Operations

The Spec Book Nobody Finished Reading

By John J. BakerNovember 19, 20269 min read


A submittal register should start from the spec book, and on a commercial job that book can run past 1,000 pages. AI can now read the whole spec and draft the register in minutes, so your project manager checks a list instead of building one by hand. Getting complete submittals in early matters because the general contractor's review queue is where the time goes.

Here is how it usually goes. An electrical sub wins a job on a new medical office building. The project manager gets the spec book, flips to Division 26, and starts a spreadsheet: panelboards, switchboards, wiring devices, lighting fixtures, fire alarm if it is in your scope. Each section asks for product data, shop drawings, samples, certificates, or some mix. The PM types each one in, guesses at a due date, and moves on to the next job that is already behind.

Somewhere in a Division 01 section or a Division 23 section about controls wiring, there is a line the PM did not read. It asks for a coordination study or a seismic certificate from your scope. Nobody finds it until the GC asks where it is, or until the gear is about to ship and the package was never approved.

Nobody was careless. The book is simply longer than anyone has time to read twice.

What is a submittal register, and why does it start with the spec book?

A submittal register is the list of every document the architect and engineer want to see before you buy or install something, with a due date for each. The spec book is where those requirements live, section by section, usually in a part of each section titled "Submittals" or "Action Submittals."

Each line on the register needs the same few things: the spec section number, the item, the type of submittal, the date it is due to the GC, and how long the material takes to arrive once it is approved.

One line of a submittal register

Every entry needs the same five things. Miss one and the date slips.

  1. Spec section

    26 24 16

    Where the requirement lives in the spec book.

  2. Item

    Panelboards

    What the architect and engineer want to see.

  3. Submittal type

    Product data and shop drawings

    What the package must contain.

  4. Due to the GC

    Week 3 after award

    Set from the schedule, not a guess.

  5. Lead time after approval

    Ask your supplier

    Drives when the order must go in.

Illustrative example of a typical register line for an electrical subcontractor, not data from a specific project. Section numbers follow CSI MasterFormat.

If a line is missing from the register, it is missing from your schedule.

Where do submittals go wrong for an electrical or mechanical sub?

They go wrong in three predictable places: a requirement that never made it onto the register, a package that goes in incomplete and comes back, and a long-lead item that gets ordered late because its approval was late.

The first one is a reading problem. The requirements that affect your scope are not always in your division. A mechanical sub may find a commissioning requirement in Division 01.

The second one is a checking problem. A package for a rooftop unit goes in with the cut sheet but without the sound data the spec asked for, and it comes back marked "revise and resubmit." That is one more trip through the line.

The third one is a timing problem, and it is the expensive one. Switchgear, transformers, generators, and large air handlers can carry long lead times. If the submittal is approved late, the order goes in late, and the install date slides no matter how hard the crew works.

What can AI actually pull out of a 1,000-page spec today?

It can pull out most of the submittal requirements, organized by section, in minutes, with some errors a person still needs to catch. The best recent evidence comes from a research group at the University of North Florida that built and tested a system for exactly this job.

In a paper published in June 2026 in the FLAIRS-39 conference proceedings, the group described a hybrid system. Plain rules split the spec into its standard CSI MasterFormat sections, and a large language model is used only for narrow extraction inside each section. On 20 real specifications, it scored 94.3 percent on F1, a standard accuracy measure that balances what it found against what it missed. The authors report 94 percent less time and 93 percent less cost than doing it by hand, and say the system has been in use with one contractor for six months.

An earlier paper from the same group, published in Frontiers in Built Environment in January 2026, describes specs of 1,000 to 1,500 pages, processing in about 6 to 8 minutes per document, and computing cost under 10 cents per file. The authors also write that "some misclassifications or redundant entries required expert review," and that these tools "augment rather than replace professional judgment."

Read both carefully. Twenty specifications is a small sample. The researchers built and tested their own system. The work was supported by a collaboration with a Florida civil contractor that employed one of the authors. That does not make it wrong, but it is not an independent test, and a 94 percent score still means some lines are missed or wrong.

The design choice matters more than the score. The AI does not decide what the spec means. Rules find the sections, the model pulls out the items, and a person reviews the list. That is the same split I described in what AI actually does: it reads and drafts, and a person decides.

Why does getting submittals in early matter when the GC's queue is the bottleneck?

Because a submittal that arrives when the GC's reviewers are buried waits longer, no matter how good it is. A case study presented in December 2025 at the Project Production Institute symposium by researchers at Stanford's Center for Integrated Facility Engineering and Mortenson makes this concrete.

On the $1.3 billion Gaylord Pacific hotel project, which had 4,383 submittals, the team analyzed 913 of them. Subcontractors submitted 69 percent on time. But 440 of the 913 entered the GC's review queue when more than 100 were already waiting. At the peak, 247 were open at once. The team found the queue ran best at around 30 open submittals, where cycle time would have been about 20 days. Actual cycle time was about 35.

Open submittals waiting in the GC’s review queue

Gaylord Pacific hotel, $1.3 billion, 4,383 submittals. 913 analyzed.

  • Healthy workloadCycle time about 20 days

    About 30 open

  • Already waiting when 440 of 913 arrivedNearly half the analyzed submittals

    More than 100 open

  • PeakActual cycle time about 35 days

    247 open

Subcontractors submitted 69 percent on time. The wait came from when packages arrived, not only from whether they were late.

Source: Bochen Zhang (Stanford CIFE) and Brian Nahas (Mortenson), Applying Operations Science and Production Management to Improve Efficiencies and Inform AI Application in Submittal Process, Project Production Institute 12th Annual Symposium, December 2025. Single project case study. The 20 day figure is the team’s estimate at a healthy workload.

This is one project, presented by the GC and its research partner, so treat it as one well-documented example, not a rule. But anyone who has waited on a submittal knows the pattern: packages pile up after award, and the incomplete ones come back and join the pile again.

A sub cannot control the GC's queue. A sub can control two things: getting a complete register built in the first week instead of the fourth, and making sure each package is complete the first time so it only goes through the line once.

What should stay with your project manager?

The PM owns the register, the judgment calls, and every package that goes out under your company's name. AI drafts the list. The PM decides what is real.

  1. 1. AI drafts

    Read the spec book

    Split it into sections and pull every submittal requirement.

  2. 2. AI drafts

    Draft the register

    One line per item, with section, type and a proposed date.

  3. 3. PM decides

    Check scope and dates

    Cut what belongs to other trades, add real lead times, flag substitutions.

  4. 4. PM decides

    Send complete packages

    Each package checked against its section before it goes to the GC.

  5. 5. GC and architect

    Review and approve

    Approved packages release the order. Rejected ones go back in line.

AI drafts the register. It never sends a package. The split follows the hybrid design in the University of North Florida Submittals Agent paper (FLAIRS-39, June 2026), where rules find the sections, a language model extracts within them, and people keep contractual interpretation.

In practice, that means a few specific calls stay with a person:

  • Is this requirement actually in our scope? The AI will find every submittal in the book. Your PM knows which ones belong to the mechanical sub, the GC, or you.
  • Is this a substitution? If you plan to propose an equal product, that is a decision, and it needs the PM and often the estimator.
  • What is the real lead time? The spec does not know what your supplier told you last week. The PM does.
  • When do we push? Chasing a late review is a relationship call with the GC.

The AI can also miss things. It can misread a scanned page, skip a requirement written in an unusual place, or list the same item twice. That is why it flags and drafts, and never sends anything. I wrote about why that line matters in the draft that broke trust: anything that goes to a GC under your name gets a person's eyes first.

Why does this matter now?

Because the work AI does well right now is exactly this kind of reading. A Census Bureau working paper published in April 2026, based on its nationally representative business survey, found that writing, document analysis, and information search are the leading generative AI tasks at the firms that use it, and that 66 percent of users rely on AI only to help with work, not to replace it.

Construction is still early. The Federal Reserve Bank of Minneapolis reported in May 2026, from the same Census survey, that fewer than 10 percent of businesses in construction and a few other labor-intensive sectors use AI. A sub that builds a clean register in the first week of a job is ahead of most of the people it works beside.

It is the same idea as the permit packet that keeps coming back, with the architect and GC in place of the city: read the whole requirement list against your paperwork before it goes in, and save a trip through the line.

The honest limits

The research here is early and small. The accuracy figures come from 20 specifications, tested by the people who built the system. The queue figures come from one large hotel project. Neither tells you exactly what will happen on your next job.

AI will not catch a requirement that is only on the drawings, in an addendum it never saw, or in a project manual section that was scanned as an image. It does not know your scope split with the other trades. It does not know your supplier's real lead times. And a faster register does not fix a GC that is overloaded. It just means your packages are complete and in line early.

How would a sub try this on one bid or one job?

Pick one job that was just awarded, or one bid where you want to understand the submittal load before you price it. Run the spec through an AI extraction, and have your PM build the register by hand the way they always do. Then compare the two lists line by line.

You will learn three things fast: what the AI found that your PM would have missed, what the AI got wrong, and how long each version took. That comparison tells you whether this is worth building into how every job starts.

Where ClearOak comes in

ClearOak & Co. is a software and operations partner for trades businesses. We start with how your project managers build registers today, not with software. I would sit down with your PMs, look at the last few jobs where a submittal went in late or came back, and find where the misses actually came from.

From there, we bring in the right builders to set up the spec read for the tools you already use, run it next to your PMs on a few jobs, and train your team to own it. What you get back is your PMs' time, and long-lead orders placed sooner, which protects your margins on the job.

If your PMs are still building registers by hand from the spec book, schedule a call and we will look at one of your recent jobs together.

Prefer email? Reach me at john@clearoakconsulting.com and tell me which submittal came back on your last job.

Sources

Have a workflow that is costing you?

Book a free call. We will tell you honestly whether we can help and what that looks like.