Project controls AI: turn schedule, cost and field data into checked decisions

Project controls AI can help organise evidence, identify exceptions, explore programme options and draft explanations. Its value depends on the next decision: resolving a constraint, checking a forecast movement or finding an unsupported progress claim. A polished narrative that cannot be traced to a source adds another review task.

Begin with a question such as: which activities need the planner’s attention this week, and why? Set the data date and define what counts as a useful answer. The tool needs to show the relevant activity or cost code, the change since the previous accepted position, its evidence and the uncertainty. Keep different interpretations visible until the responsible person resolves them.

Treat these as proposed evaluation tasks. Ordinary calculations and data validation often belong in controlled spreadsheets or the project system; a language model can help explain the checked results. Do not replace a reproducible calculation with an answer that merely sounds numerate.

For a weekly test, retain the accepted baseline, the current programme export, the previous reporting period and a change log. Include activity IDs, work breakdown codes, calendars, relationships and remaining durations where your analysis needs them. For costs, define whether the file contains actuals, commitments, accruals or forecasts. Label the currency, units and reporting cut-off.

Map field locations and package names to the same structure. ‘North plantroom’, ‘NPR’ and ‘Area 04’ may describe the same place, but the model should not silently decide that they do. Keep a reviewed mapping table. Flag unmatched records and duplicate IDs before importing data, and retain the raw source alongside the cleaned version.

Include difficult records in the test: a revised activity description, a missing calendar, conflicting progress entries and a late invoice. Decide the expected handling of each. A reliable workflow should report the gap or conflict rather than merge it into a confident summary. RICS recommends planning the data, integration and supervision needed to make a pilot routine. RICS AI in commercial property and construction report 2026.

Consider a fictional package of 100 identical supports. At Friday’s cut-off, the site record shows 40 installed, while the inspection register shows 32 accepted. The agreed plan called for 50 installed by that date. On this simple count basis, installation progress is 40% against 50% planned, a gap of 10 percentage points. Acceptance progress is separately 32%.

These are illustrative quantities, not earned-value measures or a forecast of completion. The figures do not establish productivity, critical-path delay or payment entitlement. Those questions need the relevant hours, schedule logic, measurement rules and commercial review. An AI summary should preserve the distinction and ask why eight installed supports remain unaccepted.

Schedule prediction and schedule optioneering answer different questions. nPlan describes models that predict activity performance and methods that rank activities for attention. ALICE describes comparing planning, sequencing and resource options. The first category can inform where to investigate uncertainty; the second can help explore alternatives under modelled constraints. Neither removes the need to validate the programme. nPlan’s description of its forecasting methods; ALICE construction planning and scheduling.

For progress evidence, OpenSpace describes combining captured imagery, progress analysis and human review, with connections to common schedule formats. For retrieving information and drafting work within a project platform, Procore documents source citations and a review step before proposed changes. These are vendor-described capabilities. We have not independently tested these products or verified results on your project. OpenSpace Track product documentation; Procore AI product documentation.

For cost control, first ask whether a candidate can reconcile your actual cost structure and retain links to the source transactions. A product that produces a persuasive variance explanation but cannot distinguish a timing difference from a scope change is not ready for the cost meeting. Require the same input pack and acceptance criteria across vendors.

Keep an unedited copy of each generated output and record what the reviewer changed. This lets the team distinguish a data problem from a model problem and see whether the workload is improving over time. Use the following checks before a draft becomes a reported position.

Run the assisted workflow alongside the normal process for two reporting periods. Measure preparation plus review time, unresolved data exceptions, unsupported statements and how quickly an agreed action reaches its owner. A faster first draft is a weak result if the planner spends longer correcting it. Stop and fix recurring failures before expanding the trial.

Name a controls lead to accept the outcome and a data owner to maintain inputs. Agree what happens when the service is unavailable or an output fails review. NIST frames AI risk management as part of ongoing design, use and evaluation; our practical recommendation is to keep a small test pack and rerun it after material product or workflow changes. NIST AI Risk Management Framework.

The Hardhat Toolbox provides our own readiness assessments and structured analysis utilities. Check each tool’s stated scope: a rule-based check is not a predictive AI forecast, and no generated narrative should be treated as an approved project position. Start with the diagnostic that matches the decision you are trying to improve.