How to Tag Project Sheets So Search Actually Works

What Broken Search Actually Costs

Knowledge workers spend roughly 19% of the workweek looking for internal information, per McKinsey Global Institute research on interaction-worker time use¹. IDC research from an earlier era put the figure at around 30% of the knowledge worker's day². AEC-specific data isn't measured publicly, but the order of magnitude applies to architecture firms— and arguably more, given the volume of drawings, specifications, and precedents the average project produces.

Order of magnitude: Knowledge workers lose 19–30% of working time to search. The AEC-specific number is unmeasured. Don't trust anyone who quotes one.

An architect billing $180 per hour who spends one hour a day searching for past project content is a $45,000-per-year leak— per person. Multiply that by the project architects, designers, and project managers on staff at a fifty-person firm and the number gets uncomfortable fast. Do the arithmetic yourself with your actual rates and headcount; the answer won't be small.

The broader context matters too. Construction productivity has grown only about 1% annually for two decades, while the global economy has grown 2.8%³. Digital transformation can drive 14–15% productivity gains and 4–6% cost reductions in the sector, per the same McKinsey analysis. Knowledge management is one of the few places a firm can recover that productivity gap without changing what it builds. And it's the place most firms have left mostly untouched.

The figures are old. The pattern hasn't changed.

If naming alone isn't the answer, what is? Tagging— but tagging at three different layers, and only one of them is the filename.

The Three Layers of Tagging That Actually Drive Findability

Findability in an AEC firm depends on three tagging layers: the file name (NCS or ISO 19650 format), structured metadata attributes (project ID, discipline, sheet type, revision, project phase), and extractable in-document content (title block text, sheet notes, annotation text). A modern search system— especially an AI-driven one— needs all three. Tag at one layer and you've automated a filing cabinet. Tag at all three and search starts working.

Layer 1 — Filename. The U.S. National CAD Standard sheet identifier is a discipline designator, a sheet type digit, and a two-digit sequence number— for example, A101 is an Architectural Floor Plan, sheet 01⁴. NCS is published by the National Institute of Building Sciences and incorporates the AIA CAD Layer Guidelines and the CSI Uniform Drawing System (Modules 1–8)⁵. ISO 19650 is the complementary international standard governing project information across the asset lifecycle. It defines a Common Data Environment as "an agreed source of information for any given project or asset, for collecting, managing and disseminating each information container through a managed process"⁶. Filenames identify. They don't classify.

Layer 2 — Structured metadata attributes. This is the layer most firms underinvest in. Project ID, discipline, sheet type, document type, revision, project phase, project type or building use, completion year, code edition— these live in the DMS, not in the filename. In Autodesk BIM 360 / Autodesk Construction Cloud, they're called Custom Attributes and they're searchable inside Docs⁷. Newforma has fields. SharePoint has columns. Different name, same job. Vendor analysis suggests metadata-driven search can cut document retrieval time by up to 30%⁸. That figure is from a vendor— treat it as directional, not gospel— but the principle holds.

Layer 3 — In-document content. The text inside the drawing. Title block, sheet notes, schedule text, annotations. Surfaced via OCR for PDFs and direct text extraction for vector formats. Semantic search relies on this layer when somebody types "find me precedents with a curtain wall above 40 feet" and there's nothing in the filename or the custom attributes that would catch it. The AIA file-naming guidance distinguishes Model files (drawn full size, plotted to scale) from Sheet files (title blocks, schedules, the non-model parts of the drawing set)⁹; it's the Sheet files where Layer 3 tagging matters most, because that's where the readable text lives.

Tag at three layers and the metadata becomes the source of truth for what the firm has done. Tag at one and you have a faster way to retrieve the same chaos.

Each layer has its own failure mode. The most common one— and the reason most tagging programs collapse— is the layer everyone overinvests in: the filename.

Why Naming Standards Break in Practice (The Compliance Reality)

Tagging compliance drops the moment a deadline gets tight. Any system that assumes every new sheet gets named and tagged correctly at creation will quietly fail somewhere between 60% and 80% of the way through the project— when it matters most. Design for the failure mode, not the spec sheet.

A schema that survives Thursday at 11pm is worth more than a schema that wins the audit.

The pattern is identical across firms. Convention adopted firmwide. Training delivered. Audits scheduled. And then a deadline hits, and shorthand wins. There are three categories of compliance failure, and any working schema has to survive all three:

• Shorthand filenames during deadline crunch. "A101-final-FINAL.pdf" lands in the project folder at 11pm because that's the file the contractor needed. Three months later, nobody knows which "FINAL" is the actual final.
• Metadata fields skipped at upload. Custom attributes that aren't required get blank. Required fields that slow upload get worked around. The fields most relevant to retrieval are the ones most likely to be skipped under pressure.
• Legacy archives that predate the convention. Twenty years of drawings, three or four CAD-manager regimes, multiple naming systems half-applied. No amount of new-project enforcement touches this.

The honest fix has three parts. Enforce only the fields that must be enforced— five to seven, not thirty. Automate everything else through title-block OCR and semantic extraction. And accept that 100% compliance is a fantasy worth letting go of. Friedman & Partners frame the underlying point well: AI-powered search effectiveness in AEC firms depends on the quality, structure, and completeness of the firm's knowledge foundation¹⁰. Completeness, though, can come from extraction. It doesn't have to come from enforcement. See our notes on building AI culture for more on what survives team adoption pressure.

If the fields you enforce are the ones that survive deadline pressure, the question becomes: which fields? The minimum-viable schema for a mid-sized AEC firm is shorter than most CAD managers think.

The Minimum-Viable Tagging Schema (5–7 Fields, Not 30)

A mid-sized AEC firm needs five to seven enforced metadata fields, applied at project setup rather than at every sheet upload. The rest of the metadata— sheet-level attributes, in-document content— should be extracted automatically. Enforce what humans must enter; extract everything else. Five fields you enforce will beat thirty fields you wish you enforced.

The recommended enforced fields:

1. Project ID — Anchors every artifact to a single project record. Without this, nothing downstream works.
2. Discipline — A / S / M / E / P / C / L, aligned to NCS discipline designators⁵. Drives the largest first-cut filter.
3. Project Phase — Schematic / DD / CD / CA / As-Built. Critical for precedent retrieval; "find me the DD set for our last K–12" is a daily query.
4. Document Type — Drawing / Specification / Report / Submittal / RFI. Separates the deliverable from the conversation.
5. Revision / Date — Enables current-version retrieval and decisively resolves the "A101-final-FINAL" problem.
6. Project Type / Building Use (optional, high-value) — K–12, healthcare, multifamily, lab, mixed-use. This is what makes precedent search work.
7. Code Edition / Year of Completion (optional, high-value) — Indispensable for code-research workflows and historical comparisons.

The point of a minimum-viable schema isn't to capture everything— it's to capture the five fields a $20M–$100M AEC firm cannot operate without and let extraction handle the rest.

Note: This is a practitioner recommendation drawn from how the strongest mid-market AEC firms operate, informed by NCS Module 1 sheet identification and ISO 19650's information-container structure. It is not a quoted standard. Adapt to your firm's actual project workflows.

Operationally, each enforced field lives in the DMS as a custom attribute— BIM 360 Custom Attributes, Newforma fields, SharePoint columns— populated at project-folder creation rather than at every sheet upload. Setup is a one-time act per project. Upload is a thousand times per project. Enforce where the cost of compliance is paid once; extract where it's paid daily.

Enforcing five fields on new projects is solvable. The harder question is what to do with the projects that predate the schema— which is to say, most of the firm's archive.

Fixing the Legacy Archive Without Re-Tagging 20 Years of Drawings

You cannot manually re-tag a twenty-year archive, and you don't have to. Title-block OCR converts the structured information already drawn on every sheet (project name, sheet number, discipline, date) into searchable metadata, and semantic search retrieves by meaning rather than exact-keyword match. Together they cover the legacy-archive problem at roughly 85–90% field-level accuracy out of the box.

The approach has three parts:

1. Title-block OCR (the deterministic side). Apryse's CAD Title Block Extraction uses a custom-trained model to identify and extract data from PDF title blocks and output structured JSON metadata suitable for project databases and downstream workflows¹¹. Autodesk BIM 360 / ACC includes a title-block extraction tool that uses OCR to auto-populate Custom Attributes from sheets uploaded to the Plans section⁷. The output flows directly into Layer 2 of your tagging system without anyone typing.
2. Semantic search (the inferential side). Vector embeddings retrieve by meaning. Knowledge Architecture's Synthesis is purpose-built AI search for AEC firms; the company now serves more than 160 AEC clients¹². AI search of this kind produces summaries with citations linking back to source documents, which is essential for principal-level trust— and for an architect who needs to verify a precedent before referencing it in a proposal.
3. The combination. OCR fills in the structured-metadata layer for old drawings. Semantic search uses the in-document-content layer for the queries that no filename will ever match: "find me the curtain-wall detail with the cantilevered mullion on a healthcare project." Either alone helps. Together they make the legacy archive usable.

Title-block OCR is triage, not perfection. A 2025 research pipeline combining Faster R-CNN and GPT-4o achieved 88.2% accuracy on title-block detection across both vector and historical (scanned and handwritten) drawings¹³— strong, but 12% wrong. Treat the output as a starting point; review high-value precedents by hand.

The accuracy caveat matters. At a 12% error rate, two thousand sheets will yield roughly 240 fields that need human attention. In practical terms, that's a manageable triage workload— particularly if you weight the review toward precedent-eligible projects and recent code-edition years, rather than the 1998 tenant-improvement set nobody will ever look at again. Set realistic expectations with your IT lead and budget the human-review time as part of the rollout, not as an afterthought.

Friedman & Partners' framing is the right honest caveat: AI search effectiveness in AEC depends on the underlying knowledge foundation¹⁰. OCR plus semantic search is the bridge between "no tagging" and "useful retrieval," but it doesn't replace getting the new-project tagging right. Friedman reports a firm where a sheet-index review process dropped from six hours to under one hour using AI¹⁰; treat that as an illustrative example with attribution, not as a benchmark you should expect to replicate. See our notes on measuring AI success for the metrics that actually track whether an OCR rollout is working.

Five fields enforced on new projects, OCR plus semantic search on the archive. The remaining question is where to invest first— and that depends on the firm's stage.

Where to Invest First (A Decision Framework for Principals at an Architecture Firm)

Most $20M–$100M AEC firms should sequence three investments in this order: define the 5–7-field schema and embed it in project-setup templates first, deploy title-block OCR against the archive second, and add semantic / AI search third. Skipping ahead is the most common failure mode— and the most expensive. Buying AI search before you've enforced a tagging schema is buying a faster way to retrieve the same chaos.

1. Step 1 — Define and embed the schema (30–90 days). Pick the 5–7 fields. Build a project-setup template in the DMS that forces them at project creation. Train one project-coordinator role to own the setup checklist. Cost: low. Risk: low. This is the step principals are tempted to skip because it doesn't feel like "AI." It's the step that determines whether the AI investment downstream pays off.
2. Step 2 — Title-block OCR on the archive (60–120 days). Pilot against one office or one project type before firmwide rollout. Budget for a 10–15% error rate and human spot-checks on high-value precedents. Cost: moderate. Risk: low if scoped as triage rather than truth. Don't pay anyone to chase the last 5% of accuracy; chase the precedents that drive proposals instead.
3. Step 3 — Semantic / AI search layered on top (90–180 days). Once Steps 1 and 2 are in place, the AI-search vendor's product has clean data to work with. Per AIA 2024 research, only about 8% of architecture firms have implemented AI solutions, with roughly 20% currently in implementation, driven significantly more by large firms with 50+ employees¹⁴. Translation: $20M–$100M firms are at this decision point right now. The firms that buy AI search before fixing the foundation get the worst outcomes.

The investment decision the principal is really making is whether this is a tooling problem or a workflow problem. It's mostly workflow. Tooling helps, but only if the workflow under it is sane. The hidden costs of AI projects tend to come from skipped foundations, not from the AI tools themselves. For a deeper read on how to choose between sequencing options at a strategic level, see the AI decision framework for founders.

Three things to remember: tag at all three layers, enforce only the fields humans must enter, and sequence the AI investment after the foundation— not before it. The firms that get this right will retrieve their precedents in seconds. The firms that buy AI search first will retrieve the same chaos faster.

These decisions are easier to make on paper than on a Monday morning.

When to Get Help

If sequencing this investment for your firm feels worth a conversation rather than a vendor pitch, that's what implementation partners are for. The right partner maps the schema to your actual project workflows, scopes the OCR pilot against your archive's real condition, and helps you avoid buying AI search on top of unsolved foundations. An implementation partner maps the schema to the workflow, not the workflow to the vendor.

We work with $20M–$100M AEC firms on exactly this question. If a second pair of eyes on the sequence would be useful, Dan Cumberland Labs is set up for the advisory version of that conversation.

FAQ

What is the NCS sheet numbering format?

The NCS sheet identifier is a discipline designator (one or two letters), a sheet type digit (0–9), and a two-digit sequence number— for example, A101 is an Architectural Floor Plan, sheet 01⁴. The standard is published by the National Institute of Building Sciences and incorporates the AIA CAD Layer Guidelines and the CSI Uniform Drawing System Modules 1–8⁵.

What's the difference between NCS and ISO 19650?

NCS governs U.S. sheet identification and CAD layer practice for drawing sets. ISO 19650 governs project information across the full asset lifecycle and defines the Common Data Environment as "an agreed source of information for any given project or asset, for collecting, managing and disseminating each information container through a managed process"⁶. They are complementary; many U.S. firms follow NCS for drawings and map to ISO 19650 for project-level information management.

Can AI search work on old architecture drawings that aren't tagged?

Yes. Title-block OCR converts the structured information drawn on each sheet (project name, sheet number, discipline) into searchable metadata, and semantic search retrieves by meaning rather than exact filename match⁷¹¹. A 2025 research pipeline reported 88.2% accuracy on title-block detection across vector and scanned/handwritten drawings¹³, so treat the output as triage-grade and review high-value precedents by hand.

What's the minimum metadata to enforce on new sheets?

Five to seven fields, applied at project setup rather than at every sheet upload: Project ID, Discipline, Project Phase, Document Type, Revision, and optionally Project Type and Code Edition. Sheet-level data (sheet type, title, scale) should be extracted automatically rather than enforced. Custom attributes in BIM 360 / ACC are the most common operational home for these fields⁷.

How many AEC firms have implemented AI today?

About 8% of architecture firms have implemented AI solutions, with roughly 20% currently working on implementation, per AIA 2024 research¹⁴. Adoption is driven significantly more by large firms with 50+ employees, which means $20M–$100M AEC firms are at the decision point right now.

Should we adopt a tagging schema before buying AI search software?

Yes. AI-powered search effectiveness depends on the quality, structure, and completeness of the firm's underlying knowledge foundation¹⁰. Buying AI search before enforcing a minimum schema means accelerating retrieval of the same chaos. Sequence the investment: schema first, title-block OCR second, AI search third.

References

1. McKinsey Global Institute, "The Social Economy: Unlocking Value and Productivity Through Social Technologies" (2012) — https://www.mckinsey.com/industries/technology-media-and-telecommunications/our-insights/the-social-economy
2. IDC, "The High Cost of Not Finding Information" (2001) — https://computhink.com/wp-content/uploads/2015/10/IDC20on20The20High20Cost20Of20Not20Finding20Information.pdf
3. McKinsey & Company, "Decoding Digital Transformation in Construction" (2024) — https://www.mckinsey.com/capabilities/operations/our-insights/decoding-digital-transformation-in-construction
4. Archtoolbox, "Construction Document Sheet Numbers and Order" (2020) — https://www.archtoolbox.com/construction-document-sheet-numbers/
5. National Institute of Building Sciences, "NCS Content | United States National CAD Standard - V6" (2014) — https://www.nationalcadstandard.org/ncs6/content.php
6. Autodesk University, "ISO 19650, the Common Data Environment, and Autodesk Construction Cloud" (2023) — https://www.autodesk.com/autodesk-university/article/ISO-19650-Common-Data-Environment-and-Autodesk-Construction-Cloud
7. Autodesk, "Assign Metadata to Files Within BIM 360 Document Management" (2024) — https://www.autodesk.com/support/technical/article/caas/tsarticles/ts/3laVEBfATQouerLULLKGW.html
8. BIMcollab, "How to Simplify Document Management with Metadata" (2024) — https://www.bimcollab.com/en/resources/blog/how-to-simplify-document-management-with-metadata/
9. American Institute of Architects, "Keys to Classifying Project Files" (2023) — https://www.aia.org/resource-center/keys-classifying-project-files
10. Friedman & Partners, "The Rise of AI Search in AEC Knowledge Management" (2025) — https://friedmanpartners.com/the-rise-of-ai-search-in-aec-knowledge-management/
11. Apryse, "Automate CAD Title Block Extraction from PDFs" (2024) — https://apryse.com/blog/automate-cad-title-block-extraction
12. Knowledge Architecture, "Synthesis: AI Search for Architecture, Engineering, and Construction (AEC) Firms" (2025) — https://www.knowledge-architecture.com/synthesis-ai-search
13. arXiv, "Title Block Detection and Information Extraction of Building Drawings" (2025) — https://arxiv.org/pdf/2504.08645
14. American Institute of Architects, "New Research Explores Perceptions and Opportunities for Artificial Intelligence" (2024) — https://www.aia.org/about-aia/press/new-research-explores-perceptions-and-opportunities-artificial-intelligence