# The Input Layer Is the Strategy **By Dan Cumberland** · Published May 7, 2026 · Categories: AI Strategy > Poor data quality costs the average professional services firm $12-15 million a year[^1], and AEC firms are quietly leading the loss column. Before you buy... ## The $12 Million Question No One in Your Firm Is Asking Poor data quality costs the average professional services firm $12\-15 million a year[1](/blog/blog-3-layer-architecture#ref-1), and AEC firms are quietly leading the loss column\. Before you buy another AI tool, look at the data you'd feed it\. That number isn't a forecast\. It's an invoice your firm is already paying, line by line, in re\-bills, blown estimates, and hours your principals spend reconciling reports that should reconcile themselves\. Gartner research puts the average annual cost at $12\.9 million[2](/blog/blog-3-layer-architecture#ref-2)\. The information sector loses roughly $12,161 per employee per year to dirty data, and employees spend up to 27% of their time correcting it[2](/blog/blog-3-layer-architecture#ref-2)\. Multiply that by your billable headcount\. Now layer in what AI does to that math\. Trimble research found that 85% of AI implementation failures trace back to data quality, not tool selection[3](/blog/blog-3-layer-architecture#ref-3)\. ACEC's 2025 readiness report showed 87% of contractors expect AI to meaningfully impact construction— and only 19% have actually adapted workflows[4](/blog/blog-3-layer-architecture#ref-4)\. - **$12\-15M annual cost** of poor data quality \(industry average\) - **$12,161 per employee** lost to dirty data each year - **27% of employee time** spent correcting bad inputs Most AEC firms don't have an AI problem\. They have a data problem they're about to spend money pretending to solve with AI\. The reason these losses keep compounding has a name\. It's called the [3 layer architecture](/services/ai-strategy), and most firms only think about one layer\. They're also quietly absorbing the [hidden costs of AI projects](/blog/hidden-costs-ai-projects) that get blamed on the wrong line item\. ## What Is the 3 Layer Architecture \(and Why Most Firms Skip Layer One\) A 3 layer architecture for AI has three components: the **input layer** \(the data you feed the system\), the **processing layer** \(the model that finds patterns\), and the **output layer** \(the insight or decision the system returns\)\. Input quality sets the ceiling for everything above it\. That's the whole frame\. Three layers, one dependency rule: 1. **Input layer\.** Your timesheet entries, project codes, cost codes, phase definitions, billing data— the raw material\. 2. **Processing layer\.** The AI model that ingests that material and learns patterns from it\. 3. **Output layer\.** The forecast, the recommendation, the dashboard tile your principals actually look at\. The dependency runs in one direction: output ≤ processing ≤ input\. Whatever the input layer can't see, the processing layer can't learn, and the output layer can't tell you\. This is the GIGO principle— garbage in, garbage out— applied at scale\. The research is consistent: IBM finds flawed inputs produce unreliable outputs[5](/blog/blog-3-layer-architecture#ref-5)\. Saifr quantifies it — a 15% inaccuracy rate in training data severely degrades model performance, and up to 87% of automation systems never reach production[6](/blog/blog-3-layer-architecture#ref-6)\. If your input layer is inconsistent, your processing layer learns noise, and your output layer hands you confident nonsense\. AimMultiple identifies poor data as the primary cause of AI hallucinations[7](/blog/blog-3-layer-architecture#ref-7)\. Monograph points out the practical version: AI cannot distinguish a genuine pattern from a data gap[8](/blog/blog-3-layer-architecture#ref-8)\. It just averages over the chaos\. Here's where the strategic miss happens: ```html-table
Where AEC firms investWhere the leverage actually lives
Layer 3 (output): dashboards, BI tools, AI assistantsLayer 1 (input): cost codes, phases, timesheet vocab
Layer 2 (processing): vendor model selection, pilotsLayer 1 (input): governance and standardization
New tool budgetThe data that already exists
``` The 3 layer architecture is simple: input, processing, output\. Input quality determines what the other two layers can possibly produce\. Choosing not to fix layer one is a strategic decision— you're just making it by default\. To see why the input layer dominates, look at what happens when an AEC firm asks AI a profitability question with messy data\. ## Why Inconsistent Project Codes Break AI Pattern Recognition When the same activity is coded three different ways in your ERP, AI sees three different activities\. It can't learn that they're the same— it learns that your firm is chaotic, and reflects that chaos back as unreliable forecasts\. Picture the same design development task showing up across three projects as "Design Dev," "DD," and "design\-development\." A senior PM reads those three entries and sees one activity\. An AI model reads three\. Run that across 18 months of timesheets, and the patterns that should drive your next bid— hours per phase, true cost per deliverable, recovery rates— sit buried under coding noise the model can't unbury\. > AI doesn't fix coding inconsistency\. It magnifies it\. Procore's analysis of construction cost codes makes the operational version of this point: standardized cost codes are the building blocks of estimating accuracy and reduce the gap between estimated and actual costs— which is exactly where profitability erodes[9](/blog/blog-3-layer-architecture#ref-9)\. Unanet is blunter still: an AI\-ready ERP requires standardized master data— project codes, rates, phases— before any model gets useful[10](/blog/blog-3-layer-architecture#ref-10)\. HSO's analysis of Deltek Vision to Vantagepoint migrations traces integration failures back to upstream data inconsistency that nobody caught until the new system surfaced it[11](/blog/blog-3-layer-architecture#ref-11)\. What inconsistent input does to AI output, in practical terms: - **Profitability signals** get buried under coding noise - **Forecasting models** average over real patterns and never find them - **Anomaly detection** flags everything or nothing - **Resource recommendations** rest on phantom precedents There's a second symptom: shadow data\. Construction Dive reports 77% of construction companies struggle with inconsistent quality processes, and the practical reason is that spreadsheets exist because the ERP doesn't capture how work actually gets done[12](/blog/blog-3-layer-architecture#ref-12)\. Standardized cost codes are the single highest\-leverage AI investment most AEC firms aren't making\. If the input layer is so consequential, why do most AEC firms skip it? Because the alternative looks faster\. ## The AI Tool Paradox: Why Buying Faster Means Failing Faster Adding AI to a firm with bad data inputs doesn't fix the problem— it makes the failure faster, more confident, and more expensive\. This is the AI tool paradox at the heart of every stalled AEC pilot\. ACEC's readiness research shows the gap clearly: 87% of contractors expect AI to impact their business, only 19% have adapted workflows[4](/blog/blog-3-layer-architecture#ref-4)\. Trimble found 77% believe AI will reshape construction, but only about 20% have advanced adoption, and roughly one\-third have no integrated AI strategy at all[3](/blog/blog-3-layer-architecture#ref-3)\. Saifr's number— up to 87% of automation systems never reaching production— isn't a vendor problem[6](/blog/blog-3-layer-architecture#ref-6)\. It's a data problem dressed in vendor clothing\. > AI is intellectual augmentation\. If the intellect you're augmenting is fed bad inputs, you've augmented your wrong answers\. The pattern shows up the same way every time: - Vendor pitch lands\. The demo is impressive\. - Pilot scoped against the messiest data the firm has\. - Output looks plausible but doesn't match what principals know to be true\. - Adoption stalls\. Tool gets blamed\. Budget moves on\. - The next vendor pitch lands\. This is where AEC AI readiness gets confused with AEC AI procurement\. And most firms are buying AI to skip the data work\. The data work is the AI work\. Just because it's easy to buy a tool doesn't mean it's good to buy one yet\. A working [AI decision framework for founders](/blog/ai-decision-framework-founders) starts at the input layer because everything else cantilevers off it\. The good news: the input layer is the most fixable part of this— if you sequence it right\. ## What to Fix First — A 3 Layer Architecture Action Plan The leverage points in your input layer are not exotic\. They're cost codes, project phases, and timesheet descriptions— the same operational discipline that determines profitability today\. Five steps, in order: 1. **Audit project and cost code consistency\.** Pull a year of data and count the variants\. How many ways is "design development" coded? How many phase codes are functional duplicates? You're not fixing anything yet— you're sizing the problem\. 2. **Standardize the top 20 codes that drive 80% of billable hours\.** Don't try to boil the ocean\. Procore's research is right that cost codes are the foundation of profitability tracking, and the Pareto subset is where the leverage lives[9](/blog/blog-3-layer-architecture#ref-9)\. 3. **Tighten timesheet descriptions\.** Minimum field standards, controlled vocabularies, dropdowns instead of free text where it makes sense\. This is the single highest\-frequency input in your firm\. 4. **Establish an owner\.** Even if it's part\-time\. Trimble is explicit that data governance is a critical precondition for AI— and governance without an accountable human is just a slide[3](/blog/blog-3-layer-architecture#ref-3)\. See [AI governance strategy](/blog/ai-governance-strategy) for what that role actually looks like\. 5. **Only then evaluate AI tools, with the input layer as the spec\.** When you do shop, [measuring AI success](/blog/measuring-ai-success) starts with the question: what would this tool need from our input layer to actually work? > Fix the input layer first, and your existing reports get sharper before any AI tool arrives\. The compounding benefit shows up before the AI does\. Standardized codes mean clearer profitability visibility on your next P&L\. Cleaner timesheets mean tighter estimates on your next bid\. The order is data, then strategy, then tool— not the other way around\. This is the work most firms try to skip\. It's also the work that decides whether AI ever actually works for them\. ## FAQ The questions AEC principals ask most about AI readiness all come back to the input layer\. ### What is the 3 layer architecture for AI? A 3 layer architecture has three components: an input layer \(the data you feed the system\), a processing layer \(the model that finds patterns\), and an output layer \(the insight returned\)\. Input quality determines what the other two layers can produce\. Skip the input layer and the rest of your stack inherits its problems\. ### Why does data quality matter so much for AI? AI learns by recognizing patterns in historical data\. When inputs are inconsistent, AI sees noise instead of patterns and produces unreliable output— the GIGO principle applied at scale\. IBM and Saifr both identify data quality as the most common cause of AI failure[5](/blog/blog-3-layer-architecture#ref-5)[6](/blog/blog-3-layer-architecture#ref-6)\. ### What does poor data quality cost an AEC firm? Industry research from IBM and Gartner places the average annual cost at $12\-15 million for professional services firms[1](/blog/blog-3-layer-architecture#ref-1)[2](/blog/blog-3-layer-architecture#ref-2)\. Employees spend up to 27% of their time correcting bad data[2](/blog/blog-3-layer-architecture#ref-2)\. The cost shows up as re\-bills, blown estimates, and reconciliation hours that compound quietly\. ### What should we fix first to be AI\-ready? Start with cost code standardization and timesheet description quality\. These are the highest\-leverage items for both project profitability today and AI accuracy tomorrow\. Procore and Trimble both name standardized master data as a precondition for any reliable AI use[9](/blog/blog-3-layer-architecture#ref-9)[3](/blog/blog-3-layer-architecture#ref-3)\. ### Why do so many AEC AI pilots fail? About 85% of AI failures trace to data quality, not tool selection[3](/blog/blog-3-layer-architecture#ref-3)\. Most firms buy tools to skip the data work\. The data work is the AI work\. Pilots scoped on top of inconsistent inputs produce output that doesn't match reality, and the tool gets blamed for the inputs it inherited\. ## The Strategic Reframe Your input layer is your AI strategy, whether you've decided that or not\. The firms that win with AI in the next two years aren't the ones with the best tools\. They're the ones with the cleanest inputs\. The 3 layer architecture isn't a diagram to admire— it's a decision frame\. Your data discipline today sets the ceiling for what any future model can do for you\. And the choice not to fix it is, itself, the choice\. If mapping the input layer of your firm feels like reading the label from inside the bottle, that's normal— it's hard to see your own coding chaos until someone outside it does\. That's the kind of work [our AI strategy services](/services/ai-strategy) are built for\. ## References 1. IBM, "Why AI Data Quality Is Key To AI Success" \(2024\-2025\) — [https://www\.ibm\.com/think/topics/ai\-data\-quality](https://www.ibm.com/think/topics/ai-data-quality) 2. Revefi, "The Cost of Poor Data Quality on Business Operations" \(citing Gartner\) \(2024\-2025\) — [https://www\.revefi\.com/blog/business\-operations\-poor\-data\-quality\-cost](https://www.revefi.com/blog/business-operations-poor-data-quality-cost) 3. Trimble, "Implementing AI Solutions in AEC: A Guide to Boosting Efficiency and Innovation" \(2024\-2025\) — [https://www\.trimble\.com/blog/construction/en\-US/article/implementing\-ai\-solutions\-aec\-guide\-boosting\-efficiency\-innovation](https://www.trimble.com/blog/construction/en-US/article/implementing-ai-solutions-aec-guide-boosting-efficiency-innovation) 4. ACEC, "AEC's AI Moment: New Report Reveals Readiness Gap and Roadmap" \(2025\) — [https://www\.acec\.org/podcast/aecs\-ai\-moment\-new\-report\-reveals\-readiness\-gap\-and\-roadmap/](https://www.acec.org/podcast/aecs-ai-moment-new-report-reveals-readiness-gap-and-roadmap/) 5. IBM, "Why AI Data Quality Is Key To AI Success" \(2024\-2025\) — [https://www\.ibm\.com/think/topics/ai\-data\-quality](https://www.ibm.com/think/topics/ai-data-quality) 6. Saifr, "Garbage In, Garbage Out: Why Data Quality Is Critical to AI" \(2024\-2025\) — [https://saifr\.ai/blog/garbage\-in\-garbage\-out\-why\-data\-quality\-is\-critical\-to\-ai](https://saifr.ai/blog/garbage-in-garbage-out-why-data-quality-is-critical-to-ai) 7. AimMultiple, "AI Data Quality in 2026: Challenges & Best Practices" \(2025\-2026\) — [https://research\.aimultiple\.com/data\-quality\-ai/](https://research.aimultiple.com/data-quality-ai/) 8. Monograph, "AI in Construction Estimating: Accuracy & ROI Guide" \(2024\) — [https://monograph\.com/blog/ai\-construction\-estimating\-accuracy\-roi\-guide](https://monograph.com/blog/ai-construction-estimating-accuracy-roi-guide) 9. Procore, "Construction Cost Codes: Best Practices and Industry Insights" \(2024\-2025\) — [https://www\.procore\.com/library/construction\-cost\-codes](https://www.procore.com/library/construction-cost-codes) 10. Unanet, "How AI Is Changing the Way A&E Firms Use ERP Data" \(2024\-2025\) — [https://unanet\.com/blog/how\-ai\-is\-changing\-the\-way\-ae\-firms\-use\-erp\-data](https://unanet.com/blog/how-ai-is-changing-the-way-ae-firms-use-erp-data) 11. HSO, "Moving from Deltek Vision to Vantagepoint: Expectation vs\. Reality" \(2024\-2025\) — [https://www\.hso\.com/blog/moving\-from\-deltek\-vision\-to\-deltek\-vantagepoint\-expectation\-vs\-reality](https://www.hso.com/blog/moving-from-deltek-vision-to-deltek-vantagepoint-expectation-vs-reality) 12. Construction Dive, "Data Quality Is Holding Construction Back\. AI May Help\." \(2024\) — [https://www\.constructiondive\.com/news/ai\-improve\-construction\-data/743361/](https://www.constructiondive.com/news/ai-improve-construction-data/743361/) --- Source: https://dancumberlandlabs.com/blog/3-layer-architecture/