Construction Tender Takeoff: From 28 Days to 30 Minutes

Customer Profile

One of Türkiye's leading construction firms, active in large-scale public infrastructure and commercial projects. The firm participates in 10-15 major public tenders per year. Each tender pulls a team of 4-6 senior engineers off other work for 3-4 weeks to prepare quantity takeoffs, unit-price analyses and the final bid package.

The Problem

When a public tender is published, the submission deadline is typically 28 calendar days. Within that window the team has to:

• Read hundreds of pages of technical specifications and drawings
• Extract quantities (areas, lengths, counts) for every line item
• Run unit-price analyses against the firm's historical cost database
• Perform risk assessment on margin-sensitive line items
• Produce the formal bid documents in the format the contracting authority demands

Roughly 80% of that work was manual data extraction from drawings and specs. The math is brutal: 4-6 engineers × 28 days × 8 hours = 900-1300 hours of manual labor per tender. Average error rate was 3-5% — which translated, on past bids, into 10-20% cost variance after award. That variance is the difference between a profitable contract and a losing one.

Our Solution

We built a five-stage hybrid AI system that combines computer vision, document AI and a human-in-the-loop review interface.

1. Drawing Processing Pipeline

A custom computer-vision stack ingests AutoCAD DWG files and scanned PDF/JPEG drawings. Detectron2 handles area segmentation (rooms, columns, beams, slabs); a custom-trained YOLOv8 model detects symbols (doors, windows, plumbing fixtures, electrical components). DWG files are parsed natively for vector geometry; raster drawings go through a pre-processing step that normalises rotation, scale and line thickness before detection.

2. Specification Processing

Specification PDFs are sharded by page and pushed through Azure Form Recognizer to extract tables, headings and itemised clauses. GPT-4 then categorises each clause into a discipline (structural, mechanical, electrical, landscape) and links it back to the relevant drawing regions detected in stage one.

3. Quantity Calculation

Area and dimension data from the drawings, plus unit definitions from the specs, plus the firm's internal unit-price database, feed an automated quantity table. An anomaly-detection layer flags outliers — for example: "material X usually runs around 2,000 m² for buildings of this footprint, but this drawing produces 5,500 m² — verify." Anomalies are surfaced to the operator queue with the supporting context attached.

4. Operator Review UI

Low-confidence records (typically 5-10% of total line items) are queued for a human operator. The Next.js review surface shows the original drawing on the left and the AI-extracted data on the right, with single-key approve/edit shortcuts. One operator clears 1,000+ line items in 30 minutes — the system surfaces only what actually needs human judgment.

5. Tender Document Generation

Approved quantities and unit prices flow into a templating engine that produces the formal Excel + PDF bid package in the firm's house format, ready for legal review and submission.

Results (6 months in)

• 28 days → 30 minutes (1344x speedup on the end-to-end takeoff cycle)
• 6 engineers × 28 days = 168 engineer-days → 1 operator × 30 min = 0.0625 engineer-days (≈2688x throughput per tender)
• Accuracy: 97% on line-item quantities (manual baseline was 95%)
• Cost variance after award: ±1% (manual baseline was ±4%)
• Annualised labor recovered: 12 tenders × 168 days saved = 2,016 engineer-days redeployed to billable project work
• New bid capacity: the same team now evaluates 3x more tenders per quarter, materially expanding the firm's win pipeline

Why This Approach Worked

Three architectural decisions made the difference. First, we kept the human in the loop — fully autonomous extraction sounded attractive but would have capped accuracy below the manual baseline; with operator review on low-confidence records, we exceeded it. Second, we treated the firm's internal unit-price database as a first-class data source rather than a downstream consumer, which is what closed the cost-variance gap. Third, the anomaly-detection layer was tuned with engineering judgment from the firm's own senior estimators, not generic statistical thresholds — false positives stayed below 2%, so operators trusted the queue.