# Closed-Project Cost Data Is the Most Undervalued Asset in Your Firm **By Dan Cumberland** · Published May 13, 2026 · Categories: Business Growth > Closed-project cost data is the highest-leverage asset most AEC firms own and rarely use. Nine out of ten large infrastructure projects exceed budget, with an... ## The Asset Sitting in Your Server Room Closed\-project cost data is the highest\-leverage asset most AEC firms own and rarely use\. Nine out of ten large infrastructure projects exceed budget, with an average overrun of 28 percent on typical projects[1](/blog/blog-project-cost-engineering#ref-1)\. Project cost engineering, the discipline of treating that historical project data as a structured asset, turns archived files into fee leverage\. Every firm in this revenue band has the asset\. Drawings, cost reports, change\-order logs, schedule variances, RFI histories\. A record of every project closed in the past decade\. Most of it sits on a shared drive\. Almost none of it gets queried after handover\. AACE International defines project cost engineering through the Cost Estimate Classification System \(18R\-97\)[2](/blog/blog-project-cost-engineering#ref-2), which maps project definition maturity to estimating accuracy\. The standard has been globally adopted since 1997, refreshed in 2020, and remains the reference frame for any estimating discipline conversation\. The estimating problem is a data\-structure problem firms haven't solved yet\. Your firm's source of truth on cost lives in a folder no one queries\. The cost of leaving it there shows up downstream as change orders, mispriced fees, and lost bids\. To see why archives are undervalued, you first need the framework the industry uses to measure estimating maturity in the first place\. ## What Project Cost Engineering Actually Is \(AACE Class 1–5\) Project cost engineering is the systematic discipline of using historical project data to estimate, validate, and control future project costs\. The industry\-standard framework is AACE International's Cost Estimate Classification System \(18R\-97\)[2](/blog/blog-project-cost-engineering#ref-2), which maps project definition maturity to estimating accuracy across five classes\. AACE 18R\-97 establishes that estimate accuracy is governed by project definition maturity[2](/blog/blog-project-cost-engineering#ref-2)\. Not effort\. Not seniority\. Not software\. A Class 5 estimate ranges as wide as ±50 percent\. A Class 1 estimate tightens to ±5 percent\. The class\-to\-accuracy mapping is the contract between estimating effort and the precision your client can expect\. Here is the full breakdown: ```html-table
ClassDesign CompletionAccuracy RangeTypical Use
Class 50–2% (ROM)-20% to -50% / +30% to +100%Concept screening, feasibility
Class 41–15%-15% to -30% / +20% to +50%Study, schematic
Class 310–40%-10% to -20% / +10% to +30%Budget authorization
Class 230–70%-5% to -15% / +5% to +20%Control/bid
Class 150–100%-3% to -10% / +3% to +15%Definitive, contract
``` The framework gives firms a shared vocabulary for accuracy ranges that maps to fee structures and risk allocation\. When a Class 3 estimate is the contractual commitment, both firm and client understand what variance is acceptable\. CSI standards complement AACE on the data\-structure side[3](/blog/blog-project-cost-engineering#ref-3)\. AACE 18R\-97 is the entry credential for any cost intelligence conversation in AEC\. It tells you what accuracy is achievable at a given design maturity\. It does not tell you how to extract better signal from the projects you have already closed\. That is where most firms operate two classes below where their data could take them, and where structure becomes the limiting factor\. ## Why AEC Firms Underprice Their Own Archives Firms underutilize closed\-project data for three structural reasons: archives are unstructured, cost codes vary project\-to\-project, and no one owns the cross\-project analysis\. The cost shows up downstream as change orders, mispriced fees, and lost bids\. **Structure\.** Each project carries its own cost\-code system, or a proprietary one a PM built mid\-project\. Without UniFormat or MasterFormat discipline applied consistently, cross\-project comparison is manual archaeology\. Most firms have the data on a shared drive\. They don't have it in a queryable structure\. **Ownership\.** Estimators close projects and move to the next bid\. No dedicated function audits closed projects for variance patterns\. Lessons\-learned meetings happen\. Lessons\-learned datasets rarely do\. You can't read the label from inside the bottle when no one is hired to read it\. **Cost\.** The downstream effect is measurable\. Change orders average 8 to 14 percent of total contract value across the industry, with distressed projects reaching 25 percent[4](/blog/blog-project-cost-engineering#ref-4)\. Design\-professional errors account for 3 to 5 percent of change orders nationally[4](/blog/blog-project-cost-engineering#ref-4), the subset most addressable by historical pattern analysis\. The patterns that predict them sit in your last 20 projects\. You just can't see them yet\. Construction productivity growth has averaged 1 percent annually over the past two decades, against 2\.8 percent for the global economy and 3\.6 percent for manufacturing[5](/blog/blog-project-cost-engineering#ref-5)\. Dodge tracks 16,000\-plus US projects daily across the industry[6](/blog/blog-project-cost-engineering#ref-6); individual firms rarely track their own portfolio with comparable rigor\. Your firm's [AI decision framework](/blog/ai-decision-framework-founders) should reflect what machine learning now extracts from historical cost data\. ## What AI/ML Actually Extracts From Closed\-Project Data Machine\-learning models trained on historical project data can predict future project costs with accuracy that approaches Class 2 estimates from Class 4 design inputs\. Peer\-reviewed research using XGBoost regression on construction cost data reports 9\.09 percent mean absolute percentage error and a 0\.929 adjusted R²[7](/blog/blog-project-cost-engineering#ref-7)\. In practical terms, the model explains 93 percent of cost variance across the test dataset\. Parametric estimating uses historical project parameters \(square footage, structural system, occupancy type, region, year\) plus actuals from closed projects to fit a model that predicts cost on a new project sharing those parameters\. The discipline has existed for decades; the data volume modern algorithms handle is what changed\. Peer\-reviewed comparison across construction\-cost datasets identifies the working ML toolkit[7](/blog/blog-project-cost-engineering#ref-7)\. These are the methods construction\-cost researchers benchmark against each other; each is suited to different data shapes and prediction tasks: - Fuzzy logic \(FL\) - Artificial neural networks \(ANN\) - Multiple regression - Case\-based reasoning \(CBR\) - Decision trees \(DT\) and random forest \(RF\) - XGBoost \(gradient\-boosted regression\) - Genetic algorithms XGBoost was the highest\-performing method in that comparison, posting 9\.091 percent MAPE and 0\.929 adjusted R²[7](/blog/blog-project-cost-engineering#ref-7)\. The caveat: these numbers are dataset\-specific\. Performance varies with data quality, feature engineering, and project mix\. No universally "best" algorithm exists outside of a defined dataset\. A 2024 practitioner study from construction\-management software vendor Monograph reports 20\.4 percent better accuracy, 51\.3 percent faster completion, and under 5 percent variance on bid day when models pull auto\-refreshed labor and material indices[8](/blog/blog-project-cost-engineering#ref-8)\. Treat the numbers as practitioner signal with peer\-reviewed methodology citations\. The indices themselves come from sources like ENR's Construction Cost Index[9](/blog/blog-project-cost-engineering#ref-9), which tracks materials and labor across 20 US cities going back decades\. AI here is intellectual augmentation\. The model surfaces patterns in your archive that no individual estimator has time to find\. The AACE\-trained estimator interprets context, applies judgment, and signs the estimate\. When firms are [measuring AI success in implementation](/blog/measuring-ai-success), the right KPIs track signal quality, not headcount reduction\. ## The Data Structure That Makes AI Possible \(UniFormat → MasterFormat\) AI extraction depends on consistent cost taxonomy\. CSI's UniFormat organizes costs by building element \(foundations, structure, exterior enclosure, services\) for cross\-project comparability\. MasterFormat organizes costs by product specification across 50 divisions, such as Division 03 Concrete and Division 26 Electrical, for detailed design and procurement[3](/blog/blog-project-cost-engineering#ref-3)\. Firms that adopt both turn each closed project into a row in a queryable dataset\. The two standards do different work at different stages\. UniFormat lives in early\-phase estimating, where the question is how much an elementary school of a given size and structural system costs per square foot in this region\. MasterFormat lives in detailed design, where the question is what the procurement cost is for this specific concrete mix or electrical specification\. CSI Dynamic Standards govern the crosswalk between them\. As design matures, a UniFormat element rolls up into the MasterFormat product specifications that make up that element[3](/blog/blog-project-cost-engineering#ref-3)\. A firm maintaining both spines on every project gets two queryable views of the same archive: one for benchmarking, one for procurement\. Without a shared taxonomy, every project is a one\-off\. With one, every project is a training sample\. This is the unglamorous prerequisite most firms skip when they shop for an AI estimating tool\. When the model fails, the cause sits in the data structure: 200 different schemas pretending to be one\. Standards plus models plus archives is the technical stack\. The business case is what makes it worth funding\. ## The Business Case — Win Rates, Margins, Change Orders Firms that systematize cost intelligence see measurable separation from peers: higher win rates, tighter fee accuracy, and reduced change\-order exposure\. Monograph's 2024 practitioner analysis reports digital leaders are 19 percent more likely to exceed 60 percent win rates than emerging firms, and 37 percent more likely than firms still classed as beginners[8](/blog/blog-project-cost-engineering#ref-8)\. Win\-rate improvement is the most under\-priced ROI line on the business case\. A two\-point improvement is often worth more than every other gain combined for a firm bidding meaningful volume\. Fee accuracy is the second line\. Moving from Class 4 to Class 3 estimates reduces fee leakage from scope ambiguity\. Auto\-refreshed indices from ENR keep the model current with materials and labor escalation[9](/blog/blog-project-cost-engineering#ref-9)\. KPMG's 2024 Global Construction Survey of 375 construction leaders reports cost inflation of 4\.15 percent globally for the year, with North America at 3\.6 percent[10](/blog/blog-project-cost-engineering#ref-10)\. That escalation alone is enough to wipe a firm's fee margin if estimates don't adjust quarterly\. Change\-order reduction is the third line, and the most measurable\. Change orders average 8 to 14 percent of total contract value across AEC[4](/blog/blog-project-cost-engineering#ref-4)\. Design\-professional errors account for 3 to 5 percent of those, the subset most addressable by historical pattern analysis\. Here is the math at $40M of annual revenue: > A firm running $40M of annual contract value with change\-order cost averaging 12 percent is exposed to roughly $4\.8M of change\-order activity per year\. Cutting that to 9 percent through historical\-pattern analysis recovers $1\.2M of margin\. No fee increase\. No staff reduction\. Just better signal pulled from projects the firm has already closed\. Sector productivity grows 1 percent a year[5](/blog/blog-project-cost-engineering#ref-5)\. The relative gain from being one or two standard deviations ahead of peers compounds across a decade\. Most firms have not yet accounted for the [hidden costs of AI implementation](/blog/hidden-costs-ai-projects) in this category, which is why the gap stays open longer than competitive theory predicts\. ## Where to Start \(Implementation Sequence\) The first move is auditing whether your last 20 closed projects can be expressed in a shared taxonomy\. Software comes after\. If your projects cannot be expressed in a shared taxonomy, that audit is the project\. Everything else depends on it\. Data structure precedes data science\. The firms that rush to AI before fixing their cost\-code discipline buy expensive models trained on noise\. Start with the projects you already closed, not the projects you wish you had\. Six steps, in sequence: 1. **Audit data accessibility\.** Where do closed\-project cost archives actually live? Who has them? In what format? PDF, Excel, Procore exports, paper\. Catalog before you cleanse\. 2. **Adopt a shared taxonomy\.** UniFormat for element\-level cross\-project comparison\. MasterFormat for detailed projects[3](/blog/blog-project-cost-engineering#ref-3)\. Decide who owns the mapping\. 3. **Backfill the last 20 to 50 projects\.** Manual, tedious, and the highest\-leverage work in this whole stack\. Once done, you have a training dataset\. 4. **Pilot a parametric model on one project type\.** Pick the type you bid most frequently\. Compare model output to your estimators' instinct\. Track the delta over five to ten bids\. 5. **Treat estimators as the decision\-makers\.** The model surfaces patterns; the estimator interprets context\. This is where [building an AI culture across your firm](/blog/building-ai-culture) shifts from policy memo to daily decision\. 6. **Refresh quarterly\.** ENR indices, labor rates, regional adjustments[9](/blog/blog-project-cost-engineering#ref-9)\. A stale model is worse than no model— it carries the authority of a fresh one without the accuracy\. The unglamorous, manual work in steps 1 and 3 is where the leverage is\. Steps 4, 5, and 6 are the work most consultants want to talk about\. Steps 1, 2, and 3 are the work that makes the rest possible\. ## Frequently Asked Questions Frequently asked questions about project cost engineering, AACE standards, and AI\-assisted estimating: ### What is project cost engineering? Project cost engineering is the systematic discipline of using historical project data to estimate, validate, and control future project costs\. The industry standard is AACE International's Cost Estimate Classification System \(18R\-97\), which defines five accuracy classes based on project definition maturity[2](/blog/blog-project-cost-engineering#ref-2)\. ### What does AACE Class 1 vs\. Class 5 mean? AACE Class 5 is a rough order\-of\-magnitude estimate at 0 to 2 percent design completion, with an accuracy range as wide as ±50 percent\. AACE Class 1 is a definitive estimate at over 50 percent design completion, with an accuracy range as tight as ±3 to ±10 percent\. Class is determined by project definition maturity[2](/blog/blog-project-cost-engineering#ref-2), not by effort or seniority\. ### How accurate is AI for construction cost estimating? Peer\-reviewed research using XGBoost regression on construction cost data reports 9\.09 percent mean absolute percentage error and 0\.929 adjusted R²[7](/blog/blog-project-cost-engineering#ref-7)\. The model explains roughly 93 percent of cost variance in the test dataset\. Practitioner studies from Monograph report 20\.4 percent better accuracy and under 5 percent variance on bid day when models pull auto\-refreshed indices[8](/blog/blog-project-cost-engineering#ref-8)\. ### What is the difference between UniFormat and MasterFormat? UniFormat organizes construction costs by building element and functional system \(foundations, exterior enclosure, services\), suited to early\-phase estimating and cross\-project benchmarking\. MasterFormat organizes costs by product specification across 50 divisions \(Division 03 Concrete, Division 26 Electrical\), suited to detailed design and procurement\. CSI Dynamic Standards define the crosswalk between them as design progresses[3](/blog/blog-project-cost-engineering#ref-3)\. ### What is a typical change\-order rate on AEC projects? Change orders average 8 to 14 percent of total contract value across the industry, with distressed projects reaching 25 percent[4](/blog/blog-project-cost-engineering#ref-4)\. Design\-professional errors account for 3 to 5 percent of change orders nationally, the subset most addressable through historical pattern analysis\. ## The Discipline, Not the Software The competitive advantage is the discipline that produces the dataset\. The model is downstream of that\. AI mastery in cost engineering is fundamentally about thinking skills and strategy, not software\. The firms that win the next decade are the ones who built a project cost engineering capability while their competitors were shopping for tools\. The asset already exists in your firm\. AACE gives you the framework[2](/blog/blog-project-cost-engineering#ref-2)\. UniFormat and MasterFormat give you the structure[3](/blog/blog-project-cost-engineering#ref-3)\. XGBoost and its peers give you the extraction[7](/blog/blog-project-cost-engineering#ref-7)\. None of it works without the dataset, and the dataset is already on your shared drive waiting to be read\. Sector productivity grows 1 percent a year[5](/blog/blog-project-cost-engineering#ref-5)\. A firm one standard deviation ahead of that baseline for ten years has built something durable\. If you are evaluating where cost intelligence fits in your firm's AI strategy, what to build, what to buy, what sequence makes sense for a $20M to $100M practice, that is the kind of conversation [Dan Cumberland Labs' AI strategy services for founder\-led firms](/services/ai-strategy/) are built for\. We help AEC firm leadership map AI implementation to the specific decisions principals are already making\. ⚠️ EVERYTHING BELOW IS PIPELINE METADATA — NOT PUBLISHED ## References 1. McKinsey & Company, "Megaprojects: The Good, The Bad, and The Better" \(2023\) — [https://www\.mckinsey\.com/capabilities/operations/our\-insights/megaprojects\-the\-good\-the\-bad\-and\-the\-better](https://www.mckinsey.com/capabilities/operations/our-insights/megaprojects-the-good-the-bad-and-the-better) 2. AACE International, "Cost Estimate Classification System \(18R\-97\)" \(1997, updated 2020\) — [https://web\.aacei\.org/docs/default\-source/toc/toc\_18r\-97\.pdf](https://web.aacei.org/docs/default-source/toc/toc_18r-97.pdf) 3. Construction Specifications Institute \(CSI/CSC\), "MasterFormat and UniFormat Standards" \(2024\) — [https://theconstructionstandard\.com/how\-is\-uniformat\-used\-in\-cost\-estimating](https://theconstructionstandard.com/how-is-uniformat-used-in-cost-estimating) 4. American Institute of Architects, "The Truth About Change Orders" \(2023\) — [https://learn\.aiacontracts\.com/wp\-content/uploads/2023/07/The\-Truth\-About\-Change\-Orders\.pdf](https://learn.aiacontracts.com/wp-content/uploads/2023/07/The-Truth-About-Change-Orders.pdf) 5. McKinsey & Company, "Reinventing Construction: A Route to Higher Productivity" \(2017, updated 2024\) — [https://www\.mckinsey\.com/capabilities/operations/our\-insights/reinventing\-construction\-through\-a\-productivity\-revolution](https://www.mckinsey.com/capabilities/operations/our-insights/reinventing-construction-through-a-productivity-revolution) 6. Dodge Data & Analytics, "Construction Analytics — Dodge Construction Network" \(2024\) — [https://www\.construction\.com/reports/construction\-analytics/](https://www.construction.com/reports/construction-analytics/) 7. American Society of Civil Engineers, "Journal of Construction Engineering & Management — AI & Parametric Estimating \(Vol\. 146, No\. 1\)" \(2020\) — [https://ascelibrary\.org/doi/abs/10\.1061/\(ASCE\)CO\.1943\-7862\.0001678](https://ascelibrary.org/doi/abs/10.1061/(ASCE)CO\.1943\-7862\.0001678\) 8. Monograph, "AI in Construction Estimating: Impact Study & ROI Guide" \(2024\) — [https://monograph\.com/blog/ai\-construction\-estimating\-accuracy\-roi\-guide](https://monograph.com/blog/ai-construction-estimating-accuracy-roi-guide) 9. Engineering News\-Record, "Construction Cost Index — Economics" \(2024\) — [https://www\.enr\.com/economics](https://www.enr.com/economics) 10. KPMG International, "Global Construction Survey 2023 \(15th Annual\)" \(2024\) — [https://kpmg\.com/xx/en/our\-insights/operations/2023\-global\-construction\-survey\.html](https://kpmg.com/xx/en/our-insights/operations/2023-global-construction-survey.html) --- Source: https://dancumberlandlabs.com/blog/project-cost-engineering/