP's broader AI ecosystem.
Gemini API
Fallback for non-GCP workloads. Used for image understanding in cleaning verification and equipment inspection flows.
Azure OpenAI
For enterprises with Azure commitments. Provides GPT-4 Turbo and fine-tuned model support.
OpenAI-compatible
Any OpenAI-compatible endpoint (Ollama, LM Studio, custom fine-tuned models). Organisations can bring their own LLM infrastructure.
Safety by design
Every LLM interaction is guarded.
- Domain-locked system prompts per workload (no jailbreak attempts)
- Prompt injection detection via token analysis
- Input length limits to prevent token explosion
- PII sanitisation in error messages and logs
- Output validation — LLM responses are type-checked against expected schema
- Fallback to SQL or human review if LLM output fails validation
Real Scenarios
How Intelligence works in production.
Occupancy forecasting
WTMS classifier recognizes occupancy questions. SQL queries fetch historical booking data, current reservations, and staffing levels. Novel patterns route to Vertex AI for predictive insights.
Equipment maintenance
Electrical AI logs failures and maintenance cycles. Classifier routes to SQL for recent incidents and cost history. Anomalies are flagged for LLM analysis.
Guest service resolution
Menu complaints are classified by type (food quality, price concern, menu availability). SQL joins with POS and inventory. Unresolved complaints can route to LLM for policy exceptions.
Workforce analytics
WTMS tracks attendance, task completion and performance metrics. Classifier routes to SQL for KPIs, trends and compliance reports. Novel performance questions route to Gemini.
Transparency
Every query is auditable. Every cost is visible.
We believe AI infrastructure should be explainable. You can see which questions hit SQL (free) versus LLM (paid). You can review the classification decision, the system prompt, and the actual LLM response for any query. Monthly analytics show spending trends and usage patterns by department.
This transparency builds confidence. It prevents surprise bills. It helps you optimise your query patterns. And it's a requirement for any compliance framework that cares about AI governance.
Let's build something
See Intelligence in action.
We'll walk you through how a real operational question is routed, classified, and answered—with zero LLM tokens when possible.
