The Low-Risk Test— 4 Criteria
A spec section qualifies as low-risk for AI-assisted drafting when it meets four conditions: heavily procedural content, standards-driven structure, low design judgment, and easy verification against an authoritative reference. Miss any one of these and the section moves up the risk curve.
The risk in a spec section comes from design judgment and hidden consequence— not from word count. A long, procedural Part 1 General section is safer than a short paragraph buried in a performance spec.
| Criterion | What It Means | Passes / Fails |
|---|---|---|
| Procedural | Describes how — process, format, timing — not what specific product | ✅ Submittal procedures · ❌ Mechanical performance |
| Standards-driven | References codified standards (ASTM, ANSI, MPI, ASHRAE, AWI) | ✅ Reference standards lists · ❌ Custom performance criteria |
| Low design judgment | Carries no structural, performance, or means-and-methods decisions | ✅ Toilet accessories · ❌ Concrete mix design |
| Easily verifiable | Outputs cross-check against a cut sheet, standard, or office master in minutes | ✅ Paint product schedule · ❌ Hidden MEP routing |
The framework matters more than the list. Apply the test to a section you're considering, get four yeses, and you've found a candidate. Get three yeses and a hard no on design judgment? Keep it human.
This matters because AI hallucinations— responses containing false or incorrect information— "appear indistinguishable from factually correct responses"4. Studies cited in construction law analyses put hallucination rates between 1% and 27%5. Verification only works when verification is fast. That's why the fourth criterion is non-negotiable.
Five sections pass this test cleanly today. Here they are.
The 5 Sections Where Drafting AI Works Today
Five CSI MasterFormat sections meet the four-criterion test today: 01 33 00 Submittal Procedures, 01 42 19 Reference Standards, 09 91 23 Interior Painting, 10 28 00 Toilet/Bath/Laundry Accessories, and 07 92 00 Joint Sealants. These five share the same DNA: procedural, standards-driven, manufacturer-cut-sheet content with low design judgment and fast verification.
| Section | Why Low-Risk | What AI Drafts | What the Human Verifies |
|---|---|---|---|
| 01 33 00 Submittal Procedures | Procedural; industry-standard structure | First-pass submittal list, format, timing | Owner-specific submittal requirements |
| 01 42 19 Reference Standards | Pure compilation work | Compiled list of cited ASTM/ANSI/MPI/ASHRAE | Standard numbers and current editions |
| 09 91 23 Interior Painting | Manufacturer-driven; codified VOC limits | Product schedule, MPI categories, prep notes | Product designations against current data sheets |
| 10 28 00 Toilet/Bath/Laundry Accessories | Off-the-shelf commodity products | Schedule from cut sheets | ADA/accessibility callouts |
| 07 92 00 Joint Sealants | Highly standardized (ASTM C920) | Three-part structure, sealant-type-to-application | Movement classes for unusual joint conditions |
01 33 00— Submittal Procedures
This is procedural content at its purest. What gets submitted, in what format, by when. The skeleton is industry-standard, with mature reference templates in the public domain6. AI compiles a strong first draft against that skeleton.
The human verification element matters here. Conspectus has rightly warned for years that Division 01 should not be treated as boilerplate7. Owner-specific submittal requirements, custom general conditions, and project-specific coordination clauses all override the template. AI drafts the procedural backbone. The specifier checks the project-specific overlay.
01 42 19— Reference Standards
Compilation work, almost entirely. The section lists every ASTM, ANSI, MPI, ASHRAE, and AWI standard referenced across the project. AI handles compilation and updating well.
There's one trap. AI does fabricate standard numbers occasionally— a credible-looking ASTM citation that doesn't actually exist, or an outdated edition presented as current. Verification is fast (a standards-database lookup takes seconds), but it's mandatory.
09 91 23— Interior Painting
Manufacturer-driven content with codified rules. MPI categories are well-defined. VOC limits are stable: per OTC, interior flat paints cap at 100 g/L, enamels at 150 g/L, stains at 250 g/L, and primers at 200 g/L8. Product schedules follow predictable patterns.
What the human verifies: specific product designations against current manufacturer data sheets. Manufacturers change formulations. AI doesn't know that unless it's grounded against current data sheets. Verification is a cut-sheet check.
10 28 00— Toilet, Bath & Laundry Accessories
Off-the-shelf commodity products. Manufacturer cut sheets drive the content. Design judgment is minimal— the architect picks the line, the specifier captures the requirements.
ADA and accessibility callouts are the only place this section can move. Current ADA mounting heights and clearances need a sanity check. Beyond that, it's a content category AI handles cleanly.
07 92 00— Joint Sealants
Highly standardized. ASTM C920 governs. Sealant types map to predictable applications, and the three-part section format has been stable for decades.
The verification line is movement-class and chemistry for unusual conditions. Standard expansion joints? AI's first draft is usable. A complex curtain wall interface with non-standard movement? That's specifier judgment, not template.
An equally important question follows: which sections should AI never touch?
Where Drafting AI Should Never Go
Drafting AI should never produce specification content for structural, MEP performance, life-safety, or hidden-work sections. Anywhere a hallucinated number or fabricated product reference could disappear behind drywall, the answer is no.
The danger is most significant in work that becomes invisible once covered9. Once it's behind sheetrock, behind soffit, behind grade— the error compounds quietly until a failure surfaces it. Spec what gets hidden with the same rigor you'd want if you were the one signing the structural calcs.
Categories to keep human-only:
- Division 03 (Concrete) and Division 05 (Metals) — structural, engineer-of-record judgment
- Division 23 (HVAC) and Division 26 (Electrical) — performance sizing, code-critical
- Division 33 (Utilities) — public-safety adjacent
- Hidden-work categories generally — foundations, reinforcing steel, post-tensioning, fireproofing, critical MEP routing
- Performance specifications generally — they require design-intent calibration AI doesn't have
The reason these lines hold is structural. It's in how every spec is built.
The SectionFormat Lens— Why Part 1 Beats Part 2 and Part 3
AI safety in spec drafting tracks more closely with SectionFormat Part than with section number. Part 1 (General) is procedural and low-risk. Part 2 (Products) requires manufacturer accuracy. Part 3 (Execution) carries means-and-methods judgment.
CSI MasterFormat organizes specs into 50 Divisions; each Section follows the SectionFormat structure: Part 1 General, Part 2 Products, Part 3 Execution10. That three-part shape is the lens that explains why the five sections above work.
| Part | Content | AI Safety |
|---|---|---|
| Part 1 General | Submittals, references, quality assurance, scheduling | Highest — procedural |
| Part 2 Products | Manufacturer names, model numbers, performance data | Moderate — verify against cut sheets |
| Part 3 Execution | Installation methods, field tolerances, sequencing | Lowest — means-and-methods judgment |
Even in a low-risk section, AI should draft Part 1, support Part 2, and stay out of Part 3. That's the rule that survives across the whole MasterFormat structure.
The five sections work because they're either Part-1-heavy (01 33 00, 01 42 19) or have stable, commodity Part 2 and Part 3 content (09 91 23, 10 28 00, 07 92 00). Apply the same Part-by-Part logic to any section you're considering and you'll have your answer before the first prompt runs.
Knowing where to deploy AI is half the work. The other half is the guardrails.
Guardrails— How Low-Risk Stays Low-Risk
Four guardrails keep AI-assisted drafting actually low-risk in practice: human verification of every output, citation tracing for standards and products, office-master integration rather than ad-hoc drafting, and a documented carrier conversation about AI exclusions.
- Verify every output. AI hallucination rates run 1–27% across studies5. Specifier sign-off is non-negotiable. Trust, then verify. The math doesn't care how good the first draft looked.
- Trace citations. Every standard number, every product reference, every code citation gets checked against the primary source. AI fabricates standard numbers occasionally. A two-minute lookup catches what a fast read won't.
- Integrate with the office master. AI should be drafting into firm-vetted master templates, not from a blank prompt. Vendors of office-master spec systems (RIB Software) aim to start every project around 80% complete11. Stack AI on top of that foundation and the compounding works. Stack it on a blank prompt and it doesn't.
- Have the carrier conversation. Verisk's new AI exclusion forms for professional liability policies took effect January 1, 202612. Some E&O carriers are following. Verify your firm's policy before scaling AI use across projects. This isn't a future problem.
The design professional, not the AI vendor, retains professional liability for AI-introduced errors13. A contractor "who requests information from AI and relies on the answer may assume risk it would not otherwise have assumed had it requested the information from the architect"14.
This is what makes the hidden costs of unscoped AI projects so expensive in regulated practice. The cost isn't the tool license. It's the liability that doesn't migrate to the vendor when the tool gets a number wrong.
What This Says About Your Firm's AI Strategy
Starting with five bounded, low-risk sections isn't AI's destination for your firm. It's the disciplined entry point. Firms that scale AI well begin where they can measure both productivity gains and risk exposure— then expand the muscle.
The Dodge Construction Network / CMiC 2025 survey makes the gap concrete: 87% of contractors expect AI to transform the industry, while only 19% have adapted workflows for AI15. That's not an ambition problem. It's an execution discipline problem. And execution discipline is what the five-section approach builds.
Three things compound when a firm gets this right:
- A working library of vetted AI prompts tied to office-master templates
- A review workflow specifiers actually trust (because they wrote it)
- A documented carrier and liability posture leadership can defend
"The firms that will thrive won't be those that adopt AI the fastest, but those that integrate it the most thoughtfully."16
That AIA framing maps to what we see across firm-leadership work generally. AI in spec drafting works the same way it works everywhere else: it amplifies firms that already have disciplined systems and exposes the ones that don't. Specifier domain expertise plus AI is leverage. AI without that expertise is noise. The same logic shapes a decision framework founders can apply across the firm — and it shapes building an AI-ready firm culture.
A few questions firms ask before piloting this approach.
FAQ— Drafting AI in Practice
What is "drafting AI" in architecture?
In the AEC industry, drafting AI most commonly refers to AI-assisted specification drafting— using AI to produce first-draft technical specifications that human specifiers then review and finalize. It's distinct from CAD line-drafting AI. The applicable use cases are bounded by section type and SectionFormat Part.
Which spec sections are safe for AI today?
The lowest-risk sections are heavily procedural and standards-driven: 01 33 00 Submittal Procedures, 01 42 19 Reference Standards, 09 91 23 Interior Painting, 10 28 00 Toilet/Bath/Laundry Accessories, and 07 92 00 Joint Sealants. Each pairs codified standards with low design judgment and fast verification.
Which spec sections should not use AI?
Structural specs (Divisions 03 Concrete and 05 Metals), MEP performance specs (Divisions 23 HVAC and 26 Electrical), and any sections governing hidden or safety-critical work. The danger is greatest in work that becomes invisible once covered— foundations, rebar, fireproofing, and critical MEP routing9.
Does my E&O insurance cover AI-drafted specs?
Possibly yes today, but AI exclusions are appearing. Verisk's new AI exclusion forms for professional liability took effect January 1, 202612, and other carriers are following. Verify with your carrier before scaling AI use across projects. The design professional retains liability regardless of whether AI was involved in drafting13.
Should I use ChatGPT or Claude, or a purpose-built spec-drafting tool?
For Part 1 procedural content (submittals, reference standards, scheduling), general LLMs like ChatGPT or Claude work well with proper office-master grounding. For Part 2 Products content— where manufacturer accuracy matters most— purpose-built tools fine-tuned on construction data (Avitru, Conspectus Cloud, Deltek Specpoint, SpecLink.AI) carry less hallucination risk. The right answer depends on which Part of the section you're drafting.
The Starter Set in Practice
Drafting AI works today on the spec sections where the math is in your favor: heavy procedure, codified standards, low design judgment, fast verification. Five sections meet that test now. The rest of the spec stays in human hands— for now, and for good reason.
The list grows as tools mature and office masters get smarter. Today's task is to build the muscle in the bounded zone where productivity and risk can both be measured. Drafting AI is a starter set, not the destination. The firms that win build the muscle here first.
If you're trying to map where AI fits across your firm's workflow— not just specs— that's the work we do at Dan Cumberland Labs. Pick one section from the five. Run it through your office master. Measure. Then expand the muscle. Deciding where AI belongs in a firm's practice is its own discipline. Worth doing carefully.
References
- Conspectus Inc. (Ricke, Chris & Melody Fontenot), "AI(A) Interview — Part 1, Spec Software" (2023) — https://www.conspectusinc.com/blog/ai-spec-software-interview
- American Institute of Architects (AIA) / Deltek, "Shaping the Future: Six Key Benefits of AI in Specifications" (2024) — https://www.aia.org/resource-center/shaping-future-six-key-benefits-ai-specifications
- American Institute of Architects (AIA), "AI in Practice: Strategic Insights from The Architect's Journey to Specification" (2025) — https://www.aia.org/resource-center/ai-practice-strategic-insights-architects-journey-specification
- Stoel Rives LLP (Brown, Evan A.), "Bricks and Bots: AI Technologies' Growing Impact on Construction" (2025) — https://www.stoel.com/insights/publications/bricks-and-bots-ai-technologies-growing-impact-on-construction
- Fabyanske, Westra, Hart & Thomson (Orman & Radaj), "Legal Risks of the Use of AI in the Design-Build Process" (2025) — https://www.fwhtlaw.com/blog/2025/05/16/legal-risks-of-the-use-of-ai-in-the-design-build-process/
- Whole Building Design Guide / DOD UFGS, "UFGS 01 33 00 Submittal Procedures" — https://www.wbdg.org/FFC/DOD/UFGS/UFGS%2001%2033%2000.pdf
- Conspectus Inc., "Division 01: It's Not Boilerplate!" (2012) — https://www.conspectusinc.com/blog/2012/07/division_01_-_its_not_boilerplate_1
- Behr Pro, "Section 09 91 23 Interior Painting Master Specification" (2020) — https://www.behr.com/binaries/content/assets/behrdotcomrefresh/pro/specifications/behrpro-09-91-23-int-master-specification-06-01-20.pdf
- Construction Owners (Roozbahani, Max Mahdi), "AI Chatbot Errors Could Create Long-Term Risks for Construction Projects" (2026) — https://www.constructionowners.com/news/ai-hallucinations-risk-construction-decisions
- Construction Specifications Institute (CSI), "MasterFormat" — https://www.csiresources.org/standards/masterformat
- RIB Software, "Office Masters & Smart Linking in Specification Management" — https://www.rib-software.com/en/blogs/spec-management-office-masters
- SMB Securenow, "AI Hallucination Liability: The 2025 Risk & Compliance Guide" (2025) — https://smbsecurenow.com/ai-hallucination-liability-guide/
- Fabyanske, Westra, Hart & Thomson (citing ASCE), "Legal Risks of the Use of AI in the Design-Build Process" (2025) — https://www.fwhtlaw.com/blog/2025/05/16/legal-risks-of-the-use-of-ai-in-the-design-build-process/
- Stoel Rives LLP (Brown, Evan A.), "Bricks and Bots: AI Technologies' Growing Impact on Construction" (2025) — https://www.stoel.com/insights/publications/bricks-and-bots-ai-technologies-growing-impact-on-construction
- Construction Dive (reporting on Dodge / CMiC 2025 survey), "AI nears 'tipping point' in construction as contractors pilot tech: survey" (2025) — https://www.constructiondive.com/news/builders-ai-transform-businesses-survey/807555/
- American Institute of Architects (AIA) / Deltek, "AI in Practice: Strategic Insights from The Architect's Journey to Specification" (2025) — https://www.aia.org/resource-center/ai-practice-strategic-insights-architects-journey-specification
