What You Need to Know: Automation doesn't fix bad cost data; it scales it. AI treats every record as valid, so errors and outliers flow straight into estimates, and when a tool works as a black box, estimators can't see how a number was built, which means they can't adjust or defend it. Four steps make cost data AI-ready: consistent attributes, estimator-controlled definitions of what makes projects similar, curated project subsets, and outlier flags. Just as important, tools need to let veteran and tech-savvy estimators each set their own level of automation, because trust is what drives adoption.
AI promises to take the research grind off estimators' desks. It can find comparable projects in seconds, flag risks before they become change orders, and update unit prices without anyone touching a spreadsheet, among many other tasks.
But none of that works if the data that comes out of these actions can't be trusted. Automation can’t fix bad cost data; in fact, it will simply scale it.
So here's why data quality is the real starting line for AI in preconstruction, and what teams can do to make sure they’re starting with a solid data foundation.
When an estimator pulls historical costs by hand, they apply judgment along the way. They know the hospital job from 2019 had an unusual site condition. They know the parking structure was coded wrong. They quietly leave out the numbers that don't make sense.
Automation doesn't know any of that. It treats every record as equally valid. A miscoded project, a missing escalation adjustment, or a one-off outlier gets pulled straight into the comparison, and the result looks just as confident as a good one.
That's the danger. A wrong number that looks authoritative is harder to catch than a blank cell. Feed an AI tool poor data and it will work against you, faster than any person could.
The second problem is visibility. Many AI tools work as a black box: data goes in, a number comes out, and nobody can say how it got there.
For estimators, that's a dead end. If you don't know the basis of a number, you don't know where to start adjusting it. You can't explain it to an owner. And you can't defend it when the project team asks why the budget moved.
A glass-box approach works differently. Predictions are built on concrete, traceable records. You can see which projects were treated as comparable, which attributes drove the match, and which data points were set aside as outliers. The estimator stays in control of the number instead of copying a figure into a spreadsheet and hoping for the best.
Traceability only works if the data behind it is clean. You can't show your work if your work is a mess.
Attributes are the context that makes automation possible. They tell the system what a project actually is: its type, size, location, delivery method, schedule, and whatever else matters to how you price work.
There's no magic number. Some contractors track 10 attributes; others track 400. They can be simple values, ratios, or complex formulas. Some will apply to every project, while others only matter for certain sectors. The point is consistency. An attribute that's filled in for half your projects is a gap the AI will fall into.
Most tools decide for you which projects are comparable, and you rarely get to see the logic. A better approach lets the estimator pick a baseline project and choose which factors define similarity, such as project date, size, or type.
That control matters because "similar" depends on the question. A similar project for pricing MEP systems might not be a similar project for estimating general conditions.
If your team mostly bids healthcare work, you shouldn't have to filter out warehouses every time you search. Saved collections of projects, KPIs, or cost elements act as focused mini-databases. Some update automatically as new projects match the criteria. Others stay fixed for a specific analysis.
Curated subsets cut noise and make results easier to trust, because estimators know exactly what's in the comparison set.
Every historical database has anomalies: a project with an unusual scope, a data entry error, a market spike. Outlier detection with adjustable thresholds surfaces those records so estimators can decide whether to include them, rather than letting them quietly distort an average.
The best setups let teams turn outlier filtering on or off and set their own thresholds, because what counts as unusual varies by trade and region.
Clean data and transparent tools also solve a people problem. Estimators who learned the trade by hand are often wary of automation, and for good reason. They've seen what happens when a bad number goes out the door.
Those estimators don't need to be talked into trusting AI. They need tools that let them see the numbers going in and check the logic coming out. Meanwhile, more tech-savvy team members may want to automate as much as possible. The right setup supports both, with automation that can be dialed up or down rather than forced on everyone.
When the data is trustworthy and the reasoning is visible, adoption follows. Estimators spend less time on research and more time on the high-value work: judgment, strategy, and client conversations.
It's tempting to jump straight to the most advanced AI on the market. But the teams that get real value from automation start with the unglamorous work: consistent attributes, clear comparison sets, and a process for catching bad records.
Get that foundation right, and AI becomes a true multiplier for your estimating team. Skip it, and you'll just get wrong answers faster.
Learn More About DESTINI Estimator →