How to Use AI for Specification Writing and Track What It Changed

Specification writing is one of the most practical places for AI at an architecture or engineering firm, and one of the riskiest. A tool that tidies an office master can also change a fire rating, swap "shall" for "may" or add a brand name to a public bid, and the architect or engineer who stamps the set owns the result. This guide covers where AI helps, what it gets wrong and a simple redline process that keeps responsible control on the record.

Can architects and engineers use AI to write specifications?

Yes. California's licensing laws do not prohibit AI tools, and they put responsibility for the specification on the licensed professional who signs it. Under Business and Professions Code section 5536.1, whoever prepares or is in responsible control of plans, specifications and instruments of service for others must sign them, and a licensed architect must also stamp them as evidence of responsibility (employees working for a licensee are exempt from signing). For civil engineering work, section 6735 requires specifications to be prepared by, or under the responsible charge of, a licensed civil engineer, and specifications released for construction must bear the engineer's signature and stamp.

Section 5535.1 defines responsible control as the control over content during preparation that architects ordinarily exercise under the professional standard of care. Section 6703 defines responsible charge as independent control and direction using the engineer's own judgment. Neither statute mentions AI. Both assume the licensed professional knows what is in the document, and that is the practical problem with AI in specifications: it can change wording in ways that are easy to miss.

Where does AI help in architectural specification writing?

AI helps most with the repetitive editing in specification writing, while the specifier still makes every decision. Common uses:

• Editing office master sections to fit a project, such as removing products or systems the project does not use.
• Checking that section references, related sections and submittal requirements line up.
• Comparing specification text against drawing notes or schedules to flag conflicts.
• Drafting first-pass text for a new section from the firm's prior project specifications.
• Summarizing manufacturer data for the specifier's review.

Each of these saves time. Each also involves the tool rewriting text that someone will later sign.

What kinds of changes does AI make that are easy to miss?

AI can change numbers, mandatory language, references and product names in ways that look correct on a quick read. NIST's Generative AI Profile (NIST AI 600-1) describes confabulation, where AI systems "generate and confidently present erroneous or false content." In a specification, that can look like:

• A changed number: a thickness, a tolerance, a fire rating or the edition of a test standard.
• A changed verb: "may" where the master said "shall," or a dropped "not."
• An invented reference: a standard designation or code section that does not exist or does not say what the text claims.
• A product name inserted where the master used a performance requirement.
• A paragraph deleted because the tool judged it redundant.

Product names matter on public work. California Public Contract Code section 3400 generally bars public works specifications from limiting bidding to one concern or naming a brand or trade name without the words "or equal," with limited exceptions such as matching products already in use or items available from only one source. An AI tool that has seen many specifications may add a product name the firm never intended.

How do you track what AI changed in a specification?

Treat each AI pass as a separate revision and compare it against the last human-approved version before anyone accepts it. The goal is a record of what changed, who reviewed it and who accepted it.

A workable process:

• Save a baseline. Before AI touches a section, save the current approved version under a clear version name.
• Run AI on a copy. The tool edits a working copy, never the baseline.
• Run a document compare. Use the compare feature in Word or in your specification software to produce a redline of every change between the baseline and the AI output.
• Review every change. The specifier accepts or rejects each one. Changes to numbers, standards, product names and mandatory language get a second check against the source.
• Record acceptance. Note the reviewer, date and tool in the section's revision history or a simple log in the project folder.
• Verify references. Open each cited standard or code section that AI added or changed.

This creates the evidence of responsible control a licensing board, client or opposing expert might ask for. It also makes late-stage coordination easier when a section changes close to issue.

Who owns an AI-drafted specification?

A firm can claim copyright only in the parts of a specification that people wrote, selected, arranged or modified. The U.S. Copyright Office's 2025 report on copyrightability concludes that prompts alone do not give a user enough control to be the author of AI output, while human-authored expression and creative selection, arrangement or modification of AI output can be protected. A specification built from the firm's own master, edited and accepted by its staff, carries more human authorship than one generated from a prompt, and the saved redlines document that contribution. Ownership terms for instruments of service in the client agreement still govern, so review those with counsel.

How do you keep project data out of the wrong AI tools?

Use one firm-approved AI tool under a business agreement, and block uploads to everything else. Specifications often include security details, owner standards and, on defense work, controlled unclassified information. Pasting them into a personal AI account sends that data outside the firm's control. Our guide to using AI tools without exposing client drawings covers how to set up that approved tool.

The CIS Critical Security Controls v8.1 give a simple structure: a data inventory and classification scheme (Safeguards 3.1, 3.2 and 3.7), need-to-know access (3.3), data loss prevention (3.13) and URL filtering for unapproved sites (9.3). For firms on defense projects, our article on keeping CUI out of AI tools covers the additional rules.

Remote access matters too. Specifiers often work from home or a jobsite trailer. Zero trust network access lets them reach project files without opening the office network, and lets the firm apply the same AI rules wherever they work.

What should a firm do this quarter?

Start with six steps a principal can assign this quarter:

• Name the approved AI tool for specification work and block personal AI accounts on firm devices.
• Add a short AI section to the firm's specification procedures: baseline, working copy, compare, review, record.
• Require a document compare for every AI-assisted section before it goes into a set.
• Add a check for brand names without "or equal" to the review list for public projects.
• Train specifiers to open every standard and code reference AI adds or changes.
• Update the firm AI policy. Our guide to writing an AI policy for architecture and engineering firms is a starting point.

How can s90 help?

s90 sets up the approved AI workspace, data loss prevention rules that keep project files out of unapproved tools, and zero trust access for specifiers working outside the office. If you want help building AI into your specification process with those controls in place, we are happy to talk.