AI for Construction Cost Control: What It Actually Does

Most contractors find out whether a project made or lost money after it finishes, which is a bit late to do anything about it. This guide is for PMs and cost engineers who want AI handling the arithmetic in cost control without handing it the judgment calls. It comes from two of my videos: a full walkthrough of a construction cost tracking system, and a deep dive into applying earned value management in practice, not textbook theory.

Key takeaways

• Cost control is not estimating and it is not accounting. It is tracking whether the work you already priced is making or losing money while you can still change the outcome.
• Every cost code needs a quantity and a rate, tracked against the same work breakdown structure as your schedule, or your cost value reconciliation stops meaning anything.
• AI’s job is the arithmetic: coding actual costs, computing CPI and SPI, rolling forward a forecast. Percent complete stays a human call, always.
• Subcontract cost value reconciliation runs on a different logic to self-perform work. Track variations, not productivity, and accrue the cost when the work is done, not when it is claimed.
• Forecasting off a project-wide cost performance index is the wrong way to do it. Forecast by activity, and treat unapproved variations as a risk, not a certainty.

What is AI for construction cost control?

AI for construction cost control is using a tool like Claude to code actual costs against your budget, compute earned value metrics like CPI and SPI, and roll a forecast to completion, then draft the monthly report. The two inputs that decide whether any of it is trustworthy, percent complete and the finishing rate, stay a human call.

Cost control is a different discipline to estimating. An estimate is how you calculate the price you sign a contract for. A budget is how you track spend against that price once it is signed, broken into cost codes, and every code should carry both a quantity and a rate. That distinction matters because a code can go wrong two ways: the quantity moves (a 20-week project runs 30 weeks), or the rate moves (an $8,000-a-week supervisor is actually costing $10,000), and you want to know which one it is before you can fix it.

How you break codes down should follow where the money actually sits, not a fixed template. On one commercial project I’ve broken down, materials ran 36% of cost, labour 38%, and project management as little as 2%, and I have seen projects where any one of those numbers is completely different. Break your codes down further wherever the biggest number is; a 1% cost category does not need twenty codes to track it.

Cost control is also not accounting. Accounting exists for tax reporting. Cost control exists to answer one question while you still have time to act on it: are we making or losing money on the work we have already priced? “Forecasts are always wrong, but they’re still useful,” is a line I keep coming back to, because the whole point is bringing that insight forward to a point where you can still influence the outcome, not finding out on the last invoice.

The stack I use for this is the same one that runs the rest of a project: a cost tracking sheet with a cost code per bucket of spend, Claude reading the accounting export and dockets to code actual costs against those buckets, and a monthly cost value reconciliation that turns coded actuals plus percent complete into a forecast. If none of that is set up yet, AI for construction workflows covers where this workflow sits alongside contract admin, procurement and drawings.

How does earned value management actually work?

Earned value management (EVM) compares three numbers for the same work: planned value (what you should have done by now), earned value (what you have done, in dollars), and actual cost (what you have spent). From these you get schedule performance index (SPI = EV/PV) and cost performance index (CPI = EV/AC): ahead or behind, over or under budget.

I explain it with a road trip. Melbourne to Sydney is 878 km, budgeted at $120 in fuel over a planned 9 hours. Three hours in, you should have covered 290 km (your planned value). You have actually covered 350 km (your earned value), so you are ahead of schedule. You have spent $50 on fuel, but 350 km of a 878 km trip should only have cost $48 (your earned value in dollars), so you are spending slightly more than planned even though you are further along. SPI works out to 1.21, CPI to 0.96. The whole system runs on that same idea, just applied per cost code instead of per kilometre.

The hard part is not the arithmetic, it is measuring percent complete honestly. On a worked underground electrical example, week four had a plan of 200 metres of trenching, actual progress of 180 metres, and an actual cost of $32,000. That gives an SPI of 0.9 and a CPI of 0.84, both because the crew hit rock and had to hire a rock breaker. Two numbers, one root cause, and neither one tells you the cause on its own, which is the actual value of tracking both together.

Subcontract packages need a different lens. If a subcontractor has pulled 60% of the cable but only claimed 42% of the value, your CPI will read as 1.24, which looks great and is meaningless. Progress claims lag the work. The fix is to accrue the subcontractor’s cost when the work is physically done, not when the claim arrives, or your cost performance index just measures billing timing instead of performance.

Where does AI fit in cost value reconciliation, and where does it stop?

The split holds across every step of the monthly cost value reconciliation. AI reads the accounting export, codes the transactions, does the CPI and SPI arithmetic, and rolls the numbers into a forecast and a report. You set the percent complete, judge the rate you will finish at, and decide what an amber or red variance actually means.

CVR task	What AI does	What stays with you
Cost-code budget setup	Restructures the estimate into cost codes on the same WBS as the schedule	Confirming every code ties to real scope, setting contingency
Coding actual costs	Reads the accounting export plus dockets and timesheets, matches each cost to a code	Reviewing anything that will not code cleanly
Percent complete	Nothing. This one is off the table for AI	Setting the true percent complete per activity, eyes on the work
CVR arithmetic	Computes CPI, SPI and the cost variance once percent complete is set, notes when a code crosses the breach threshold	Judging whether the variance is a productivity, quantity or claim-timing problem
Forecast to completion	Recalculates cost-to-complete from the judged rate and quantity remaining	Judging the rate you will finish at and the demobilisation tail
Monthly cost report	Formats the structured data into the report	Checking it before it goes to the client, a polished report can still be quietly wrong

Two things are worth naming on their own. First, percent complete never comes from budget spent. “We spent 60% of the budget, so it’s 60% done” is the worst way to measure progress, because it is circular, spend does not tell you what got built. Use a clean quantity where one exists (metres of trenching), or a rule of credit system where it does not (a pit is 20% at excavation, 30% at base, and so on through to the lid). Second, watch the S-curve. Every activity starts slow, gets productive in the middle, and closes out slow again with defects and demobilisation, so 95% of the quantity installed is not 95% of the activity complete. Ignore the tail and your forecast comes in light every time.

Those cost-actuals and CVR-forecast skills, along with the cost tracking spreadsheet template, are in the ContractorOS community for members setting up their own version of this loop.

How do you forecast cost to complete without kidding yourself?

Forecasting cost to complete means adding what you have spent to a judged estimate of the remaining work, and the method changes by cost type. Self-perform uses cost-to-date plus a judged rate times quantity remaining, including the demobilisation tail. Subcontract is a lump sum, so you only forecast variations. Overheads are burn rate times remaining duration.

One shortcut is genuinely wrong and worth naming: taking your overall project CPI and dividing your total remaining budget by it (estimate at completion = budget at completion ÷ CPI). That formula only holds at the activity level, where the cause of the variance is consistent. Blend it across a whole project and you are averaging away exactly the information you need, a trenching crew that hit rock and a subcontract package running to plan are not the same story, and a blended number tells you neither.

Pending variations deserve the same discipline. The moment you forecast an unapproved variation as if it is certain, you have overstated your position, and if the client knocks it back or prices it lower you have to walk the number back. Treat the profit on a pending variation as an opportunity in your risk register instead, weighted by how likely it is to be approved, not as revenue you can already count.

If a subcontract cost is running into real disagreement over entitlement rather than simple arithmetic, that moves into contract administration territory, notices, time bars and the negotiation itself, covered in AI for construction contract administration. Cost control tells you something changed. Contract admin is how you recover for it. And because so much of a forecast depends on how long the work will actually take, keeping the schedule current matters just as much as the cost side, which is its own workflow covered in AI Gantt chart for construction.

Common mistakes to watch for

• Measuring percent complete by budget spent. It is circular and it is the single fastest way to make a forecast lie to you.
• Letting the schedule and the budget use different groupings. If your schedule has one line for “underground electrical” and your budget splits it into trenching, pits and foundations, nothing reconciles cleanly.
• Valuing subcontract earned value off the claim instead of the work done. Claims lag reality and will make your CPI look better or worse than it is.
• Forecasting with a project-wide CPI. Estimate at completion = budget ÷ CPI only holds per activity. Applied to a whole project, it hides the actual cause.
• Forecasting unapproved variations as certain. They belong in the risk register at a weighted value, not in the budget as confirmed revenue.
• Ignoring the S-curve tail. A task that is 95% through its quantity is not 95% complete once you account for defects and demobilisation.

None of this replaces a cost engineer. It gives one a head start on the parts of the job that are pure structure, so the hours go into the judgment calls instead of retyping numbers from an invoice into a spreadsheet. The full worked examples, including the cost tracking sheet and the earned value calculations line by line, are in the videos above.