When people hear "AI in construction," they usually think of generative AI: design tools that produce building plans, chatbots that write RFIs, or vision systems that detect safety hazards on job sites. These are real applications. They're not what builders need most.
What builders need most is an AI that watches their project finances continuously and alerts them to problems before those problems become expensive. That's a different kind of AI, running on a different kind of data, doing a different kind of work.
The Event Stream Model
BLT's AI doesn't analyze reports. It reads an event stream — a continuous log of every action taken in the system. Every expense submitted. Every receipt OCR'd. Every budget line modified. Every hour logged. Every subcontractor invoice recorded. Every photo uploaded.
Each event is structured: event type, entity type, entity ID, who did it, project, metadata. The AI reads this stream and builds a live model of every project's state. Not a snapshot. A continuous model that updates in real time as data flows in.
What the AI Actually Watches For
Budget threshold alerts — not just "category X is at 85%" but "category X is at 85% with $2,400 in pending expenses that will push it to 93% when approved." The AI knows about pending state, not just current state.
Cross-project anomalies — comparing spending patterns across concurrent builds at equivalent phases. If Foundation Materials on Project 3 is running 40% higher than the same phase on Projects 1 and 2, that's signal. It might be justified (different soil conditions, different slab spec). It might not be. Either way, the manager should know.
Pace alerts — if hours are being logged but no expenses are appearing in a phase that should be generating material costs, something's off. If a phase has been active for a week with no progress photos, the AI notices. These aren't accusations — they're prompts to check in.
Subcontractor drift — cumulative invoice tracking against original estimates, with alerts at configurable thresholds. When a sub crosses 80% of their estimate with significant scope remaining, the manager needs to know before the final invoice arrives.
Natural Language Queries
Beyond monitoring, BLT's AI answers questions in plain English. "What's our average cost per square foot for foundation work across all three projects?" "Show me all Home Depot spend in the framing phase this month." "Which budget categories are most likely to overrun before the project closes?"
For each query, a context builder assembles the relevant slice of data — project state, budget snapshots, recent events, historical comparisons — and sends it to the AI. The response is direct and actionable, not a wall of data the manager has to interpret.
What the AI Doesn't Do
The AI doesn't make decisions. It doesn't approve expenses, modify budgets, or authorize purchases. Every action in BLT requires a human to confirm it. The AI's role is to ensure the human has the information they need to make good decisions — and to surface the information they didn't know they needed.
The builder stays in control. The AI makes sure they're informed.