Email and documents
Customer email to draft quote
Classifies inbound email, extracts line items, matches them to material and pricing history, and drafts the quote for review. Preparation time went from hours to minutes.
March to September 2026
Systems in production or pilot on live ERP, CRM and customer data. They are internal, so they are described by function rather than by name.
Email and documents
Classifies inbound email, extracts line items, matches them to material and pricing history, and drafts the quote for review. Preparation time went from hours to minutes.
ERP integration
Signed CRM quotes become ERP sales quotations without re-keying. Credit holds and master-data gaps go to a review queue. One engine, configured per division.
Assistant
Answers plain-language questions from live ERP, CRM and a document corpus indexed per brand, with citations. Tool access is enforced per user on the server. In pilot with about 20 users.
Audit
Compares what the customer specified, what was quoted and what was booked. Every artifact is kept as a timestamped snapshot, and discrepancies are caught before submittal.
Vision
Detects callouts and opening tags on shop drawings and validates them against schedules. The prototype measures 80 to 100% recall on real projects.
Sales intelligence
Classifies filings, permits and interconnection queues to surface large construction projects 6 to 18 months before a request for proposal. A person decides what becomes a lead.
Service desk
Predicts category, priority and likely resolution for incoming tickets, and runs in shadow against analyst decisions before any automation.
ERP
A configure-to-order configurator that generates bills of material, routings and pricing from a product specification, connected to the same integration layer as the assistant.
Data
Co-leading an assessment of data ownership and use-case readiness, measured against the procedure corpus and live transactional data. Each gap has an owner.
Model Context Protocol servers expose ERP, CRM, helpdesk and product data as named tools. Per-user credentials, gated writes, read-only services where required.
Enterprise agreement with a data-processing contract, zero retention, per-project keys and a central usage log. Two vendors compared blind on a hundred real documents.
Results below 70% confidence go to a review queue. Model output is stored beside the human answer as ground truth.
Every assistant turn traced with tools, tokens, latency and cost. An evaluation set of roughly 230 golden questions, scored on tool evidence and attribution.
Frontier API, cloud-hosted models where tenancy matters, and a self-hosted open model on a single GPU. Backend selectable per user.
New engineers shipping in five days, one project template and registry, a deny-first policy for coding agents, prompt anonymization and a tracked secrets inventory.