AI assistant for the construction site manager

The pain we solve

The site manager spends the day on site, not in the office. And precisely for that reason they live disconnected from much of the information that already exists about their own site in Business Central. The paradox is stark: the ERP has a live budget, up-to-date progress certifications, consumed work items, contracted subcontractors with their insurance policies and TC1s, signed contracts with their clauses, change orders pending approval. It is all there. But accessing it from the site is slow and clunky, and for trivial questions it ends up being faster to call the technical office than to open the ERP.

The questions the site manager resolves every day are operational and seemingly minor: how much material do I have left on the concrete work item? Is the electrician arriving tomorrow compliant with their TC1? What does the contract say about debris removal responsibility? When does the civil liability insurance of the roofing subcontractor expire? Which sites in the group have unsigned change orders older than 30 days? The answer to each of these exists in BC or SharePoint, but retrieving it breaks the site workflow.

The aggregate cost is invisible but heavy. Every interruption (call to office, email wait, manual SharePoint search) takes the manager's attention away from what they should be doing: coordinating trades, supervising quality, deciding on incidents. An average builder with five site managers active at once loses, on a reasonable hypothesis, two to four manager-hours a day to information access friction. That is several thousand manager hours a year spent searching for data that was already stored.

But the real cost goes beyond time. When the answer does not arrive fast, the manager decides with incomplete information or improvises. Materials ordered in excess just in case. Subcontractors entering to work without anyone actually verifying their TC1. Change orders executed unsigned, causing problems at certification. Contract clauses the manager does not remember ending up in dispute with the customer. The quality of on-site decision falls, and it shows in the project final result.

What the AI does here

The core of the case is a conversational assistant embedded in the dvproject mobile app. The site manager opens the app, taps the assistant icon and asks in natural language, by voice or text. The question reaches the Azure OpenAI Service pipeline through Copilot Studio, which routes the query to the relevant data sources: Business Central via REST API for the live figures, RAG over the site documentation (specs, contracts, measurements, progress certifications) for clause questions, and the app local database for queries that can be resolved from cache.

The assistant has site context. It knows which site the user is on (by geolocation or active selection in the app), knows which subcontractors that site has under contract, its budget chapter, its responsible technical director and its current phase. When the manager asks "how much material do I have left on item X", the assistant understands they mean the active site, looks up the item in the live budget for that site in BC, calculates the balance based on registered production entries and delivery notes, and answers in natural language with the figure, the unit of measure and an estimated end based on recent consumption.

For documentary questions (what does the contract say about X? what does the technical specification require on Y?), the assistant uses Retrieval Augmented Generation over the site documentary repository. The documentation is indexed with embeddings and vectorised for semantic search. The assistant locates the relevant excerpt, quotes the clause verbatim, indicates the page or section of the original document and proposes an interpretation. The interpretation is guidance: the legal decision still belongs to the technical director or the builder's legal department.

The entire infrastructure runs inside the company's Azure tenant. Contractual documentation, which is sensitive in construction, does not leave to public models. Embeddings are generated and stored on-tenant. Assistant responses are logged with complete traceability: which site, which user asked what, which document the assistant consulted, what it answered. Useful both for internal audit and for resolving subsequent discrepancies with the customer.

The default operating mode is read-only. The assistant answers and proposes drafts but does not write to BC without an explicit action by the manager in the standard forms. This is a design decision: on site, where the consequences of a mistake can be material, we prefer a reliable assistant with human confirmation over an autonomous assistant with risk.

Before and after

How we deliver

When this case is NOT a fit

Some scenarios make the case lose meaning or render it unviable. We say it directly.

• If you do not use dvproject or dvproject-construccion. The assistant relies on the site structure of the dvproject suite. With a different project module we can talk, but the value proposition drops sharply.
• If your site documentation is on paper or in loose files on a local disk. The documentary RAG part requires an indexable digital repository, ideally SharePoint or equivalent. The digital archive needs to be sorted first.
• If your site managers do not carry a corporate smartphone. The assistant lives in the dvproject mobile app. Without a corporate device and data plan, the case is not consumed. The manager's standard equipment needs to be standardised first.
• If you expect the assistant to replace the technical director's presence. The assistant answers operational questions; it does not make critical technical decisions. Supervision by the technical director remains as necessary as before.