A distributed infrastructure provider cut data retrieval time by 90% with a custom knowledge assistant

The company operated at immense scale with data fragmented across thousands of daily status reports, regional spreadsheets, and email chains. Senior management lacked quick access to accurate project snapshots. Getting answers to straightforward questions about project status, budget allocation, or resource utilisation took hours or even days of manual collation by junior staff.

The scale of the problem was hard to appreciate from the outside. The company managed over 1,400 active projects across eight regional divisions. Each project generated a daily status report, a standardised Word document filled in by a site coordinator and emailed to a regional inbox. Over the course of a week, roughly 10,000 of these reports landed in various inboxes across the organisation. Nobody had a complete picture of anything.

Senior leadership faced a specific and recurring frustration. When a board member or external stakeholder asked for an update on a particular project or region, the process of producing an accurate answer involved emailing three or four regional managers, waiting for replies, reconciling conflicting figures from different spreadsheets, and assembling a summary document manually. This typically took between four hours and two full working days, depending on how responsive the regional teams were. For time-sensitive decisions, procurement approvals, resource reallocation, schedule changes, this lag was genuinely costly.

The organisation had previously attempted to solve this with a centralised project management platform, but adoption had been inconsistent. Field teams continued submitting Word documents because that was the established workflow, and the platform became another silo rather than a single source of truth. What leadership needed was a system that could work with existing reporting formats rather than requiring the entire organisation to change its habits.

We built a secure web application with a central dashboard that ingested and unified all project data sources into a single reliable database. An AI-powered assistant allows managers to ask questions in plain English and receive instant, accurate answers with source citations. The system continuously indexes new reports and updates as they arrive.

The project began with a four-week data audit across all eight regional divisions. We catalogued every reporting format, 14 distinct Word document templates, six different Excel tracking sheets, three SharePoint libraries, and two email distribution lists. We mapped the fields that appeared consistently across formats: project ID, location, phase, budget spent, budget remaining, schedule status, and open issues. We then built a normalisation layer that could extract structured data from documents regardless of template version or regional variation.

The ingestion pipeline runs every two hours, monitoring regional email inboxes and SharePoint folders for new documents. It extracts structured data using a combination of document parsing and a fine-tuned extraction model, validates extracted fields against a schema, and writes clean records to a central PostgreSQL database. Documents that fail validation are flagged for manual review with the specific extraction failure highlighted. The pipeline processes an average of 800 documents per day with a 97% straight-through success rate.

The AI assistant is built on a retrieval-augmented generation architecture. When a manager asks a question, "What is the current budget utilisation for the Northern region?" or "Which projects are more than two weeks behind schedule?", the system queries the structured database for relevant records, pulls the most recent status text from matching reports, and synthesises a response with figures, context, and source citations. Every answer includes a list of the specific reports it drew from, with links to the original documents. Managers can drill down from a high-level summary to the underlying source in two clicks.

The dashboard layer provides pre-built views for the most common management queries: regional project health heatmaps, budget burn-rate charts, schedule variance summaries, and open-issue trackers. These update automatically as new reports are ingested, giving leadership a live operational picture without any manual data entry.

Time for senior managers to retrieve project data dropped from hours or days to seconds. Leadership gained real-time visibility into resource allocation across all active projects. The system transformed thousands of static reports into a dynamic, searchable intelligence platform accessible to the entire leadership team.

The immediate impact was on senior management time. Before deployment, the CEO and regional directors spent an estimated four to six hours per week fielding data requests and waiting for manually compiled reports. After deployment, those same requests were answered through the AI assistant in under a minute. The time saving eliminated the downstream burden on junior staff who had previously been tasked with collating data on demand.

The quality of decision-making improved alongside the speed. Managers reported that having access to accurate, current data changed the nature of their operational meetings. Rather than spending the first thirty minutes reconciling figures from different sources, leadership could focus on decisions themselves. Several resource reallocation decisions that would previously have been deferred pending data collection were made in the meeting itself.

The system also surfaced patterns that had been invisible before. Because all project data now sat in a single queryable database, leadership could run analyses that had never been possible: identifying which project types consistently ran over budget, which regional divisions had the best schedule adherence, and which site coordinators repeatedly flagged the same category of issue. These insights fed into a structured quarterly review process that had not previously existed.

Adoption was near-universal within the first month. Because the system worked with the existing reporting workflow, site coordinators continued submitting the same Word documents, there was no behaviour change required from the field teams. The change was entirely on the consumption side, and senior managers embraced it quickly because the alternative was waiting two days for a manually compiled report.