"Can AI actually search ten years of our project files and give engineers a real answer?"

Institutional knowledge stored in SharePoint and shared drives is a tax that compounds quarterly. Senior engineers spend hours per new project hunting for the right precedent. Junior engineers cannot find what they do not know exists. The cost looks invisible on the income statement and shows up everywhere on the timesheet.

Through 2024 and the first half of 2025, retrieval augmented generation over real engineering project data was a research demo. The accuracy was not there. The citations were inconsistent. The systems hallucinated when the answer was not in the source set. By mid 2025, several things converged. Embedding quality improved. Retrieval frameworks matured. Models stopped guessing when the context was thin. RAG over a firm's own files became reliable enough to put in front of paying engineers without an apology.

September 2025 was the month that conversion happened at Alpha Structural, California's largest hillside structural repair and geotechnical engineering firm in its niche, with a thirty year archive of soils reports, foundation designs, and bid histories sitting in SharePoint. Pilot scoped, deployed, used, and validated. Two engineering teams in production. The receipts below.

How long does a desktop knowledge assistant deployment take?

Five to seven weeks for a focused pilot scoped to a single business unit or two engineering teams. The work splits roughly into four blocks. Ingestion of historical files takes one to two weeks depending on volume and how clean the storage hierarchy is. Retrieval tuning against real engineer queries takes two weeks because we run the assistant against the actual prompts the team uses, not synthetic ones. Security review and tenant deployment overlap the back half. Rollout with training closes the engagement. After that the system runs in production with light maintenance.

What does the engineer actually see on screen?

A search box that accepts plain English questions. A results pane that lists matched historical projects ranked by relevance, with a one line summary of why each project matched. A detail view that surfaces the original drawings, scope notes, final cost, photos, and outcome for any project the engineer clicks. Citations link back to the source file in SharePoint so the engineer can open the original record without copying or rekeying. The interface looks like a focused search tool because that is what it is.

How does the firm protect proprietary project data?

The retrieval index sits inside the firm's tenant. The model never trains on the firm's data. Access controls inherit from SharePoint so an engineer only sees what they were already permitted to see. Audit logs record every query and every retrieved document. The deployment fits inside an ISO 42001 compliant governance framework so legal and IT can sign off without rewriting the policy stack. The detailed governance posture is in the ISO 42001 Governance Guide if your compliance team wants the full read.

Who needs to champion this internally for it to work?

An internal AI champion with cross functional authority. At Alpha Structural the EVP of Marketing took that role because she had visibility across engineering, estimating, and operations, and the political mandate to push the rollout. The technology choice matters less than the champion. Without one, the deployment becomes another search tool nobody opens. With one, the assistant becomes the default first stop for any engineer starting a new scope. We screen for that champion in the first discovery call and walk away if the seat is empty.

How AI bid review against historical project data protects margin on technical projects. The same firm from this issue runs every new bid through an agent that flags assumptions out of family with the firm's own actuals. The estimator still owns the number. The agent makes sure they see what they would have otherwise forgotten.

Subscribe below and the October issue lands in your inbox the day it publishes.