# A Civil Firm's 14-Day PER Boilerplate Experiment Produced the ROI Number a Board Couldn't Ignore **By Dan Cumberland** · Published May 13, 2026 · Categories: Business Growth > A CSI number is a six-digit code from the CSI MasterFormat system that identifies a specific construction work result. The first two digits name the division,... ## What a CSI Number Actually Is A CSI number is a six\-digit code from the CSI MasterFormat system that identifies a specific construction work result\. The first two digits name the division, the next two name the section, and the last two name the subsection\. MasterFormat 2020, published jointly by the Construction Specifications Institute \(CSI\) and Construction Specifications Canada \(CSC\), is the current edition and contains 50 divisions covering everything from procurement to waterway construction\.[1](/blog/blog-csi-number-construction#ref-1)[2](/blog/blog-csi-number-construction#ref-2) The structure reads as three pairs of digits: - **First pair \(DD\):** Division \(the trade or work category\) - **Second pair \(SS\):** Section within that division - **Third pair \(SS\):** Subsection providing detailed classification A worked example clarifies it\. Procore's guide to MasterFormat uses 06 41 93 = Cabinet and Drawer Hardware: Division 06 \(Wood, Plastics, and Composites\), Section 41 \(Architectural Wood Casework\), Subsection 93 \(Cabinet and Drawer Hardware\)\.[3](/blog/blog-csi-number-construction#ref-3) Civil engineers see analogous codes constantly— 33 30 00 for Sanitary Sewerage Utilities, 31 23 00 for Excavation and Fill\. MasterFormat is sometimes called the Dewey Decimal System of construction— the industry standard organizing system in North America for construction specifications\.[3](/blog/blog-csi-number-construction#ref-3) It shows up everywhere the work meets paper: project manuals, submittals, bid packages, estimating sheets, and the PER documents that anchor federal water and waste infrastructure funding\. Six digits don't matter much until you see how the 50 divisions group up— and how three of them happen to live where civil engineers spend most of their week\. ## The 50 Divisions, Mapped to Civil Work The 50 MasterFormat divisions are organized into four subgroups \(Facility Construction, Facility Services, Site and Infrastructure, and Process Equipment\) plus Division 00 \(Procurement and Contracting\) and 01 \(General Requirements\)\.[4](/blog/blog-csi-number-construction#ref-4) For civil engineers, the work concentrates in Divisions 31 through 35, the Site and Infrastructure subgroup\.[5](/blog/blog-csi-number-construction#ref-5) The system wasn't always this big\. MasterFormat expanded from 16 to 50 divisions in November 2004, responding to industrial and heavy\-civil complexity that the original taxonomy couldn't carry\.[1](/blog/blog-csi-number-construction#ref-1) That 2004 expansion is what gave civil firms their own subgroup instead of leaving water, wastewater, and earthwork buried under "Sitework\." Here's the civil engineer's home turf: ```html-table
DivisionScopeWhat It Covers
31 EarthworkSite preparation and gradingExcavation, fill, soil treatment, dewatering
32 Exterior ImprovementsSite features outside the building footprintPaving, landscaping, fencing, site furnishings
33 UtilitiesOutside-the-fence utilitiesWater, wastewater, stormwater, energy distribution
34 TransportationMovement systemsRoadways, rail, bridges, traffic control
35 Waterway and MarineWater-edge and aquatic infrastructureDams, levees, coastal work, marine construction
``` Division 33 deserves a closer look because it's the heart of civil PER work\. According to The Construction Specifier, Division 33— Utilities addresses all types of "outside\-the\-fence" utilities, including water, wastewater, stormwater, hydrocarbons, hydronic and steam energy, high\- and extra\-high voltage electrical, and communications\.[6](/blog/blog-csi-number-construction#ref-6) In practice, those sections include 33 10 00 Water, 33 30 00 Wastewater, and 33 40 00 Stormwater— the exact section families that drive most federal water\-infrastructure proposals\.[7](/blog/blog-csi-number-construction#ref-7) Once you accept that civil work concentrates in five divisions— and that those divisions structure your PERs, your proposals, and your submittals— a quiet question surfaces\. What fraction of all that text is actually different from project to project? ## Why CSI Structure Is the AI Opportunity Hiding in Plain Sight Approximately 80% of specification content is consistent across projects, according to a 2024 analysis published by Deltek through the American Institute of Architects\.[8](/blog/blog-csi-number-construction#ref-8) That consistency is exactly what makes CSI\-organized documents tractable for AI assistance\. The six\-digit code isn't bureaucracy\. It's a machine\-readable spine that lets you point AI at a single section, hold quality constant with human review, and measure throughput\. > Roughly 80% of specification content is consistent across projects— and CSI numbering is what makes that consistency machine\-readable\. Why does structure matter so much? AI works best on bounded, labeled, repeating problems\. A six\-digit code gives you all three at once\. You can take five years of past Division 33 30 00 wastewater sections, identify the structural template inside them, and use that template as the spine of an AI prompt\. The model has a defined target\. The reviewer has a defined output\. And the firm has a defined unit of work\. Where does the boilerplate actually live? PER sections\. Proposal qualifications\. Submittal cover pages\. Spec front\-end matter\. It also lives where federal funding agencies set the rules— and there's a structural reason that matters for AI work\. A standardized PER format was jointly adopted by USDA, EPA, and HUD in April 2013 for water and waste infrastructure funding applications\.[9](/blog/blog-csi-number-construction#ref-9) When three federal agencies share a document template, the boilerplate quotient inside that document goes up, not down\. The AIA/Deltek analysis is architecture\-leaning, and we should be honest about that\. But the structural argument— that federally templated PERs concentrate even more repeatable text than typical specs— points the same direction\. The question for a civil firm leader stops being "should we try AI?" and becomes "which CSI sections do we automate first?" That's a smaller, cheaper, more defensible question to answer\. If the 80% number sounds abstract, the easiest way to make it concrete is to scope it down\. Pick one document type, pick one boilerplate\-heavy section, and run a measured pilot\. Here's what a 14\-day version looks like in civil practice\. ## The 14\-Day PER Boilerplate Experiment A 14\-day PER boilerplate pilot scopes AI assistance to a single, repeating section type— typically Division 33 \(Utilities\) facility\-description or alternatives\-analysis text— and measures time\-per\-section while holding quality constant with engineer review\. It is not a complete PER cycle\. A complete PER typically runs 9 to 12 months according to industry practice\.[10](/blog/blog-csi-number-construction#ref-10) It is a contained throughput test on the most structurally repeatable content in the document\. > A 14\-day pilot doesn't measure a complete PER\. It measures throughput on a single CSI\-organized section type, with engineer review holding quality constant\. The scope matters more than the timeline\. A civil firm running this kind of pilot picks one document type \(PER\), one section family \(e\.g\., Division 33 wastewater alternatives analysis\), and one team to execute\. That's the unit of work\. Everything else flows from that decision: tool selection, model choice, prompt engineering\. The 14\-day rhythm breaks into four blocks: 1. **Days 1–3: Gather and template\.** Pull five to ten representative past PER sections from the firm's archive\. Extract the structural template— what every section has in common across projects\. 2. **Days 4–7: Build and back\-test\.** Draft prompt scaffolding and a one\-paragraph quality rubric\. Run AI drafts on three historical PERs \(this is back\-test work, not live deliverables\)\. 3. **Days 8–11: Run on live sections\.** Generate AI drafts on two live, non\-submitted PER sections\. Engineer review and red\-line each one against the rubric\. 4. **Days 12–14: Measure and write up\.** Document time\-per\-section before and after\. Capture red\-line volume as a quality proxy\. Write the one\-page memo for the principal or board\. What gets measured during the pilot: - **Hours per section:** Baseline \(without AI\) vs\. AI\-assisted draft \+ engineer review - **Engineer red\-line volume:** Words changed, sentences rewritten, structural edits— a proxy for quality - **Reusable artifacts produced:** Prompt library, quality rubric, structural template - **Sections completed:** How many PER sections moved through the pipeline in two weeks Why does this design work? Because federal PER format is jointly adopted by USDA, EPA, and HUD— structurally templated boilerplate is the highest\-leverage starting point\.[9](/blog/blog-csi-number-construction#ref-9) You're not asking AI to invent a methodology\. You're asking it to fill in a known structure with the firm's voice and the project's specifics\. That's the task the current generation of frontier models is good at\. A note on framing\. This article narrates the pilot shape and uses industry data to back the ROI argument\. The specific dollar figures and outcomes referenced in the next section come from Bluebeam's 2025 survey of more than 1,000 AEC decision\-makers— aggregate data, not a single named engagement\. The 14\-day rhythm above is the structural design we'd recommend for any civil firm running its first AI pilot on PER work\. Throughput numbers from one pilot only matter if they line up with what the broader industry is reporting\. They do\. ## The ROI Numbers a Board Couldn't Ignore Bluebeam's October 2025 survey of more than 1,000 AEC technology decision\-makers found that 68% of firms using AI have saved at least $50,000, and 46% have reclaimed 500 to 1,000 hours on critical tasks like document analysis and planning\.[11](/blog/blog-csi-number-construction#ref-11)[12](/blog/blog-csi-number-construction#ref-12) Only 27% of AEC firms currently use AI, but 94% of those that do plan to expand investment in 2026— meaning the firms running pilots now are setting the competitive floor\.[13](/blog/blog-csi-number-construction#ref-13)[14](/blog/blog-csi-number-construction#ref-14) > 68% of AEC AI adopters have saved at least $50,000\. 46% have reclaimed 500–1,000 hours\. 94% plan to expand investment\. A clean board\-readable view: ```html-table
What AEC AI Adopters ReportNumberSource
Saved at least $50,00068%Bluebeam 2025 (n=1,000+)
Reclaimed 500–1,000 hours46%Bluebeam 2025
Currently using AI27% of AEC firmsBluebeam 2025
Plan to expand investment in 202694% of adoptersBluebeam 2025
``` The historical baseline reinforces the direction\. McKinsey's pre\-generative\-AI analysis estimated that AI could increase construction productivity by up to 20%, reduce costs by up to 15%, and improve project delivery times by up to 30%\.[15](/blog/blog-csi-number-construction#ref-15) That work pre\-dates the current generation of models, so treat it as the older anchor— the conversation the industry had before Bluebeam ran a 1,000\+ firm survey on actual outcomes\. Pair them as historical baseline and current reality\. For a civil firm board, the translation is straightforward\. Pilot\-result\-per\-engineer multiplied by engineers, repeat sections, and proposals per year produces an annualized hours\-saved number\. That's what fits on a slide\. A pilot that proves a 40% time\-per\-section reduction on a Division 33 wastewater section becomes a serious capital\-allocation conversation once you see how often the firm writes those sections\. For framing on what matters in those board conversations, see [how to measure AI success](/blog/measuring-ai-success)\. > A board doesn't need a forecast\. It needs a number from a 14\-day pilot and a chart showing where the industry is heading\. Aggregate numbers look generous because they describe successful adopters\. The other half of the picture is that most AI pilots don't make it to ROI at all— and the reason has nothing to do with the models\. ## Why Most AI Pilots Fail \(and How to Avoid It\) Most AI pilots fail because of workflow integration, not because the models are inadequate\. The pilots that succeed are narrowly scoped, measured against a baseline, and embedded in the document type the team already produces every quarter\. The pilots that fail try to "do AI" across the firm at once\. > Most AI pilots fail on workflow integration, not model quality\. Scope kills more pilots than capability does\. The American Society of Civil Engineers' 2025 survey of more than 2,200 industry professionals found that lack of skilled personnel is the most cited barrier to AI adoption among AEC firms\.[16](/blog/blog-csi-number-construction#ref-16) In other words: the constraint is people\-and\-process, not technology\. And the diagnosis matches what we see in [why engineering firms struggle with AI adoption](/blog/engineering-firms-ai-adoption/)— pilots stall when there's no one whose calendar owns them\. Where pilots typically break down: - **Scope explosion\.** "Let's try AI on everything" becomes "we tried AI on nothing well\." - **Missing baseline\.** Without a measured starting point, the team can't prove the delta\. - **No engineer\-in\-the\-loop\.** Models draft, no one reviews against a rubric, and quality slips invisibly\. The 14\-day design defends against each: - **Small scope\.** One section type, one document type, one team\. - **Measured baseline\.** Days 1–3 capture before\-numbers from past work\. - **Engineer review\.** Every AI draft gets red\-lined against a one\-paragraph rubric— a discipline that mirrors how skilled civil engineers already QC junior staff\. For tooling context across the broader category, the [AI construction software landscape](/blog/ai-construction-software/) maps where models, plugins, and platforms intersect with the civil workflow\. But tool selection rides on scope, not the other way around\. If the design works, the next question is operational: how do you actually run one of these next month without disrupting a single live deliverable? ## How to Run Your Own 14\-Day Pilot Next Month Running a 14\-day PER boilerplate pilot starts with picking the right section and the right historical comparables— not with picking an AI tool\. Choose one CSI\-organized section your firm writes at least quarterly, gather five to ten of the firm's past versions, and define what "passing quality" looks like before you write a single AI prompt\. > Pick the section before you pick the tool\. The CSI code is the unit of work\. If you can't define passing quality in one paragraph before the pilot starts, you don't have a pilot— you have a demo\. Pre\-work before Day 1: 1. **Pick the CSI section\.** Best candidates are sections your firm writes quarterly or more\. Division 33 30 00 Wastewater alternatives analysis is the canonical civil PER candidate\. 2. **Gather 5–10 past versions\.** Real deliverables from the last 24 months\. Anonymize client names if needed\. 3. **Write a one\-paragraph quality rubric\.** What does a passing section look like? What red\-line patterns mean "send back"? 4. **Pick the AI tool\.** Any current frontier model \(ChatGPT, Claude, or Gemini\) handles long\-form technical prose competently\. Tool choice matters less than scope discipline\. This is also a good moment to think through your overall [AI decision framework for founders](/blog/ai-decision-framework-founders)\. What survives the pilot— these are the artifacts the firm keeps regardless of outcome: - A reusable prompt library scoped to the pilot section type - A one\-paragraph quality rubric the firm uses going forward - A measured time\-per\-section delta \(before vs\. AI\-assisted\) - A one\-page memo summarizing the result for the principal or board > What to leave OUT of the first pilot: anything that touches a live federal submission\. Run the pilot on past work or non\-binding sections only, until the rubric is stable\. If scoping this kind of pilot feels heavier than your team can absorb between live deliverables, an outside consultant who has run similar engagements can compress the design phase from weeks to days\. That's exactly the kind of work [AI strategy services for engineering firms](/services/ai-strategy/) is built to handle— peer\-to\-peer, no vendor lock\-in, deliverables you own at the end\. The right partner does the scoping, you keep the artifacts\. Whether you run the pilot yourself or with help, the strategic move is the same: pick one CSI section, run a measured 14\-day test, and walk a board memo into the room\. Everything else is a footnote\. A few common questions surface every time a firm starts mapping CSI numbers to AI work\. Here are the short answers\. ## FAQ ### What is a CSI number in construction? A CSI number is a six\-digit code from the CSI MasterFormat system that identifies a specific construction work result\. The structure reads as division, section, subsection\. Example: 06 41 93 = Cabinet and Drawer Hardware \(Division 06, Section 41, Subsection 93\)\.[3](/blog/blog-csi-number-construction#ref-3) ### How many CSI divisions are there? Fifty, organized into four subgroups \(Facility Construction, Facility Services, Site and Infrastructure, and Process Equipment\) plus Division 00 \(Procurement and Contracting\) and 01 \(General Requirements\)\.[4](/blog/blog-csi-number-construction#ref-4) ### Which CSI divisions matter most to civil engineers? Divisions 31 through 35: Earthwork, Exterior Improvements, Utilities, Transportation, and Waterway and Marine Construction\.[5](/blog/blog-csi-number-construction#ref-5) Together they form the Site and Infrastructure subgroup— civil engineering's home turf in MasterFormat\. ### What does CSI 33 30 00 mean? Division 33 \(Utilities\), Section 30 \(Sanitary Sewerage Utilities\), Subsection 00 \(the overall section\)\. Division 33 covers "outside\-the\-fence" utilities, including water, wastewater, stormwater, and energy distribution\.[6](/blog/blog-csi-number-construction#ref-6)[7](/blog/blog-csi-number-construction#ref-7) ### What's the current MasterFormat edition? MasterFormat 2020, published jointly by CSI and CSC\.[2](/blog/blog-csi-number-construction#ref-2) The system has been at 50 divisions since the November 2004 expansion from the original 16\.[1](/blog/blog-csi-number-construction#ref-1) ### What ROI are AEC firms getting from AI? 68% of AEC AI adopters have saved at least $50,000, and 46% have reclaimed 500–1,000 hours, per Bluebeam's October 2025 survey of more than 1,000 industry decision\-makers\.[11](/blog/blog-csi-number-construction#ref-11)[12](/blog/blog-csi-number-construction#ref-12) Only 27% of AEC firms currently use AI today, but 94% of those that do plan to expand investment in 2026\.[13](/blog/blog-csi-number-construction#ref-13)[14](/blog/blog-csi-number-construction#ref-14) ⚠️ EVERYTHING BELOW IS PIPELINE METADATA — NOT PUBLISHED ## References 1. Wikipedia, "MasterFormat" \(2025\) — [https://en\.wikipedia\.org/wiki/MasterFormat](https://en.wikipedia.org/wiki/MasterFormat) 2. Wikipedia, "MasterFormat" \(2025\) — [https://en\.wikipedia\.org/wiki/MasterFormat](https://en.wikipedia.org/wiki/MasterFormat) 3. Procore, "MasterFormat: The Definitive Guide to CSI Divisions in Construction" \(2024\) — [https://www\.procore\.com/library/csi\-masterformat](https://www.procore.com/library/csi-masterformat) 4. Wikipedia, "50 Divisions" \(2025\) — [https://en\.wikipedia\.org/wiki/50\_Divisions](https://en.wikipedia.org/wiki/50_Divisions) 5. Wikipedia, "50 Divisions" \(2025\) — [https://en\.wikipedia\.org/wiki/50\_Divisions](https://en.wikipedia.org/wiki/50_Divisions) 6. The Construction Specifier, "Exploring the Utility of the New Division 33" \(2016\) — [https://www\.constructionspecifier\.com/exploring\-the\-utility\-of\-the\-new\-division\-33/](https://www.constructionspecifier.com/exploring-the-utility-of-the-new-division-33/) 7. The Construction Specifier, "Exploring the Utility of the New Division 33" \(2016\) — [https://www\.constructionspecifier\.com/exploring\-the\-utility\-of\-the\-new\-division\-33/](https://www.constructionspecifier.com/exploring-the-utility-of-the-new-division-33/) 8. Deltek / American Institute of Architects, "Shaping the future: Six key benefits of AI in specifications" \(2024\) — [https://www\.aia\.org/resource\-center/shaping\-future\-six\-key\-benefits\-ai\-specifications](https://www.aia.org/resource-center/shaping-future-six-key-benefits-ai-specifications) 9. USDA Rural Development, "Electronic Preliminary Engineering Report" \(2018\) — [https://www\.rd\.usda\.gov/programs\-services/water\-environmental\-programs/electronic\-preliminary\-engineering](https://www.rd.usda.gov/programs-services/water-environmental-programs/electronic-preliminary-engineering) 10. Morrison\-Maierle, "How to Prepare a Preliminary Engineering Report" \(2023\) — [https://m\-m\.net/insights/how\-to\-prepare\-a\-preliminary\-engineering\-report/](https://m-m.net/insights/how-to-prepare-a-preliminary-engineering-report/) 11. Bluebeam, "Building the Future: Bluebeam AEC Technology Outlook 2026" \(2025\) — [https://press\.bluebeam\.com/2025/10/new\-bluebeam\-report\-shows\-early\-ai\-adopters\-in\-aec\-seeing\-significant\-roi\-despite\-uneven\-adoption/](https://press.bluebeam.com/2025/10/new-bluebeam-report-shows-early-ai-adopters-in-aec-seeing-significant-roi-despite-uneven-adoption/) 12. Bluebeam, "Building the Future: Bluebeam AEC Technology Outlook 2026" \(2025\) — [https://press\.bluebeam\.com/2025/10/new\-bluebeam\-report\-shows\-early\-ai\-adopters\-in\-aec\-seeing\-significant\-roi\-despite\-uneven\-adoption/](https://press.bluebeam.com/2025/10/new-bluebeam-report-shows-early-ai-adopters-in-aec-seeing-significant-roi-despite-uneven-adoption/) 13. Bluebeam, "Building the Future: Bluebeam AEC Technology Outlook 2026" \(2025\) — [https://press\.bluebeam\.com/2025/10/new\-bluebeam\-report\-shows\-early\-ai\-adopters\-in\-aec\-seeing\-significant\-roi\-despite\-uneven\-adoption/](https://press.bluebeam.com/2025/10/new-bluebeam-report-shows-early-ai-adopters-in-aec-seeing-significant-roi-despite-uneven-adoption/) 14. Bluebeam, "Building the Future: Bluebeam AEC Technology Outlook 2026" \(2025\) — [https://press\.bluebeam\.com/2025/10/new\-bluebeam\-report\-shows\-early\-ai\-adopters\-in\-aec\-seeing\-significant\-roi\-despite\-uneven\-adoption/](https://press.bluebeam.com/2025/10/new-bluebeam-report-shows-early-ai-adopters-in-aec-seeing-significant-roi-despite-uneven-adoption/) 15. McKinsey & Company, "Artificial intelligence: Construction technology's next frontier" \(2018\) — [https://www\.mckinsey\.com/capabilities/operations/our\-insights/artificial\-intelligence\-construction\-technologys\-next\-frontier](https://www.mckinsey.com/capabilities/operations/our-insights/artificial-intelligence-construction-technologys-next-frontier) 16. American Society of Civil Engineers, "Architecture, engineering, construction sector slow to adopt AI, survey shows" \(2025\) — [https://www\.asce\.org/publications\-and\-news/civil\-engineering\-source/article/2025/12/18/architecture\-engineering\-construction\-sector\-slow\-to\-adapt\-ai\-survey\-shows](https://www.asce.org/publications-and-news/civil-engineering-source/article/2025/12/18/architecture-engineering-construction-sector-slow-to-adapt-ai-survey-shows) --- Source: https://dancumberlandlabs.com/blog/csi-number-construction/