Every construction software vendor now has an AI story. Schedules that fix themselves, cost reports that write themselves, risk registers that predict the future. Owners are being asked to pay for these features, and some of their contractors and consultants are already using them on their projects, sometimes without saying so.
The useful question isn't whether AI is coming to project controls. It's where it actually helps today, where it doesn't, and what an owner needs to put in place so the tools don't create new risks. This post takes a practical, owner-side look at all three.
Why project controls looks like a good fit
Project controls runs on data: schedules with thousands of activities, cost reports, change logs, RFIs, submittals, daily reports and meeting minutes. Much of the work is checking that data for errors and spotting patterns in it. That's the kind of work software is good at.
The catch is the quality of the data itself. SmartPM, a schedule analytics vendor, reported in its 2025 State of Construction Scheduling Report that 88% of the baseline schedules it analyzed failed to meet industry-recognized quality benchmarks. The report drew on more than 70,000 schedules and a survey of over 3,500 professionals. A vendor has an interest in that finding, so read it with care, but it matches what we see: on programs we've supported, the first problem with any analytics tool is that the schedule underneath it wasn't built well enough to analyze.
That's the theme of this whole topic. AI can make good data more useful. It can't make bad data reliable.
Use 1: Schedule quality checks
This is the most mature use, and much of it isn't really AI. The best-known framework is the 14-point schedule assessment developed by the U.S. Defense Contract Management Agency (DCMA), which software tools such as Deltek Acumen apply automatically. Its documented thresholds include:
- Activities missing logic: no more than 5% of incomplete tasks
- Leads: none
- Lags: no more than 5% of logic links
- Finish-to-start relationships: at least 90% of logic links
- Hard constraints: no more than 5% of incomplete tasks
- Activities with very high float or very long durations (over 44 working days) flagged for review
These are rule-based checks. A tool runs them in seconds, and they catch the structural problems that make a schedule unreliable. Newer tools add a layer on top: comparing updates month to month, flagging logic changes and duration cuts, and writing a plain-language summary of what changed. That last part is where language models help, because the raw output of a schedule comparison is unreadable for most executives.
View the data
| Check | Result | Threshold | Outcome |
|---|---|---|---|
| Missing logic | 8% | 5% maximum | Review |
| Lags | 3% | 5% maximum | Pass |
| Hard constraints | 4% | 5% maximum | Pass |
| Long durations | 6.5% | 5% maximum | Review |
Passing a 14-point check doesn't make a schedule realistic. A schedule can have perfect logic and still assume a crane that isn't booked or a permit that won't arrive. The check is a filter, not a verdict. Our baseline review checklist covers the judgment the software can't apply.
Use 2: Document review
Large language models are good at reading. Practical uses on the owner's side include:
- Summarizing a long contract, specification section or claim submission
- Finding every clause that deals with a topic, such as notice periods or liquidated damages
- Comparing a submittal against the specification and listing apparent mismatches
- Grouping a thousand RFIs by theme to show where the design is weakest
- Drafting meeting minutes or a first cut of the monthly report narrative
The risk is that the output looks right and isn't. The Government of Canada's guide on generative AI warns that generated content can look credible while being wrong, incomplete or outdated, and tells public servants: "Don't consider generated content as authoritative." That's the right rule for owners too. A model that misreads a notice clause can cost far more than the hours it saved. Use these tools to find and sort, then have a person who knows the contract check anything that matters.
Use 3: Forecasting
Machine learning models can be trained on past projects to predict outcomes: likely final cost, likely completion date, which activities tend to slip. In principle, this is a powerful complement to the methods in our guide to cost forecasting.
In practice, most owners don't have enough clean, consistent history to train a model on. A retailer that has completed dozens of similar fit-outs with consistent cost codes and schedules might. A developer with six unique buildings over ten years doesn't. Vendor models trained on other companies' projects may not reflect your market, your contracts or your risks.
There's also a structural limit. RLB's Middle East project management director noted in 2023 that because AI learns from historical data, it struggles with sudden, unprecedented events, citing the COVID-19 pandemic. A forecasting model can tell you what usually happens. It won't tell you about the thing that's never happened before, which is often the thing that matters.
Use 4: Risk identification
AI can help draft a first risk register from the project description, find similar risks in past projects' registers and lessons learned, and flag early warning signs in the data: RFIs ageing on one trade, a supplier's submittals coming back rejected, float eroding on a near-critical path. That's useful, because risk registers often go stale and nobody reads the logs closely enough to see trends.
What it can't do is judge which risks your organization can accept, what they'd cost you, or who should carry them. That takes people who understand the contract, the market and the owner's appetite for risk. Our post on the construction risk register covers that side.
The limitations, plainly
The same RLB article lists the main limits, and they hold up:
- Lack of context. Tools miss site conditions, stakeholder politics and the history behind a decision.
- Data quality. Construction data is often unstructured, incomplete or inconsistent. Unreliable inputs give unreliable outputs.
- Cost and adoption. Tools take time, money and training, and people resist them.
- Dependence. Teams that rely on a tool need a plan for when it fails or gives a wrong answer.
- Judgment. Negotiation, conflict resolution and decisions under pressure still need people. The article calls AI a "co-pilot," which is about right.
View the data
| Use | Where it helps | Where it fails | Human check |
|---|---|---|---|
| Schedule checks | Fast, consistent rule checks every update | Can't judge whether the plan is realistic | Scheduler reviews flags and narrative |
| Document review | Finds, sorts and summarizes quickly | Confident but wrong readings of clauses | Contract expert checks anything relied on |
| Forecasting | Spots patterns in large, clean history | Needs data most owners don't have | Compare with bottom-up and index forecasts |
| Risk identification | Drafts registers, flags early warnings | Can't set appetite or allocate risk | Owner and team decide response and owner |
Data governance: the part owners can't skip
The biggest risk for owners isn't a bad forecast. It's project data going somewhere it shouldn't. Drawings, contracts, pricing, security layouts and tenant information are often confidential. The Government of Canada's guide notes that some AI suppliers may store what users enter or use it for training, and that personal information shouldn't go into public tools.
Ask every contractor, consultant and vendor on your project four questions:
- Which AI tools are you using on our project data, and for what?
- Where is the data stored and processed, and is it used to train anyone's model?
- Who checks the output before it reaches us or affects a decision?
- What happens to our data when the project ends?
For a framework, the U.S. National Institute of Standards and Technology released its voluntary AI Risk Management Framework 1.0 on January 26, 2023. It organizes AI risk work into four functions: Govern, Map, Measure and Manage. An owner doesn't need to adopt it formally, but it's a useful structure for a short internal policy.
Where to start: a small, measurable pilot
Owners don't need an enterprise AI strategy to get value. A better approach is a narrow pilot with a clear before-and-after measure. Two that work well:
- Automated schedule review on every update. Run the quality checks each month and track the number of flagged issues over time. If the contractor's schedules improve, the tool has paid for itself in avoided disputes.
- RFI and change log analysis. Have a tool group open RFIs and pending changes by trade, drawing area and age, and compare its output with what the project team already knew. If it surfaces patterns the team missed, expand it. If it doesn't, you've learned that cheaply.
Set the measure before the pilot starts, keep the data inside an approved environment, and decide in advance what result would justify going further. Vendors will happily run a pilot that proves nothing.
What owners should do
- Fix the data first. Require schedules and cost reports that pass basic quality checks before paying for analytics on top of them.
- Start with checks, not predictions. Automated schedule and document checks deliver value now. Predictive models need history most owners don't have.
- Keep a human on anything that matters. Contract interpretation, claims, forecasts to the board and risk decisions need a qualified person to review the output.
- Write AI and data terms into your contracts. Require disclosure of AI tools used on project data, set limits on where data goes, and get legal advice on the wording.
- Adopt a short internal policy using a framework such as NIST's, covering approved tools, confidential data and review requirements.
The bottom line
AI is already useful in project controls for checking, sorting and summarizing. It's far less proven for predicting, and it doesn't replace judgment. Owners who get value from it treat it as a fast, tireless junior analyst whose work always gets checked, and they protect their data before they hand it over.
Want a practical view of where AI could help your project controls, and where it couldn't? Book a 30-minute call with Benjamin R. Consulting.
Sources: For Construction Pros, SmartPM Technologies' New Report Reveals Gaps in Construction Scheduling and Data Use (2025); Deltek, DCMA 14-Point Assessment Metric Descriptions; Government of Canada, Guide on the use of generative artificial intelligence; RLB, How can AI enhance construction project management and what are its limitations? (2023); NIST, NIST Risk Management Framework Aims to Improve Trustworthiness of Artificial Intelligence (2023). The schedule check example is illustrative. This article is general information, not legal advice.