How AI Is Changing Pre-Construction Validation in AEC
How AI Is Changing Pre-Construction Validation in AEC
What automated checks can catch before a set leaves the office, and what still needs a person.
Almost every firm reviews its drawings before they go out. Ask what that review covered, and the answer usually gets vague.
In most firms I've worked with, the set gets checked by whoever has time in the last week before an issue. One project architect looks at egress. Someone else checks door clearances if the deadline allows. How good the review is depends on the person and the week.
That's the part AI can change. The review still needs people. What AI can take away is the excuse for doing it differently every time.
Take a common example. A corridor meets the IBC's 44-inch minimum width in schematic design. In design development, a column enclosure and a duct chase eat four inches on one side, at one point. Nothing on the sheet looks wrong. Maybe the plan reviewer catches it. Maybe the framer does. Either way, your firm fixes it after it has already paid for those drawings once.
Or a rated corridor wall is drawn with the right wall type while the specification calls for a different tested assembly. Each document is correct on its own, and nobody put them side by side.
Both examples follow rules. A rule can run against the model every time the model changes, instead of once at the end by whoever happens to be free.
Where the industry is today
Architects are curious about AI. Very few have put it to work.
In the AIA's 2025 study on AI adoption, produced with Deltek and ConstructConnect, 53% of architecture professionals said they were experimenting with AI. Only 6% used it regularly. Among firm leaders, 8% said AI was integrated into their practice, and another 20% were in the middle of implementing it.
Most of that use is chatbots and writing tools. They save time, but they sit outside production. If you run a firm, the gap worth watching is the one between trying AI and letting it check a set before that set is issued.
- Use AI regularly6%
- Experimenting53%
- Other41%
- Integrated8%
- Implementing20%
- Considering35%
- Other37%
What bad information costs
This argument isn't new. In 2021, Autodesk and FMI surveyed more than 3,900 construction professionals worldwide and estimated that decisions made on bad data caused 14% of all rework in 2020, roughly $88.69 billion globally. It's a self-reported, global figure, so it's better read as a direction than as a precise U.S. number.
The direction still matters. A good share of rework begins with information that was already wrong, or already contradicted itself, before anyone built from it. Your drawings, schedules and specifications are that information. Checking them before you share them is the cheapest moment you'll get to step in.
What the evidence says about AI checking
The research on AI-assisted code checking is encouraging, and it's early. A 2025 study in Electronics connected several large language models to Revit. The models read 12 rules from the International Residential and Mechanical Codes, wrote checking scripts and flagged violations. Checks took less time. Results also swung widely from one model to another.
The authors were candid about the limits. The framework worked best on well-structured models with accurate room tags, boundaries and naming, and people still had to run the scripts and fix errors.
Two practical lessons follow. AI checking is only as reliable as the model it reads. And the architect's review stays; it just starts from a shorter list that's already sorted.
Six ways to build validation into production
Firms that get real value from automated validation start by deciding what to check. Then they make their models checkable and put someone in charge of the results.
1. Write down what gets checked
Start with a rules register, a short list of the checks that matter most for the kind of projects you do. For a multifamily or light commercial practice, that might mean 20 to 30 rules: corridor and door widths, maneuvering clearances, stair headroom, fire ratings on the egress path, plumbing fixture counts against occupant load.
Give each rule its source (a code section, an owner standard or your own firm standard), the phase where it's first checked, and an owner. Without that list, "we use AI for QA" ends up meaning whatever each person felt like trying.
2. Make the model checkable
A check can only read what the model holds. If rooms aren't tagged, walls carry no fire-rating parameter and doors have no clear width, no tool can validate them, with AI or without it.
In Revit, this usually comes down to a handful of firm standards. Rooms are placed and tagged with their occupancy. Rated assemblies live in type parameters. Door families carry clear width and clearance zones. Naming is consistent. ISO 19650 and most BIM execution plans already ask for this. The hard part is holding the line on every project, and the 2025 study landed where practitioners already are: checks break on messy models.
3. Rule-based checks first, AI second
Most dimensional and geometric rules don't need AI at all. A Dynamo script, Revit's model checking add-ins, or dedicated model checking software will test corridor widths or door clearances the same way every time, and anyone can audit the result.
AI is worth it where a rule is hard to write down. Reading the specification against model data. Comparing schedules with sections. Catching notes like "match existing" or "see spec" that are really an open decision in disguise. The order matters: measurable rules go to deterministic checks, and AI helps with what needs interpretation. When an owner or a plan reviewer asks why something was flagged, you can explain it.
4. Tie checks to milestones
IA check that only runs when someone remembers is still inconsistent. Attach each rule in your register to a moment that already exists in your process: phase gates, issues to consultants, permit submission. We cover phase gates in The Real Cost of Uncontrolled RFIs in Design-Build Firms.
5. Put a person on every flag
Automated checks throw false positives, and AI can be wrong with complete confidence. Every run needs a named reviewer who looks at each flag, decides whether it's a real issue, an acceptable condition, or a tool error, and closes it.
That's also where professional responsibility sits. The architect of record signs the set, and good tools give that signature better information to stand on. A short note on who reviewed which flags, and when, is worth having if questions come up later.
6. Measure what the checks catch
Every flag that turns out to be real is a problem that stayed out of the field. Log them by rule, phase, and discipline, and you'll see which checks earn their keep and which rules keep failing.
After a few projects, that log is your business case. It shows repeat issues going down and where your model standards still need work. It also tells you when a new tool is worth buying, and when the real problem is the process underneath it.
| Rule | Source | Check type | First phase | Owner |
|---|---|---|---|---|
| Corridor clear width | IBC egress | Rule-based | SD | Project architect |
| Door maneuvering clearance | ADA / ICC A117.1 | Rule-based | DD | Project architect |
| Rated wall type matches spec assembly | Firm standard | AI-assisted | DD | Spec writer |
| Open notes (“see spec”, “match existing”) | Firm standard | AI-assisted | CD | BIM manager |
Where to start
If your firm has 10 to 50 people, start this month with a model readiness test before buying anything:
Pick one active project in design development.
Choose five rules from your register that should be checkable in the model, such as corridor widths, door clear widths, or fire ratings on the egress path.
For each rule, check whether the model holds the data it needs: tagged rooms, rating parameters, clear widths.
Run the checks the data supports, and list every rule the model couldn't answer.
That second list tells you more than the first. It shows what your model standards need before any tool, AI or not, can check your sets reliably.
What changes
AI won't review a set by itself, and firms expecting that will be disappointed. What it can give you is consistency: the same rules on every project, at every milestone, with someone accountable for the result. In my experience, that consistency is what keeps errors in the office, where they cost hours, and away from the field, where they cost margin.
If your review depends on who has time before the deadline, the first step is finding where validation breaks down. Start with a Production Control Diagnostic™.
Frequently asked questions
Does AI replace plan review or my firm's QA/QC? No. It changes where your reviewers spend their time. Automated checks take the repeatable rules so people can focus on judgment calls. The architect of record is still responsible for the documents.
Do we need new software to start? Not at first. Plenty of dimensional checks run in Revit with Dynamo or add-ins you may already have. Dedicated checking tools and AI start paying off once you have a rules register and model standards in place.
How accurate are AI code checks today? That depends on the tool, the model, and the rule. So far, research shows useful results on well-structured models, weaker ones on inconsistent models, and people still correcting the output. I treat AI flags as a list to review, and a person makes the call.
Which rules should we automate first? Start with the ones that repeat and cost you the most, usually egress widths, door clearances, fire ratings along exit paths, headroom, and fixture counts. Your own RFI and rework history will point you there faster than any checklist.
Who should own validation in a small firm? A senior project architect or BIM manager with the authority to hold a set at a milestone. Without that authority, a failed check becomes a suggestion.
Validation that keeps your judgment in the loop.
We set up the checks that catch errors early, without handing your decisions to a black box.