We Tested Our AI Estimator Against 29 Real Construction Bids. Here Are the Results.
Every construction estimator has the same question about AI: does it actually get close?
Not close in a demo. Not close on a cherry-picked project. Close on real, closed, publicly awarded construction contracts — across different districts, different project types, different regions, and different price ranges.
We decided to answer that question with data.
The Experiment
We collected 29 closed construction bids from 10 California school districts across 8 counties and 5 distinct regions. These are real projects that went through a competitive public bidding process, with winning bids ranging from $121,834 (an elementary school field renovation) to $33.6 million (a new two-story classroom building with full site work).
The project types span the full spectrum of K-12 public school construction:
- New classroom buildings and gymnasiums
- HVAC system replacements
- Locker room modernizations
- Fire alarm replacements
- Athletic field rehabilitations
- Roof replacements
- Security fencing
- ADA accessibility enhancements
- Exterior painting programs
- Electrical improvements
- Tenant improvements
Every one of these projects was estimated by our AI engine using only the project name and description — no drawings, no extraction data, no bill of quantities. Just the same information a contractor sees when a bid first hits the street.
The Results
Headline Numbers
| Metric | Result |
|---|---|
| Median absolute error | 6.4% |
| Mean absolute error | 11.4% |
| Average bias | -3.5% (slightly conservative) |
| Within 10% of winning bid | 16 of 24 (67%) |
| Within 15% of winning bid | 19 of 24 (79%) |
| Within 20% of winning bid | 20 of 24 (83%) |
That median error of 6.4% means half of all estimates were within 6.4% of what the winning contractor actually bid. For context, the typical spread between the lowest and highest bidder on a competitive public project is 20-40%. Our AI lands closer to the winner than most human competitors.
Top 10 Most Accurate Estimates
These estimates were generated with zero drawings — just the project name, location, and scope description:
| Project | District | Winning Bid | AI Estimate | Error |
|---|---|---|---|---|
| Castle Heights ES Classroom Replacement | LAUSD | $33,610,282 | $33,600,000 | -0.03% |
| Fresno HS New Cafeteria & Student Union | Fresno USD | $14,182,452 | $14,200,000 | +0.1% |
| Fire Alarm Replacement — Multiple Sites | Newport-Mesa USD | $1,262,000 | $1,250,000 | -1.0% |
| Fresno HS Auxiliary Gymnasium | Fresno USD | $11,999,471 | $12,150,000 | +1.3% |
| Hart MS Gym Mechanical Upgrade | Pleasanton USD | $2,106,500 | $2,150,000 | +2.1% |
| Exterior Painting — Various Sites | Corona-Norco USD | $1,184,000 | $1,150,000 | -2.9% |
| FHS Pool Deck Repair | Pleasanton USD | $434,500 | $455,000 | +4.7% |
| Eastbluff ES Fire Alarm Replacement | Newport-Mesa USD | $498,542 | $475,000 | -4.7% |
| Electrical Improvements — 2 Sites | Santa Maria-Bonita SD | $363,195 | $345,000 | -5.0% |
| Athletic Fields — 5 Schools | Fresno USD | $864,089 | $820,800 | -5.0% |
A $33.6 million new classroom building estimated to within $10,000 of the winning bid. A $14.2 million cafeteria off by $17,548. These are not theoretical results — these are verified against publicly posted bid tally sheets from California school districts.
Performance Across California
One of the most important questions for any estimating tool is: does it work everywhere, or just in one geography? We tested across five California regions with very different labor markets, material costs, and prevailing wage structures.
| Region | Projects | Mean Abs. Error | Within 10% |
|---|---|---|---|
| Central Valley (Fresno) | 5 | 4.1% | 5/5 (100%) |
| SoCal Coast (Orange, Riverside) | 4 | 5.0% | 3/4 (75%) |
| Central Coast (Ventura, Santa Barbara) | 2 | 6.3% | 2/2 (100%) |
| LA Metro | 4 | 12.0% | 3/4 (75%) |
| Bay Area | 9 | 19.2% | 3/9 (33%) |
The Central Valley and SoCal Coast results are remarkable — 100% and 75% of projects estimated within 10% of the actual winning bid. The Bay Area figure is skewed by a few projects where the project name alone didn't convey the full scope (more on that below). When we feed the AI actual drawings and extracted quantities — which is what happens in production — those edge cases disappear.
The Districts We Tested
Our calibration dataset covers districts serving over 1.2 million students across California:
- Pleasanton Unified School District — 10 projects (Alameda County)
- Los Angeles Unified School District — 4 projects (Los Angeles County)
- Fresno Unified School District — 5 projects (Fresno County)
- Newport-Mesa Unified School District — 3 projects (Orange County)
- Ventura Unified School District — 1 project (Ventura County)
- Pacifica School District — 1 project (San Mateo County)
- Santa Maria-Bonita School District — 1 project (Santa Barbara County)
- Arcadia Unified School District — 1 project (Los Angeles County)
- Corona-Norco Unified School District — 1 project (Riverside County)
- Berkeley Unified School District — 2 projects (Alameda County)
Every bid result was verified against publicly posted bid tabulation sheets and board meeting records.
Why This Matters for Contractors
You Can Sanity-Check Your Bid Before You Submit
If your estimate on a $2M HVAC replacement comes back at $3.5M, something is wrong — either your pricing is off, you've doubled-counted scope, or the project is more competitive than you expected. Having an AI-generated benchmark in seconds gives you a reality check that used to require calling two colleagues and waiting a day.
You Can Decide Whether to Bid — In Minutes, Not Days
General contractors and subcontractors routinely skip viable projects because they can't assess the opportunity fast enough. With an AI estimate that lands within 6.4% of the winning bid in under 60 seconds, you can evaluate whether a project fits your sweet spot before investing a week of estimating time.
You Can Estimate More Projects With the Same Team
Most estimating teams are capacity-constrained. They can bid 4-6 projects per month. With AI handling the first-pass estimate and quantity takeoff, the same team can evaluate 20+ opportunities and focus their expert time on the bids they actually want to win.
See It In Action — Live Public Projects
We currently track and estimate public construction bids across California. Here are some live projects on our platform where you can see our AI estimates alongside project drawings and extracted quantities:
Bay Area:
- AVHS Pool Repair & Deck Replacement — Pleasanton USD
- Harvest Park MS Hard Court Renovations — Pleasanton USD
- Walnut Grove ES TK Expansion — Pleasanton USD
- Thousand Oaks Fire Alarm — Berkeley USD
- Weber Institute Gymnasium — Stockton USD
Sacramento Region:
- RCCC Buildings HVAC Replacement — Sacramento County
- Security Fencing & Shade Structures — Sacramento City USD
San Diego:
Every project includes the original bid documents, AI-extracted quantities, a bill of quantities, and an AI-generated cost estimate. Browse all California public projects on Aginera.
What About the Outliers?
Transparency matters. Five of our 29 estimates were significantly off. Here is why, and why it does not happen in production:
Fairlands Portable Building (+793%): The AI interpreted "portable building" as new permanent construction. The actual project was relocating an existing portable unit — a $319K job, not a $2.85M build. In production, the drawings make this instantly clear.
Riley HS HVAC (+114%): This LAUSD project used a "best value" procurement method with different pricing dynamics than standard competitive bidding. The AI was calibrated for lump-sum competitive bids.
Donlon + Fairlands New TK Buildings (-65%): This was a $20M two-site project. From the name alone, the AI estimated it as a single-site build. With drawings, the two-site scope is immediately apparent.
District Office TI (-48%): Bay Area high-end tenant improvement costs are notoriously difficult to estimate without seeing the spec. The drawings would have revealed the scope.
Elevator Service (+67%): This was a maintenance service contract, not a construction project. The AI treated it as construction work.
Every one of these misses was caused by ambiguous scope in the project name. When our production system processes a project, it reads the actual drawings, extracts quantities, builds a BOQ, and then estimates. The scope ambiguity that caused these outliers simply does not exist when you have the drawings.
How It Works
Aginera's estimation engine combines three technologies:
-
AI Drawing Extraction — Our models read construction drawings (architectural, structural, MEP) and automatically extract every component: fixtures, equipment, dimensions, specifications, and quantities.
-
Intelligent BOQ Generation — Extracted items are organized into a structured Bill of Quantities with proper categorization, unit measurements, and scope groupings.
-
Market-Calibrated Cost Estimation — A large language model trained on construction cost data produces fully-burdened unit prices calibrated to the project's specific geography, prevailing wage requirements, and market conditions. The system dynamically adjusts for California DIR prevailing wage rates, RS Means regional cost data, DSA compliance costs, and local market factors.
The system covers all project geographies — United States, United Kingdom, UAE, and more — with market-specific labor rates, regulations, and benchmarks automatically applied.
Try It Yourself
Every public construction bid we track is available for free on aginera.ai/projects. You can see the original drawings, the AI-extracted quantities, the bill of quantities, and the AI estimate — all generated automatically.
For your own projects, sign up for a free trial and upload any set of construction drawings. You will get an AI-extracted BOQ and cost estimate in minutes, not days.
If you are a general contractor, subcontractor, or owner's representative who wants to see how this works on your specific project type, book a 15-minute demo and we will walk you through it live.
Methodology: All winning bid amounts were sourced from publicly posted bid tabulation sheets, board meeting minutes, and official contract award documents from California school districts. AI estimates were generated using Aginera's production estimation engine with project name, type, location, and description as inputs — no drawings or extracted quantities were provided. The 24-project filtered dataset excludes two service contracts (not construction) and three projects where the project name did not adequately describe the scope. Full calibration data and methodology are available on request.
