Reliable takeoffs and FRA compliance with AI
A US rail infrastructure prime contractor running 6 active sites reduced full plan-set takeoff time by 87%, recovered 11 hours per week per Project Manager, and intercepted 31 specs-to-plan inconsistencies before FRA inspection — across 3 agents deployed in 9 weeks.
The starting point: 220 plan files, a revision every 11 days, and no system to analyze them
A US rail construction general contractor operated 6 active sites at peak — rail and road infrastructure projects running day and night operations with dispersed crews. Each project carried a plan set of 140 to 220 PDF and DWG files. Revisions arrived on average every 11 working days. A single full-set takeoff, covering layout plans, longitudinal profiles, and turnout detail drawings, required between 4.5 and 6 hours of manual work per Project Manager — repeated with each revision cycle.
Three documented bottlenecks accounted for the bulk of lost office time.
Takeoffs: manual, repetitive, and error-prone
Quantity takeoffs were done line by line in spreadsheets cross-referenced against printed or screen-viewed plans. Identifying changes between revision n and revision n+1 required side-by-side visual comparison. The process introduced transcription errors and regularly missed delta items on complex turnout configurations, particularly on No. 10 and No. 15 crossovers with multiple simultaneous geometry changes.
FRA compliance review: informal and late
Specification and FRA compliance verification was conducted on a best-effort basis, with no systematic cross-referencing between written documents and graphical plans. Non-conforming geometry, clearance discrepancies, and missing work zones were typically identified during field inspection, at which point corrections were operationally disruptive and schedule-costly.
Document retrieval: slow and unreliable across teams
Revised plans, meeting minutes, inspection reports, RFIs, and technical sheets were stored across SharePoint folder hierarchies and Procore. Locating the current version of a specific document and confirming it was the current version required 18 to 25 minutes per search on average. A measurable share of field teams were periodically working from superseded plan revisions.
Existing systems — Procore for document management, Primavera P6 for scheduling, AutoCAD Civil 3D for plan production — stored and versioned documents correctly. None analyzed content or cross-referenced sources.
The AI agents deployed: architecture and functional scope
Three agents were deployed on the contractor's existing document infrastructure. No existing system was replaced or migrated. Mirage engineers connected to the tools already in place and built the automation inside them.
All three agents were trained on 3 years of the contractor's historical plan corpus — over 1,400 DWG and PDF files — to recognize project-specific notation conventions, title block formats, and layer naming standards.
Automated quantity takeoff from track plans
The agent reads PDF and DWG plan files — layout plans, longitudinal profiles, and turnout detail drawings — and extracts quantities by line item: main track and service track linear footage, number and type of turnouts (No. 8, No. 10, No. 15 crossovers), and crossing lengths. Each extracted value is linked back to the precise location in the plan it was drawn from.
On revision release, the agent automatically compares the new set against the prior version and generates a structured delta report flagging only the changed items. No manual side-by-side comparison is required. Output feeds directly into the contractor's estimating spreadsheets.
Integration: AutoCAD Civil 3D, Bluebeam-annotated PDFs, and SharePoint versioning. Every quantity in the output carries a revision number and plan reference.
Consistency checking between written documents and track plans
The agent cross-references the statement of work, technical specifications, and FRA Part 213 compliance requirements against the graphical plans. It runs automatically on each new revision commit and flags divergences before inspection.
Flagged categories include: geometry in the plans that does not match the tolerance stated in the specification; a work zone referenced in the statement of work that is absent from the plan view; and clearance figures that contradict between the spec sheet and the longitudinal profile.
Every flagged item is routed to the responsible Project Manager or Superintendent for review. The agent produces no corrective action autonomously. Its output is a structured inconsistency log with document name, page reference, revision number, and the specific clause or plan view involved.
Natural-language document search
All project documents are indexed: revised plans, meeting minutes, technical data sheets, inspection reports, RFIs, and FRA correspondence. Any team member — including night-shift Superintendents in the field — can ask a question in plain English.
The agent returns the answer with an exact source citation: document name, page, and revision number. It does not interpret or summarize beyond what the source document states. Retrieval latency is under 4 seconds per query. Access is available on any device connected to the project network, with no separate login required beyond existing Procore credentials.
Implementation approach: 9 weeks, strict governance
The deployment ran in three phases. Human approval was required on every committing decision throughout. No agent output entered the estimating file, triggered a design change, or was transmitted to a client without explicit sign-off from the responsible PM or Superintendent.
Mirage engineers audited the existing document infrastructure, mapped all active repositories across SharePoint and Procore, and assembled the historical training corpus. No plan files were modified. Access was limited to read permissions on the document stores.
All three agents ran in parallel with the manual process on a single active site. Every agent output — takeoff figures, inconsistency flags, document answers — was independently verified by the responsible PM or Superintendent before any action was taken. Discrepancies between agent output and manual result were logged and used to refine extraction rules and model behavior.
Agents were extended to all 6 active sites. A weekly review committee — Project Manager, Superintendent, and Mirage project lead — reviewed flagged inconsistencies and document search quality. Team onboarding required one half-day session per role group. No additional software was installed on field devices.
Measurable results: 87% reduction in takeoff time, 11 hours per week recovered per PM
A full plan-set takeoff that previously required 4.5 to 6 hours was completed by the agent in 35 to 45 minutes, including PM review and sign-off. That represents an 87% reduction in elapsed time per revision cycle.
Across 6 active sites, Project Managers and Superintendents recovered an average of 11 hours per week previously consumed by takeoffs, document comparisons, and manual revision reconciliation. Those hours were reallocated to field supervision, subcontractor coordination, and schedule management.
Delta revision processing — identifying changes between two plan revisions — dropped from 2 to 3 hours of manual comparison to under 20 minutes, including agent output review by the PM.
Compliance control: 31 inconsistencies intercepted before FRA inspection
Over the first 7 months of production operation, the consistency- checking agent flagged 31 divergences between written specifications and graphical plans that would not have been caught before field inspection under the prior process.
All 31 items were reviewed and resolved by the responsible engineer before site execution. None required post-inspection remediation or generated an FRA non-conformance notice. The operational value is straightforward: corrections made at the document stage cost a fraction of what they cost at the field stage.
The agent does not determine whether an item is a genuine non-conformance. It flags a divergence and routes it to the engineer. The engineer decides. That distinction is not incidental — it is the design.
See how the same compliance logic is applied in civil infrastructure contexts on the construction page.
Schedule impact: day and night operations protected by instant document access
Rail construction track possession windows leave no margin for document confusion. A night-shift Superintendent working from the wrong geometry revision — or spending 22 minutes locating a ballast specification — translates directly into crew idle time or, worse, execution against outdated parameters.
Since deployment of the natural-language search agent, average document retrieval time dropped from 18–25 minutes to under 4 seconds. Every query returns the current revision number alongside the answer, eliminating the revision-verification step. The rate of crews working from superseded plan revisions fell to zero on all indexed sites.
On one project with a 6-day track possession window, the ability to answer field questions in real time — without calling the office to locate a document — was cited by the lead Superintendent as a direct contributor to completing the full possession scope on schedule.
Night-shift teams access the system through existing Procore credentials on tablets already in use on site. No new device, no new login, no additional training.
Deployment context: when these agents make sense (and when they don't)
These agents produce measurable value under specific conditions. They are not general-purpose tools and do not apply to every rail project.
Prerequisites
A functioning document management infrastructure — SharePoint, Procore, or equivalent — that versions files consistently. Agents index what exists. They cannot reconstruct missing or unversioned documents.
A sufficient historical corpus for the takeoff agent to learn from. Minimum: 18 to 24 months of project plan files in DWG or PDF format, reflecting the notation conventions and layer standards the project team actually uses.
A governance structure with clear human ownership of every flagged item. Without an assigned review process, inconsistency logs accumulate without resolution and the compliance value disappears.
Relevance thresholds
Projects of at least 12 months duration, with plan revisions arriving at least bi-weekly. Below that frequency, the manual process is fast enough that the efficiency delta is limited.
A minimum of 3 concurrent active sites or 80 or more plans in the active set. Smaller scopes do not generate the volume required to justify the deployment and onboarding investment.
At least 2 Project Managers or Superintendents each spending measurable hours on document work per week. Single-PM projects with a small, stable plan set typically have simpler document workflows that do not justify this type of deployment.
Cases where ROI is limited
A single short-duration project with a stable, small plan set and infrequent revisions. Organizations without a functioning DMS — agents cannot substitute for document infrastructure. Projects where FRA compliance review is already handled by a dedicated compliance officer with a structured, systematic process; in that context, the consistency-checking agent adds limited incremental value.
The agents described in this study are operational tools, not general-purpose AI. They do precisely what they were trained and configured to do, on the documents they were given access to. Their value is proportional to the volume and variability of the document work they replace. When those conditions exist, the economics are clear. When they do not, the honest recommendation is to wait until they do.
If the volume and document profile described here match your current projects, the construction deployments page provides a parallel view of how the same agent architecture runs on civil infrastructure projects.
“The takeoff used to be the first thing I'd block three hours for after a revision drop. Now it's done before I finish my coffee. What matters is that I still review every number before it goes anywhere.”