How Can AI Help Construction Project Managers? Submittals and Sub Chasing

On a small job AI can draft and track paperwork: build a submittal log from the specification, write subcontractor chase emails, and confirm deliveries. It cannot approve a submittal, judge site safety, serve contract notices or agree money. Keep AI on drafting and tracking, and keep approvals, safety, payment and anything with legal effect with a named person.

Two habits decide whether this works. The first is arithmetic: chat assistants are unreliable at date maths, so the log lives in a spreadsheet and the AI writes formulas, not dates. The second is the source: the specification is the contract, and the AI's reading of it is a draft to check. Both show up in the worked log below, which uses a small fit-out job.

Follow me on Instagram@sagnikteaches

What stays with the project manager

Settle the boundary before the first prompt. The table splits seven everyday tasks into what a chat assistant can do and what a person keeps.

Connect on LinkedInSagnik Bhattacharya

The pattern is the same in every row: AI handles the typing and the tracking, and a person owns the facts, the decisions and the signature. If a task is not in the table, ask who would be answerable if the output were wrong. That person reads it first.

Subscribe on YouTube@sagnikteaches

Build the submittal log from the spec, then fix the dates yourself

A submittal is a document, drawing or sample the contractor sends to the design team for review before ordering or installing something: product data, shop drawings, colour samples, test reports. The specification lists what each section needs. A log tracks each item from request to approval, and its real job is to show when each item must go out so that review and delivery fit inside the programme.

Paste one specification section at a time, never the whole book. Long documents invite skipped paragraphs, and a single section is a unit you can check by eye.

Extract every submittal requirement from the specification section below.
For each one give: item; type (product data, shop drawing, sample, test report,
warranty, other); whether it is for action or information if the text says;
and the exact sentence it came from. If the section has no submittal
requirement, say so. Do not infer anything the text does not state.
Return a table I can paste into a spreadsheet.

[paste one specification section here]

The model returns rows. You then add the columns it cannot know, and there are four numbers per item, all yours: the week the item must be on site, the supplier's lead time, the design team's review period from the contract, and a buffer for one rejected first submission. Two plain formulas do the rest. Order-by is on-site week minus lead time. Submit-by is order-by minus review period minus buffer. Ask the assistant to write those formulas for your column letters, then let the spreadsheet do the sums. Lead times come from supplier quotes, so never ask the model for them: it will supply a plausible figure with nothing behind it.

Here is an illustrative log for a tenant fit-out. Today is week 3, the review period is two weeks (three for lighting, air and glazing, which several consultants check) and the buffer is one week. Use the review period in your own contract, not these. The rows are sorted by submit-by.

Check one row by hand, as you should every time. Light fittings: on site in week 14, ten weeks of lead time, so the order must be placed in week 4. Approval is needed by then, and review takes three weeks, so the package should reach the design team in week 1. Take off one more week for a possible rejection and the target is week 0. At week 3 it is already past.

That top line is the value of the log. Two items are late and a third is due this week, and nobody sees that from a pile of emails. The responses are decisions for the project manager and the design team: ask for a faster review of the lighting package, ask whether a stocked alternative with a shorter lead time is acceptable (that is the design team's call), or change the install sequence. The bottom of the list matters just as much. Paint and carpet can wait, so you stop chasing the wrong things first.

Check the log against the spec before you rely on it

An extraction that looks tidy can still be missing a paragraph. Three checks take about ten minutes per section and catch many gaps.

1. Count the keywords. Search the section for "submit", "sample", "mock-up", "test report", "warranty" and "certificate". Every hit should land on a row or have a reason it does not.
2. Run it twice. If two runs return different numbers of rows, the text is ambiguous or too long. Read that section yourself.
3. Build one section by hand. Pick three sections across different trades, list the submittals yourself, and compare. The target is zero missing.

A realistic miss shows why. For a door hardware section, a model returned one row, product data. The section also asked for one sample of each lever finish and a keying schedule, in a separate paragraph. The keyword search found five hits against one row, so the gap surfaced in two minutes. The fix was to rerun with an added instruction: "List every sentence containing 'submit' or 'sample' first, then build the rows from those sentences." Cross-references cause the other common miss. A section that says "submit as required in the general requirements" points to a document the model has not seen, so read for those references yourself.

The log tracks the work, and the specification governs it. When the two disagree, the specification wins, and the log gets corrected.

Packages out and comments back: a checklist and an action list

The log tells you when to send a package. Two more prompts help you send a complete one and deal with what comes back, and neither replaces the review that matters.

Before a package goes to the design team, build a completeness checklist from the same specification text and tick it against the package. Ask: "From the text below, list everything the door hardware submittal must contain as short checklist items, quoting the sentence each came from." A typical result lists product data for each hardware type, one sample of each finish, a keying schedule, a finish schedule that matches the drawings, and the manufacturer's installation instructions. The checklist catches the missing keying schedule, which is a common reason for a first rejection. It cannot tell you whether the product complies, so you still read the data against the requirement, for example that the stated rating matches the one the specification asks for.

When the package comes back, the comments arrive as scribbles on a PDF or a paragraph in an email. Paste them in and ask for an action list: "Turn these reviewer comments into actions for the subcontractor, one per line, with who must act and what evidence closes it. Do not interpret the comments. If one is unclear, write UNCLEAR." With three illustrative comments (hinges to be ball-bearing type, finish sample 2 not matching the schedule code, keying to show a grand master system), the output looks like this:

Read each line against the reviewer's own words, and take any line marked UNCLEAR to the design team before the subcontractor spends a week guessing. Then update the log: the returned date, the resubmittal due date, and whether the buffer you allowed has now been used up.

One boundary matters here more than any other. A status such as "approved as noted" has a meaning defined by the contract, and only the design team's stamp counts. The assistant can summarise comments, but it cannot decide that the subcontractor may now order. Confirm the status in writing before a purchase order goes out.

Chasing a late submittal without damaging the relationship

A late submittal rarely means a subcontractor is ignoring you. More often the supplier has not sent the data sheet, or the sub is waiting on a colour decision, or the email reached the wrong person. A good chase finds the blocker and states what the delay protects, in programme terms. The log row holds both facts, so the AI has something specific to write from, and the message stops sounding like a nag.

Agree a ladder, timed from the submit-by date in the log:

1. One week before: a heads-up listing the items due and exactly what you need for each.
2. On the date: a short reminder.
3. Three working days late: a direct request for a send date, with the programme effect, for example that the order window for the item closes in week 13.
4. Five working days late: a phone call, followed by a two-line email recording what was agreed.
5. Anything about delay claims, notices or withheld payment: a person decides, with the contract open.

Write a chase message to [first name] at [subcontractor] about this log row.
Use only the facts in the row. Say what we need, by when, and what it
protects in the programme. No contract clauses, no blame, no threats.
Give me an email (under 80 words) and a text message (under 35 words).

Item: Door hardware, product data and one sample of each lever finish
Submit by: week 10 | Order window closes: week 13 | Lead time: 6 weeks
Status: not received | Last contact: reminder sent Monday

An illustrative email from a good run: "Hi [first name], the door hardware package is due with us in week 10 so the design team can review it and keep your order inside the six-week lead time. We need the product data and one sample of each lever finish. Could you confirm a send date today? If a supplier is slow, tell me which one and I will chase from our side." The text version: "Hi [first name], door hardware submittal is due this week. Can you confirm when the data and samples will reach us? Thanks."

Set that against the usual end-of-shift version: "Hi, where are we with the door hardware stuff? Need it asap". It names no item, gives no date and offers no reason, and a vague chase tends to get an equally vague "on it" back.

Check the draft for two habits that creep in. Models like to add contract language, such as "as required under clause 14" (an invented clause) or "failure to submit may result in delay claims" (a threat with legal weight). Cut both. Some drafts also promise things on your behalf, such as "we will extend your programme", which is not the model's call and may not be yours either. Ask for a firmer variant only after the second chase, and read it as if it were going on the file, because one day it might be.

Delivery confirmations: the message, the day-before check, the docket

A missed delivery costs a crew its morning, and many misses come from details nobody wrote down: the time window, the offloading method, who signs. The cure is two short messages and a check at the gate. The assistant drafts the messages from the order line; you own the check.

The first message goes out when the order is placed. It carries the order reference, the item and quantity, the delivery date and window, the site contact, and how the load will be taken off the truck. It asks the supplier to reply to confirm the slot. A template you can paste under the prompt "turn this order line into a delivery confirmation under 70 words":

Order ref [PO number]: 40 sheets of gypsum board, standard thickness.
Delivery: Thursday, 7:00 to 9:00 on site. Site contact: [name, phone].
Offloading: [forklift / by hand / crane]. Please reply to confirm the slot
and send the driver's number the evening before.

The second message goes the evening before: "Confirming tomorrow's 7:00 to 9:00 delivery for order [PO number]. Please send the driver's number and tell me if the slot has moved." A third, firmer version for a supplier with a poor record can say what you will do if the window is missed, such as moving the delivery to the next day, but that wording is a decision, so write the consequence yourself and let the assistant polish the sentence.

The docket check is where the assistant earns its place at the gate. Photograph the delivery docket, paste your purchase order lines under it, and ask for a line-by-line comparison that lists every mismatch in quantity, product code or description. On an imagined job, the order says 40 sheets with code GB-STD, and the docket says 36 sheets with code GB-MR. The assistant reports a shortage of four sheets and a different product code, which on this imaginary job means moisture-resistant board where standard board was ordered.

Then the human part. You count the sheets on the pallet, look at the labels, write the shortage on the docket before the driver leaves, and photograph the load. The assistant can read paper but cannot see a pallet, so sign for what you counted, never for what the docket says. Photographed text also gets misread occasionally, a 6 for an 8 for instance, so check each figure it reports against the original by eye. For the wider question of reading dockets into stock records, see whether AI can read delivery notes and update stock.

A small job's week, before and after

Here is a quick sum for an illustrative project manager running two small jobs. The counts and minutes are guesses to be replaced by your own timings, and the "with AI" column includes the time to check every output.

That is about 1 hour 45 minutes a week, so a modest saving, and the docket check saves the least because the physical count cannot be shortened. Two things matter more than the minutes. A tracked log stops items going late unnoticed, as the week 3 table showed, and consistent chasers mean no sub is chased harder than another by accident. An individual ChatGPT or Claude plan lists at $20 a month, or a business seat at $25 a user a month, so the time saved repays the tool within the first week at ordinary labour rates. Time your own tasks for two weeks first, because your numbers will differ from these.

Spreadsheet first, platform later: where the log should live

Start with a spreadsheet. It costs nothing, everyone can open it, and an assistant writes the formulas in a minute. Its weaknesses are real: no audit trail, two people can overwrite each other, and it will not send the reminders. Construction software fixes some of that for a price, and the price depends on which plan holds the feature you need.

Fieldwire is a useful example because its pricing page prints the split. Its Basic plan is free, Pro lists at $39 per user a month billed annually, and Business at $64. Submittals, RFIs, change orders and budget appear only on Business Plus, at $89 per user a month billed annually. If a submittal register is what you are buying, check which plan carries it on any product you consider, and ask what its own AI features do with submittals before assuming they do anything. Whichever tool you choose, the prompts above still work, because they produce the rows you import.

Whatever holds the log, paste only what the job needs into the chat. A specification section is often fine; a full bid pack with prices and personal contacts is not. Use a business plan that does not train on your data by default, and keep a note of which tool saw which documents.

Where RFIs and look-aheads fit

Requests for information to the design team, and the look-ahead programme that shows the next few weeks trade by trade, are separate jobs with their own checks. Rather than squeeze them in here, read AI tools for construction project management: RFIs, logs and reports and weekly early-warning flags for cost and delay.

The three jobs on this page feed each other in a chain. A late submittal makes a late order, a late order makes a late delivery, and a late delivery wrecks the next look-ahead. That is why they belong together, and why a single tracker with a status column per step beats three separate lists. For the scheduling of the subcontractors themselves, see AI scheduling for small builders. For the paperwork that must be in place before a sub starts, such as insurance and licences, the subcontractor paperwork checklist is the companion piece. And if you run the whole job as the general contractor, ten jobs ranked by effort shows where this work sits among the rest.

One last rule closes the loop. Every message the assistant drafts goes out under your name, so read it as yours. If you would not say a sentence to the subcontractor's face, delete it before it is sent.

Questions project managers ask about this

Can AI approve or review a submittal for me?

It can check a package for completeness against a checklist built from the specification, which catches missing items. Approval belongs to the design team the contract names, and your own review before sending stays yours. Use AI to find gaps, never to decide that a product complies.

Which AI tool is best for a submittal log?

Any current assistant that accepts pasted text or PDFs can extract the rows, with a spreadsheet holding the log. Choose on data terms and on what you already pay for. Check whether your construction software has a submittal register before building one by hand.

Why not let the AI calculate the dates?

Chat assistants make arithmetic and calendar mistakes that look confident, such as miscounting weeks across a holiday. Ask for spreadsheet formulas instead, put your own lead times and review periods in as numbers, and let the sheet do the maths. Check one row by hand.

Is it all right to paste specification sections into a chat?

Often yes, but read your contract's confidentiality terms and the client's rules first, paste only the sections you need, strip names where you can, and use a business plan that does not train on your data by default rather than a personal account.

Further reads

• AI for Building Contractors: Chasing Subcontractors and Progress Updates — Chasing subcontractors on progress as well as paperwork.
• Subcontractor Management: Track Insurance Expiry and Onboarding With AI — Track insurance expiry before a sub starts.
• How to Use AI in a Construction Company: Estimator, PM and Office — How AI fits the estimator, PM and office roles.
• AI in the Construction Field: Photos, Voice Notes and Punch Lists — Photos and voice notes into punch lists.
• AI Tools for Home Builders: Build Schedules, Delays and Sub Chasing — Schedules, delays and sub chasing for home builders.
• AI Tools for Shop Drawings: Checking Fabrication Drawings Against the Spec — How a steel or mechanical subcontractor can use ChatGPT, Claude and Bluebeam to compare drawings with the spec, and the checks that must stay human.
• AI Software for Construction Scheduling: Programme Updates After Delays — One small job walked through a late delivery and two weather days: what the assistant does, what you calculate yourself, and what you tell each sub.
• ChatGPT for Electrical Contractors: Bid Clarifications and RFIs — Prompts and sample outputs for clarification lists, RFI wording and submittal letters on commercial jobs, with confidential-data rules.
• AI in Steel Detailing: Gains for a Small Office, and Tasks to Avoid — Four admin jobs a small steel detailing office can give to AI, the tasks that stay with detailers and checkers, and a pre-send checklist.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Fieldwire pricing page; ChatGPT and Claude plan pages for list prices; OpenAI enterprise privacy page and vendor privacy statements for business plans.