AI in Construction Projects: Weekly Early-Warning Flags for Cost and Delay

A project manager can paste the week's progress notes, last week's flags and a short list of thresholds into an AI assistant and ask it to return only the items that cross a threshold, each with a quoted line from the notes as evidence. The manager then checks every flag before it reaches the risk register.

The routine catches slow leaks. A slip on a pour or a cost line can sit under any alarm one week at a time, yet three small slips add up. In the example below a deck pour moved by 2, 2 and then 1 working days, and only the float left on the activity tripped a flag. One limit shapes the whole routine: the AI sees only what the notes say, so silence reads as good news unless you also ask what went missing.

Follow me on Instagram@sagnikteaches

How five days of slip hid inside three quiet weeks

Everything here is illustrative: a three-storey block of 24 flats, a $6.8 million contract and a 52-week programme, now in week 17 (about a third of the way through). The level 3 deck pour was baselined for Monday of week 19 with 6 working days of float, which is the time the pour can slip before it moves the end date. Here is how its forecast drifted in the weekly notes.

Connect on LinkedInSagnik Bhattacharya

A manager reading three separate weekly notes sees a pour that is "probably a couple of days later" each time. An alarm on weekly movement never fires, because no single week moved by 3 days. A routine that carries last week's forecast forward and tracks the float catches it at week 16, with two days still in hand. That is the whole case for doing this weekly and in writing, and it is why the first job is deciding the thresholds, before any note goes into a prompt. The tutorial on programme updates after delays covers rebuilding the programme once a flag has turned into a real delay; this one stops at spotting it.

Subscribe on YouTube@sagnikteaches

Thresholds to set before you paste anything

Thresholds are the manager's, not the AI's, and they should hang on something concrete: float, contingency, contract notice periods and lead times. The ones below are a starting point, to be replaced with your own after a look at the contract and the programme.

Two notes on choosing. First, use "whichever is hit first" for cost, because a percentage alone lets a big line drift by tens of thousands, and a dollar figure alone means nothing on a small package. Second, keep the list short. Eight rules the manager can recite beat thirty that the AI applies unevenly. Write them with numbers, and store them with the watch list in a Project in ChatGPT or Claude so you paste them once.

The inputs: site notes, a watch list and last week's flags

The AI gets three things each week. The first is the site manager's notes, the second is a watch list you keep, and the third is last week's flag list, so it can tell what is new and what is persisting. The watch list is what stops the AI guessing: it holds the numbers that the narrative never states.

A filled-in watch list for week 17 (illustrative). You update the forecast column from the notes, as you would update the programme anyway, and a sheet fills in the three right-hand columns. The AI never counts days.

Cost lines sit beside it as plain text: concrete package, budget $420,000, 40% still to buy; scaffold hire, budget $86,000, now $3,600 a week; contingency $300,000, of which $54,000 is committed so far.

Notes need no special format, but they need dates, quantities and names of the activity. Compare a note the AI can't use with one it can (illustrative):

Before: "Pour looking like next week-ish, depends on rebar. Brickies a bit short. Windows still haven't been signed off."

After: "Level 3 deck: formwork 80% done; rebar delivery moved by the supplier from Monday to Thursday; pour now forecast Monday week 20. East elevation brickwork: four bricklayers on site, six planned, until the 20th. Window samples sent to the client's architect in week 13, no answer."

The second note is no longer, and the site manager can write it in two minutes if the template asks for activity, forecast date and the reason it moved. A voice note on the walk-round does the same job, and the tutorial on daily logs shows one way to turn it into text. Without that, the flags will be only as good as "next week-ish".

Date arithmetic is the one thing I would take out of the AI's hands. Ask it to quote the dates from the notes and let a sheet count the working days. In the watch list sheet that is one formula per row, =NETWORKDAYS(baseline, forecast, Holidays)-1, where Holidays is a named list of non-working days. The reason is practical: a pour that moves from a Friday to the following Tuesday is two working days, not four, and a language model counting calendar days will report four.

The weekly prompt, and the flags it returned for week 17

Here is a prompt I would start from. It asks for evidence, refuses estimates and keeps recommendations out, because the evidence is what you check and the recommendations are your job.

You help a construction project manager prepare weekly early-warning flags.
Use ONLY the four sections below: THRESHOLDS, WATCH LIST, LAST WEEK'S FLAGS, THIS WEEK'S NOTES.
1. Return a table of items that cross a threshold, or where an instruction has no agreed price.
   Columns: area, item, amber or red, the rule it breaks, an exact quote from the notes,
   the figures you used and the sum, one question for the project manager.
2. Use slip, movement and float figures exactly as the watch list gives them; never recalculate
   them. If the notes give a date that differs from the watch list, flag the mismatch.
3. If a flag needs a number I have not given (budget, quantity, rate), write NEEDS and
   name the number. Never estimate it.
4. After the table, list (a) items on last week's flags that this week's notes do not mention,
   and (b) activities on the watch list that the notes do not mention.
5. Do not recommend actions, owners or contract positions.
[paste the four sections]

The week 17 notes it was given (illustrative):

"Level 2 slab stripped and signed off Tuesday. Level 3 deck: formwork 80% done, rebar delivery moved by the supplier from Monday to Thursday, pour now Monday week 20 weather permitting. Frame subcontractor says the stair core drawings are at revision D and wants to talk about extra labour. East elevation brickwork: four bricklayers on site, six planned, until the 20th. Window samples went to the client's architect in week 13, no answer yet. Concrete supplier letter in: price up 6% from the 1st of next month, haven't checked whether we are covered. Scaffold hire extended by two weeks for the east elevation. RFI 31 (balcony drainage) still open, raised week 6."

And the flags that came back, trimmed to the essentials (illustrative output):

1. Red, programme: level 3 deck pour. Rule: float left 1 working day or less. Quote: "pour now Monday week 20 weather permitting". Figures from the watch list: slip 5, float left 1 of 6. Question: what else could move this date?
2. Red, information: RFI 31. Rule: open more than 15 working days. Quote: "still open, raised week 6". Figures: week 6 to week 17 is 11 weeks. Question: has anyone answered it informally?
3. Red, cost: scaffold hire. Rule: forecast over budget by 5% or $25,000, whichever is hit first. Quote: "extended by two weeks". Figures: 2 weeks at $3,600 is $7,200, which is 8.4% of the $86,000 budget. Question: was the extension already in the hire forecast?
4. Amber, cost: concrete package. Rule: 3% or $10,000, whichever is hit first. Quote: "price up 6% from the 1st of next month". Figures: 40% of $420,000 is $168,000 still to buy; 6% of that is $10,080, or 2.4% of the package, so the dollar limit trips first. Question: is the package fixed-price?
5. Amber, procurement: window samples. Rule: sample outstanding more than 2 weeks. Quote: "went to the client's architect in week 13, no answer yet". Figures: 4 weeks. NEEDS: window lead time, to test the red rule.
6. Amber, changes: stair core drawings. Rule: instruction with no agreed price. Quote: "stair core drawings are at revision D and wants to talk about extra labour". Question: is revision D a formal instruction, and has a price been offered?
7. Amber, recurring cause: brickwork labour. Rule: same cause in 3 of the last 4 weeks (weeks 14, 16 and 17, using the weeks-named count carried on last week's list). Quote: "four bricklayers on site, six planned, until the 20th". Figures: slip 0, float 8 of 8.

After the flags come the two lists it was asked for: from last week's flags, "roof insulation order (amber)" is not mentioned this week; from the watch list, no date is given for finishing the east elevation brickwork. Neither is a flag yet. Both are questions for Monday's walk-round.

Three things to notice. The AI quoted the notes for every flag, which is what makes checking quick. It wrote NEEDS for the window lead time instead of inventing one. And it applied "whichever is hit first" correctly on the concrete line: 2.4% is under the 3% limit, but $10,080 is over the $10,000 one. It did not decide anything. Whether the scaffold flag is real is the next section's question.

Checking each flag: six questions that catch false positives

A flag is a prompt for a question, not a finding. Put each one through the same six checks, in this order, because the early ones are the quickest:

1. Is the quote really in the notes, and does it say what the flag says? Search for the quoted words. A flag built on a paraphrase is the first sign of an error.
2. Was the sum done on the right figures? Redo it on a calculator, with the budget and quantity from your own sheet.
3. Is there an allowance or agreement the AI could not know about? Contingency lines, fixed-price packages, agreed extensions.
4. Is it already in the risk register or the RFI log? A flag that duplicates an owned item needs a note, not a new entry.
5. Who said it, and how firm is it? A delivered fact, a supplier's promise and a subcontractor's hope are different strengths of evidence, and the AI treats them alike.
6. What does it cost to be wrong each way? Chasing a false flag costs a phone call. Dismissing a real one can cost a week of programme.

Here is what those checks did to the week 17 list (illustrative):

Four confirmed, two dismissed, one on watch. The two dismissals teach more than the confirmations. The RFI flag was right about the note and wrong about the risk: the log was out of date, which is a housekeeping fault, though the one the AI could never see. The scaffold flag was correct arithmetic against the wrong budget, and the fix goes into the watch list as "known allowances", so it doesn't come back next week. For more on keeping the RFI log itself honest, see AI tools for RFIs, logs and reports.

Two other false positives are worth knowing because they come back. The first is calendar days against working days. If you let the AI count, a pour that moved from Friday to Tuesday can be reported as 4 days late when it is 2, which can push a flag over a threshold it should not have crossed. The second is a flag built on a forecast that was optimistic to begin with: a subcontractor says "back to six bricklayers on the 20th", the AI treats that as fact and clears the flag, and three weeks later the gang is still four. For the second, keep a column in the flag log for who gave the date, and trust a date less each time it moves.

From a confirmed flag to a risk register entry

Only confirmed flags go into the register, and each one becomes an entry with an owner and a date. The AI can draft the entry from the evidence, as long as you fill in the likelihood, the impact and the response. Here is the deck pour as an entry (illustrative; the scoring is the manager's judgement):

• ID and title: R-17, level 3 deck pour delayed beyond float.
• Cause: reinforcement delivery moved from Monday to Thursday; the pour is weather-dependent.
• Effect: 1 working day of float left (of 6); a further move delays level 4 formwork and the frame subcontractor's stair cores.
• Likelihood and impact: likely; moderate (a short delay to the topping-out date if the float runs out).
• Response: confirm the delivery with the supplier in writing; agree a fallback pour day with the concrete supplier; ask the frame subcontractor whether the stair cores can start ahead of the deck.
• Owner and review: site manager owns the response; project manager reviews at the next weekly meeting.
• Trigger: forecast moves again, or the delivery slips past Thursday.

If your register is a simple sheet, the tutorial on a simple AI risk register for small businesses has a template to adapt. Contract notices are a separate matter: if the pour move could trigger a notice or an extension-of-time claim, the contract sets the deadlines, and the commercial lead decides. The AI can add "check contract for a notice requirement" to a flag. It should never say whether one applies.

Silence is not good news: the look-ahead check

The routine's weakest point is what the notes leave out. Site managers tend to write about what went wrong or what is in front of them, and an activity that is quietly slipping may not be mentioned at all. Step 4 of the weekly prompt lists watch-list items with no mention, but you can ask a sharper question each week by pasting the three-week look-ahead from the programme:

Below are (1) the three-week look-ahead and (2) this week's site notes.
List every look-ahead activity that the notes do not mention at all, and every activity
mentioned only in passing with no date or progress figure. Do not guess why. Output two lists.
For each item, write one question for the site manager to answer on Monday.

In week 17 (illustrative) the look-ahead had eight activities and the notes touched five. The unmentioned three were level 1 mechanical and electrical first fix, level 1 fire stopping and external drainage. The questions went to the site manager on Monday, and the answer on the first was: "three days behind, the electricians were pulled to another site". Nothing in the notes had hinted at it. That is a programme flag that no amount of reading the notes would have produced, and it is why the silence list belongs in the routine rather than being an extra. If the same activity stays silent for two weeks, treat the silence itself as an amber.

Keeping score: a flag log, and when to move a threshold

After the first month the question is whether the thresholds are right. A flag log answers it in five minutes a week: for each flag, one row with the week, the item, who dated it, the outcome (confirmed, dismissed or watch) and, for dismissals, the reason. Count the outcomes. Here is a six-week tally for the illustrative project:

Twenty of 44 flags were confirmed (45%) and a further six were worth watching, so 26 of 44 (59%) earned the manager's attention. Dismissals fell from 5 to 2 a week because each one fed a fix: known allowances went on the watch list, the RFI log was cleaned, and one threshold moved. Read your own table the same way. If you dismiss more than half the flags three weeks running, tighten a threshold or add the missing context to the watch list. If you dismiss none, either the thresholds are loose or the checking has become a habit, so pick one confirmed flag and trace it back to the evidence.

A quick contingency check belongs in the same session, because the contingency rule is the one that moves slowly. With $54,000 committed of $300,000, that is 18%, against 33% of the programme elapsed (17 weeks of 52). Add the confirmed concrete rise of $10,080 and it is $64,080, or 21%. The rule flags only when the committed share runs 10 points ahead of the programme share, so there is no flag this week. It is worth watching because the stair core labour has no price yet, and an unpriced instruction is the kind of item that can move this line quickly.

The weekly time budget, and what the AI does not do

For one project the routine runs to about 45 minutes a week, illustrative and mostly human: the site manager's structured notes (10 minutes), updating the watch list forecasts and the sheet (10), running the prompt (2), checking the flags (15) and updating the register and the log (8). The AI's own part is the smallest and fastest. Set-up is a couple of hours once: the thresholds, the watch list template and the saved prompt. If you want a second set of prompt wording to compare, the tutorial on Claude prompts for weekly status and a risk log takes a similar route from a different angle.

The assistant plan is a small cost. ChatGPT Plus is $20 a month and Claude Pro is $20 ($17 a month billed annually), and both list Projects, which is where the thresholds and the prompt live. Many companies have an approved tool already, and that is the one to use: project notes carry commercial figures and subcontractor names, so check your company's rules before anything is pasted, and take names out where you can. Claude's pricing page says its Team plan does no model training by default, while ChatGPT's page lists training as opt-out on its personal plans.

What stays human is worth saying plainly. The AI doesn't set thresholds, count float, judge how reliable a subcontractor's date is, decide whether a flag is real, take a contract position or talk to the client. It reads text faster than you do and doesn't tire of the Friday notes. The manager still owns every decision on the list, and the routine's job is to put the right six or seven items in front of you each Monday, early enough to do something about them.

Questions project managers ask about AI risk flags

Can AI read a programme file or cost report directly?

Some assistants accept spreadsheet and PDF uploads, which helps with cost tables. A native programme file is harder, so export the few fields you need (activity, baseline date, forecast date, float) into a sheet first. Whatever you upload, check the figures it quotes back against the source, because a misread column is an easy slip to miss.

How many flags a week is too many?

There is no fixed number. If you dismiss more than half, tighten the thresholds or improve the notes. If you dismiss none for several weeks, check that you really are testing each flag and that the thresholds are not so loose that real slips pass. Aim for a list you can clear in a short weekly meeting.

Is it safe to paste project notes into an AI tool?

Only if your employer and client allow it. Notes often hold commercial figures and subcontractor names. Use the tool your company has approved, strip names where you can, and read the plan's data terms before pasting anything, since they differ between personal and team plans.

Does this replace earned value reporting?

No. Earned value and your scheduling software measure progress against a baseline with formal methods. This routine reads the narrative around them, the things site teams mention in passing, and turns it into flags. Use it to catch what the numbers have not caught yet, not to replace them.

Further reads

• Construction Job Costing With AI: Find Which Jobs Actually Made Money — Find which jobs actually made money once costs are in.
• Construction Cash Flow Forecast: Build One in a Spreadsheet With AI — Turn the cost flags into a forward cash view.
• Construction Meeting Minutes: Record, Summarise and Chase Actions With AI — Chase the actions your flags create.
• How Can AI Help Construction Project Managers? Submittals and Sub Chasing — Submittals and subcontractor chasing, the usual causes of slip.
• Construction Payment Notice Template: Draft With AI, Check the Deadlines — Deadline checks for notices when a slip becomes a claim.
• Construction Handover Pack Checklist: Gather the Paperwork With AI — Gather the handover paperwork before the end date presses.
• AI in Construction Accounting: Job Costs, WIP and Retainage Checks — Four month-end checks for a small contractor: code invoices, catch job-cost errors, recalculate WIP and track retainage, then hand your accountant a clean pack.
• Benefits of AI in Construction: What a Small Firm Can Actually Measure — Five benefits of AI a small building firm can track with a before-and-after sheet: quote time, site diaries, payment chasing, scope gaps and safety drafts.
• Construction WIP Report: What It Shows and How AI Builds One — What a work-in-progress schedule shows, a three-job worked example of over and under billing, and where AI helps and where it must not.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: claude.com/pricing and chatgpt.com/pricing (plans and data-training rows, read 1 November 2026). Thresholds, figures and project details are illustrative.