Why is preconstruction where the margin is decided?
Because by the time crews mobilize, the decisions that determine profitability have already been made by a handful of people working against a deadline. How scope was interpreted. Which quantities were trusted. What risk was carried and what was passed along.
The industry’s overrun record suggests those decisions go wrong often. KPMG’s global construction survey found that only around 31% of projects came within 10% of budget over a three-year period. Multiple industry analyses put the share of US projects exceeding their initial budget near 70%, with preconstruction estimating cited as a leading contributor rather than field execution.
What follows is a breakdown of the specific problems inside that phase, and an honest account of which ones automation actually addresses.
Where does an estimator’s week actually go?
Not where you would want it to.
A large share of preconstruction hours goes to work that consumes expertise without requiring any. Measuring linear footage. Counting devices. Typing subcontractor numbers into a tally sheet. Reconciling two proposals that describe identical work in different words. Industry reporting puts manual quantity takeoff at three to five days per project for many teams, before any pricing judgment is applied.
This is the clearest case for automation in the whole phase, because the work is repetitive, high-volume, and mechanical once the drawings are understood. The estimator’s job shifts from producing the count to checking it.
The adoption data suggests the category has already passed the argument stage. BuiltWorlds’ 2026 Annual Preconstruction Benchmarking Report found that nearly 90% of surveyed contractors have implemented an estimating solution, and 64% now use one on every project, up from roughly 13% in 2022. Bid management adoption passed 70% regular use. More than 40% of regular estimating users are piloting additional tools alongside their primary platform, which tells you where teams still feel underserved.
Where this lands in practice is a division of labor rather than a replacement. AI-assisted quantity takeoff with estimator verification is one form of it: the model produces the measurement, a credentialed estimator confirms it, including scope the drawings imply but do not show.
Can AI improve estimating accuracy, or only speed?
Both, but not in the way the marketing usually implies.
AI does not make a cost baseline more accurate. A historical cost database built from completed projects is what tells you whether a number is plausible for this building type, in this market, at this scale. That anchor is the product of accumulated project records, and no model substitutes for it.
What automation changes is how much gets checked against that anchor, and how early. When quantity extraction takes days, assumptions get set once and revisited only if something breaks. When it takes hours, an estimating team can reconcile successive estimates, explain what changed between them, and catch drift while there is still time to price it.
There is a second-order effect worth noting. Subcontractor proposals carry live market signal on labor availability, material pricing and lead times, and most of that signal arrives during the busiest week of the project and is never structured into anything durable. Historical data tells you what normal looks like. Incoming bids tell you whether this quarter is normal. Teams that capture both are working with a fuller picture than teams that capture only one.
Does bidding faster help, or does it just mean bidding more?
This is the question most speed claims avoid.
Faster preparation only protects margin if it is used to bid selectively. The failure pattern is well documented: when the pipeline feels thin, the instinct is to bid everything that arrives, estimators get buried, reviews get rushed, and the details that protect margin start slipping. A full pipeline is not a profitable one.
Market conditions in 2026 have sharpened this. Contractors are reporting historically high backlogs alongside compressed margins, which rewards firms that are selective rather than firms that chase volume. Automation is genuinely useful here, but the benefit is capacity to say no rather than capacity to say yes more often.
The practical test for any team adopting these tools: did the hours saved go into more pursuits, or into better scrutiny on fewer? Only one of those protects margin.
What breaks when five estimators use five methods?
Consistency, and then everything downstream of it.
Every estimator develops personal habits. One relies on calendar reminders, another keeps handwritten notes, a third builds a bespoke spreadsheet for each bid. Individually these work. Collectively they produce five different waste factors, five different exclusion conventions, and a chief estimator merging separate spreadsheets by hand at the worst possible moment.
Shared structure fixes most of this without any AI at all. A single template, locked conventions, and one place where the current state of a bid lives. What AI adds is the ability to enforce that structure automatically rather than relying on discipline during bid week, which is exactly when discipline degrades.
Why do scope gaps survive review?
Because they are not reading failures. They are reasoning failures, and they are invisible in the documents themselves.
An exclusion is disclosed. A subcontractor writes that fire alarm devices are not included, and an experienced estimator who reads the full proposal will catch it and price it. The problem is time: on the eighth package on Friday, nobody reads page 38 of a nine-page qualifications appendix as carefully as they read page 38 on Monday.
A scope gap is worse, because nothing in any proposal points at it. Core drilling falls between the electrical and concrete packages, so neither subcontractor carries it. Both proposals look complete. The gap surfaces in the field.
Consider a three-bid electrical package where the apparent low bidder has excluded fire alarm devices in prose. Subs B and C both carry it. Priced back in, the low bid is no longer low. Nothing about catching that requires sophistication. It requires someone to have read one sentence, at the end of a week when eight packages closed.
This is where automated extraction earns its place. AI bid leveling for general contractors produces a comparison with one row per scope item and one column per bidder, and where a bidder is silent, flags the omission rather than leaving the cell blank. An empty cell reads as zero cost. A flag demands a decision. On a thirty-package project, one ENR Top 10 contractor reported roughly 150 hours returned on tally entry, about five hours per package, with review running under thirty minutes per package.
The judgment stays where it was. Whether a gap matters, what a plug value should be, and whether to issue a clarification are all estimator calls.
What happens when experienced estimators retire?
This is the problem underneath all the others, and it is the least discussed.
The AGC-NCCER 2025 Workforce Survey found 92% of contractors reporting difficulty filling skilled trade positions. Roughly one in five US construction workers is approaching retirement age. Estimating departments are running more simultaneous pursuits than they were three years ago with broadly the same headcount, and the people carrying the institutional knowledge about how a firm prices work are the ones closest to leaving.
No software solves that. What it can do is extend the reach of the estimators a firm still has, and capture some of the reasoning that currently lives only in someone’s head.
An emerging structural response is worth noting, because it is the opposite of the replacement narrative. In July, the estimating firm MicroEstimating announced a partnership with MeltPlan that puts AI takeoff inside its workflow, with credentialed estimators reviewing the takeoff output and pricing the verified quantities against a cost database built from hundreds of completed projects.
The arrangement is unremarkable as technology and interesting as a design decision. It assumes the software produces a draft rather than an answer. “AI speeds up the process, but human expertise validates every detail,” is how the announcement frames it. MicroEstimating founder Henry Tooryani described the goal as removing a trade-off rather than a job: speed has always been available to estimating teams willing to give up rigor, and rigor has always been available to teams willing to give up the schedule.
Whether that model holds up commercially is a question the next year will answer. As a response to a scarcity problem rather than a productivity pitch, it is the right shape.
What gets lost between estimating and operations?
The reasoning, almost always.
A project manager inherits a number. What they frequently do not inherit is the set of assumptions behind it: which exclusions were accepted knowingly, which vendor quotes were compared and rejected, which allowances were deliberately thin, and which scope concerns were raised and resolved. Without that, operations cannot tell whether the framing budget is tight or whether there is padding in the electrical allowance, and cannot make sound decisions when field conditions change.
The classic version is the call three weeks in: someone asks whether the driveway was included, the estimator who priced the job is on the next pursuit, and nobody recorded that it was explicitly excluded during a walkthrough.
Preconstruction is a chain, and most of the waste sits at the joints between its links. Every handoff is a re-entry point, and every re-entry point is somewhere an assumption can quietly disappear. Mapping how takeoff, estimating and bid leveling actually connect is a useful exercise for any team trying to locate where its own losses occur. Systems that carry the leveled analysis through award and buyout rather than stopping at the comparison address this directly, because the record of why arrives with the number.
How accurate is AI in preconstruction, really?
Less accurate than the marketing and more useful than the skeptics allow.
Worth saying plainly: the software gets things wrong. It misreads handwriting. It misreads tables that break across a page. Occasionally it will normalize two line items that any estimator would recognize on sight as different work. Extraction degrades on poor scans, on tables rendered as images, and on highly non-standard layouts, which is to say it degrades on exactly the documents that are hardest to read manually.
Published accuracy figures in this category are almost entirely self-reported and rarely tested independently. Treat any percentage as a starting point for your own evaluation rather than a specification.
The more useful question is what happens when the tool is wrong, and how quickly you find out. If every value traces back to the passage it came from, verification means opening a source rather than re-reading the document set, and a package can be checked in well under an hour. If it does not, you will end up checking manually anyway, and you will have paid for the privilege.
That principle, rather than any accuracy claim, is what separates tools worth evaluating from tools worth avoiding.
What should you ask any AI preconstruction vendor?
Five questions, and they apply regardless of which product is in front of you.
Two further points of process. Run the evaluation on your own worst package, ideally one with several bidders, a handwritten submission and a late addendum, rather than on the vendor’s sample data. And ask directly which capabilities are shipping today and which are roadmap, because this category is moving quickly enough that demonstrations routinely mix the two.
What does this mean for the estimator’s role?
It is worth being precise, because the replacement narrative has not aged well and it was never accurate.
What automation removes is transcription. Measuring quantities off a drawing set. Typing bid numbers into a spreadsheet. Reconciling proposals that describe identical work in different language. None of that requires expertise, and all of it currently consumes it. McKinsey has estimated that generative AI can improve productivity in knowledge-heavy roles by 20 to 40%, and preconstruction sits squarely in that category because so much of the week is spent moving information rather than judging it.
What it does not touch is the part that earns the salary. Knowing that a particular subcontractor bids low and recovers through change orders. Judging whether an allowance is adequate for what the market is doing this quarter. Reading how much risk a specific owner will actually tolerate. Deciding which gaps are real and which are paperwork.
The estimating profession has heard a great deal of promise about automation in the past two years, and readers of this publication are better positioned than most to evaluate it. The narrower claim is the defensible one: the extraction and normalization work in preconstruction can be automated and probably should be, because it consumes scarce expertise and produces nothing that requires it. The interpretation cannot be. What good software offers is a trustworthy foundation for judgment, arriving early enough in the week to be useful.
Quick answers
What is bid leveling?
The process of normalizing subcontractor proposals to a common scope so they can be compared on equal terms. The leveled totals, not the submitted prices, show who is genuinely lowest.
What is the difference between a scope gap and an exclusion?
An exclusion is work a bidder states it is not carrying, so it is disclosed and can be priced. A scope gap is work no bidder carried, usually because it fell between two packages. Exclusions are a reading problem. Gaps are a reasoning problem.
What is a plug number?
An estimator’s judgment of what a missing scope item would have cost had the bidder included it, used to bring an incomplete bid up to full scope for comparison. It should be recorded as an estimate, never as a confirmed price.
Does AI replace construction estimators?
No. It replaces the data entry inside the role. Scope interpretation, risk assessment and award decisions remain with the estimator, and every serious tool in this category is built around a verification step rather than an autonomous output.
Does AI bid leveling replace bid distribution software?
Not usually. The two solve different problems, and increasingly they connect: Melt Bid, for example, integrates with Autodesk BuildingConnected so invitations can be issued from within the workflow, while its focus stays on the analysis after proposals arrive.
Where should a firm start?
With the step where hours are being lost to work that does not require expertise. For most commercial general contractors that is takeoff or bid leveling, because both operate on documents the team already has and require no field infrastructure.
Sources: KPMG Global Construction Survey; BuiltWorlds 2026 Annual Preconstruction Benchmarking Report; AGC-NCCER 2025 Workforce Survey; McKinsey & Company research on generative AI productivity in knowledge work.
