Construction & capital project schedule optimisation and delay-risk forecasting

Large construction and infrastructure projects are planned once, over years, with thousands of logic-linked activities, shared crews and equipment and contractual milestones. Planners typically build one baseline by hand and discover slippage late. Two AI approaches address this: generative scheduling (a solver explores sequencing, crew and equipment alternatives from the model and constraints) and ML delay-risk forecasting (models trained on past programmes predict which activities will slip and by how much). Together they support bids, recovery planning and portfolio risk.

Problem & value

• Today: schedules are authored manually in Primavera P6, Microsoft Project or Asta Powerproject; alternatives are explored by experience, risk is handled by a small number of workshop-based estimates or Monte Carlo on expert-assigned durations. Recovery after a delay is mostly manual re-planning.
• AI value: (1) explore many more sequencing/resourcing scenarios against time, cost and resource levelling; (2) replace subjective duration uncertainty with data-driven distributions from comparable historical programmes; (3) re-plan quickly after changes; (4) give owners an independent view of the contractor's programme quality.
• Evidence: vendors publish deployments rather than audited benchmarks. ALICE Technologies cites contractors such as NGE and Zachry; nPlan cites programmes such as HS2 and Network Rail [1][2]. Treat vendor-reported time/cost savings as claims to verify in a pilot on your own projects.

Variants

• Generative scheduling (contractor): build-sequence and crew/equipment optimisation from BIM and P6.
• Delay-risk forecasting (owner, lender, PMO): P50/P80 completion dates and drivers across a programme or portfolio.
• Bid-stage risk: risk-adjusted tender programme and contingency sizing.
• Rail/road/utility programmes: linear projects with possession windows and track access constraints.
• Residential/repetitive building: takt planning and flow-line scheduling.
• Recovery planning: re-sequencing and acceleration options after delay events.
• Delay claims support: forensic analysis of programme impact (human expert signs off).

How it works

1. P6 / MS Project export + BIM (IFC) + resource and calendar data
2. data cleaning (logic checks, open ends, constraints, duplicate/merged activities)
3. activity-to-element mapping (BIM quantities
4. durations via productivity rates)
5. (a) optimisation: CP-SAT / MILP / metaheuristic over RCPSP with crews, equipment, zones
6. (b) risk model: ML on historical programmes
7. duration/delay distributions per activity
8. Monte Carlo schedule simulation (network logic + learned uncertainty)
9. P50/P80, criticality
10. scenario comparison and recommendations (planner reviews, adjusts constraints)
11. approved schedule written back to P6 / MS Project; updated with weekly progress

• Optimisation core: a resource-constrained project scheduling problem (RCPSP) with multi-mode options (crew sizes, methods), precedence logic, zones and calendars. Exact and constraint-programming solvers handle mid-size networks; large networks need decomposition or heuristics. ML supplies productivity and duration estimates; the solver provides the guarantees.
• Delay-risk model: gradient boosting or neural models on activity features (type, trade, season, predecessor structure, planned duration, project context) trained on historical baselines vs. as-built dates. Output is a distribution, not a point value, and feeds Monte Carlo simulation.
• LLM layer: natural-language what-if questions, schedule-quality review, translating contract clauses into constraints (see OPT-018). The LLM never edits the logic without planner approval.
• Human-in-the-loop: the planner owns the schedule; AI proposals are options. Contract-relevant baselines and delay claims need sign-off by a qualified scheduler.

Data preparation

Data needed: baseline and updated schedules (Primavera P6 XER, MS Project, Excel), bills of quantities, resource and cost data, daily site reports, progress photos, weather records and lessons learned from past projects. In an SMB contractor or public owner these sit in scheduling tools, the ERP or construction software, file shares, e-mail attachments and paper site diaries.

Typical gaps: schedules without logic links or consistent activity codes, progress reported as free text, scanned site diaries, duplicated project versions, no structured history of delays and their causes, and restricted access to subcontractor data.

Preparation steps: export schedules to a structured format and check logic quality (open ends, constraints, negative float); harmonise activity codes and a work-breakdown standard; digitise site diaries with OCR and extract delay causes; build a reference table of finished projects with planned versus actual durations; agree access rights with subcontractors. The reference history is the most valuable asset and also the most time-consuming.

Expect a high share of effort, about 40-50 percent. Helpful: PREP-006 for scanned diaries, PREP-011 for spreadsheet clean-up, PREP-019 for delay-cause categories and DOC-018 for document extraction. See GUIDE-003.

Implementation paths

Which path for whom

• enterprise (A): Contractors and owners already use Primavera, MS Project or Asta; established add-ons and risk-analysis products integrate into these tools and are the fastest route.
• developer (B): Teams building their own scheduling intelligence combine constraint solvers such as OR-Tools with cloud LLM or ML APIs for duration forecasting and schedule parsing.
• public (B): Public owners need transparent, reproducible results and procurement-compliant hosting; a build on EU cloud with open solvers avoids lock-in to a single scheduling vendor.

Path A — Buy (recommended for most contractors/owners)

• Generative scheduling: ALICE Technologies (US; P6/MS Project import, scenario "what-if" engine).
• Delay-risk / schedule analytics: nPlan (UK; ML forecasts from a large library of historical programmes, P6 integration). Other schedule-analytics tools: Oracle Primavera Cloud with risk analysis, Bentley Synchro 4D, Autodesk Construction Cloud, Hexagon/Trimble Connect integrations, Safran Risk, Deltek Acumen (schedule quality and risk), Oracle Primavera P6 Risk Analysis.
• EU/UK options: nPlan (UK), Powerproject (Elecosoft, Sweden-based group), Nemetschek (Germany) 4D/BIM tooling, Dalux and Planradar (EU site-progress tools feeding actuals).
• Choose A when: you need a proven integration with P6/BIM within weeks, you lack ML/OR staff, and the vendor's data-use terms (e.g. use of your programmes to train shared models) are acceptable.

Path B — Build on cloud APIs

• Solver: Google OR-Tools CP-SAT (open source, runs anywhere) or commercial Gurobi / FICO Xpress / Hexaly (France) / CPLEX for large RCPSP models [3].
• ML: managed AutoML/ML platforms in EU regions (Azure ML, Vertex AI, SageMaker Frankfurt).
• LLM copilot: frontier models (Claude, GPT, Gemini, Mistral) with tool calls into the solver and the schedule database; XER parsing with open libraries (e.g. PyP6Xer) or P6 Web Services/REST API.
• Choose B when: you want custom constraints (zones, takt, possession windows) with moderate data science effort.

Path C — Self-host (open source)

• Stack: OR-Tools CP-SAT, scikit-learn/LightGBM/XGBoost, a Monte Carlo engine (NumPy), IfcOpenShell for IFC quantities, PyP6Xer for XER, PostgreSQL; optional open-weight LLM (Qwen, Mistral, Llama) for the copilot.
• Hardware: HW1 (CPU server) suffices for solver and tabular ML; HW2 only if the LLM copilot is local.
• Ops effort: moderate; main cost is data engineering and domain modelling, not compute.

Path D — Custom training

Worth it for owners or large contractors with years of programmes (hundreds of projects with baselines and as-built dates). Requires harmonised activity taxonomies, as-built records and metadata on project type, location and contract form. Without this history, use vendor models or priors plus your own calibration.

State of the art (2026)

• Delay-risk forecasting from historical programme libraries is productised and in use on major public infrastructure programmes [2]. Accuracy depends on how similar your projects are to the training data.
• Generative/automated scheduling is production-grade at early adopters, but adoption is limited by BIM quality, missing activity-to-element links and planners' trust; most schedules are still hand-built [1].
• Solver + learned durations hybrids and LLM copilots for what-if analysis and schedule review are the current trend.
• Limits and failure modes: garbage-in schedules (open ends, constraint abuse, lag misuse) produce confident but wrong output; historical data may embed past bias (padded or optimistic baselines); optimisation ignores unmodelled site realities (access, permits, supply delays, labour availability); ML cannot foresee scope change or contractual/political events; shared-model vendors may be opaque.

Shared resources & combinations

• Same stack (optimization-solver + tabular-ml): OPT-002 production scheduling, OPT-019 field-service dispatch, OPT-003 workforce scheduling, OPT-008 supply-chain planning share solver skills and infrastructure.
• PRED-023 (project schedule and cost-overrun risk forecasting) is the closest sibling; cost-side forecasts and this delay-risk module should share data.
• Inputs from: DOC-030 (quantity takeoff from plans), VIS-040 (scan-to-BIM), ROBO-009 (automated progress monitoring providing actuals for re-planning).
• OPT-018 (natural-language optimisation modelling) supplies the LLM-to-solver copilot pattern.

Effort, cost & KPIs

• Path A: weeks to a few months for pilot on 1–3 projects; per-project or enterprise licences.
• Path B/C: 4–9 months for a production-grade pipeline incl. data cleaning and P6 integration; team of scheduler + OR/data scientist + engineer.
• Cost drivers: schedule data quality, BIM-to-activity linking, licences, change management with planners.
• KPIs: schedule accuracy (forecast vs. actual completion, calibration of P50/P80), planner hours per re-plan, number of scenarios assessed, delay days avoided or recovered, contingency released/consumed, schedule-quality score (open ends, hard constraints), bid hit rate on risk-adjusted programmes.

People & change

Management and the project or construction manager own the decision and the schedule baseline. Planners and site managers are the end users and hold the tacit knowledge, so involve them from the first workshop and let them judge the quality of suggestions. IT covers integration with scheduling and construction software; subcontractors and suppliers must agree to data sharing. Legal and compliance check contract consequences, for example delay claims or milestone obligations, and the data owners of schedule archives must release the history.

Works council: depends. A pure planning optimiser does not monitor staff, but if site progress is tracked per crew or via cameras and wearables, co-determination under §87(1) no. 6 BetrVG applies; the works council should in any case be informed in good time about AI in work processes (§90 BetrVG). Public clients involve the staff council (Personalrat).

Data protection: involve the DPO if site photos, access data or worker productivity are used; a DPIA is likely for person-related monitoring. Pilot on one running or recently finished project, compare AI forecasts with actuals, and address the common concern that the tool replaces planners by presenting it as a what-if aid with the planner deciding. Offer AI literacy sessions for planners. See GUIDE-009, GUIDE-008 and GUIDE-004.

Risks, regulation & compliance

• EU AI Act: minimal risk in typical use (planning tool, no Annex III category). Caveats: use of the same tools to allocate tasks, monitor or evaluate individual workers on the basis of behaviour or personal traits is an Annex III employment use and can be high-risk; keep resource optimisation at crew/trade level. AI literacy duty (Art. 4) applies [4].
• GDPR: schedules are largely non-personal. Personal data arises with named workers, subcontractor individuals, GPS/site-access logs or progress photos; minimise, pseudonymise, and involve works councils where individual-level data is used (Germany: BetrVG §87).
• Contract and liability: AI schedules are not contractual until accepted; delay claims (FIDIC, VOB/B, NEC) rely on recognised methods (e.g. SCL Protocol, AACE guidance) and expert judgement, not black-box outputs.
• Commercial confidentiality: programme data is sensitive; check vendor terms on data reuse and model training, hosting region and exit.
• Mitigations: schedule-quality checks before use, calibration tests on past projects, versioned scenarios, audit trail of changes, planner approval.

Getting started

1. Export 5–10 completed projects (baseline + updates + as-built) from P6 and assess data quality.
2. Back-test a delay-risk model or vendor tool: does the P80 date cover actual finish dates?
3. Pilot scenario generation on one upcoming project phase with a planner; compare with the manual plan.
4. Define governance: who approves AI schedules, data-use terms, contractual status.
5. Link progress actuals (ROBO-009, site apps) for weekly re-forecasting.

Sources

1. ALICE Technologies — generative construction scheduling — https://www.alicetechnologies.com/
2. nPlan — schedule delay risk forecasting — https://www.nplan.io/
3. Google OR-Tools — CP-SAT scheduling guide — https://developers.google.com/optimization/scheduling
4. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
5. Regulation (EU) 2016/679 (GDPR) — https://eur-lex.europa.eu/eli/reg/2016/679/oj
6. Oracle Primavera P6 — https://www.oracle.com/construction-engineering/primavera-p6/
7. IfcOpenShell — https://ifcopenshell.org/
8. Society of Construction Law — Delay and Disruption Protocol — https://www.scl.org.uk/resources/delay-disruption-protocol
9. ALICE Technologies — NGE expands use of ALICE Core (Jan 2025) — https://blog.alicetechnologies.com/news/nge-expands-ai-driven-success-in-construction-planning-and-scheduling-with-alice
10. Construction Dive — Zachry Construction implements ALICE (press release, 21 Jan 2025) — https://www.constructiondive.com/press-release/20250121-zachry-construction-corporation-implements-alice-to-drive-innovation-in-hea/