Agentic AI Platforms for Construction Planning and Schedule Simulation

Summary

We build agentic planning assistants that sit between a design model and a construction schedule. A system of this kind reads a building model, derives the activities implied by its elements, attaches quantities and productivity rates, builds or amends a programme, and answers questions about what happens when labour, budget, material availability or sequencing constraints change. The useful output is not prose. It is a schedule an experienced planner can open in their own tool, together with the assumptions that produced it.

The Challenge

Model data is rarely clean enough to trust on arrival: authoring tools export IFC inconsistently, base quantity property sets are missing or malformed as often as not, and an element that carries no classification cannot be mapped to an activity at all. Scheduling itself is the second trap: a language model asked to emit dates will produce plausible ones that no critical path calculation supports, and durations drift against real working days as soon as calendar definitions are ignored. Buildability is the third: an optimiser that does not treat curing periods, crane reach, floor handover sequence and site access windows as hard inputs will propose a programme no site team can execute. Planners will not abandon the tools they already use, so anything that loses baselines, calendars, resource assignments or activity codes on export is dead on arrival. And because commercial rates and float are sensitive, advice with no audit trail is not reviewable or reversible.

The Solution

Reading the model: takeoff and activity mapping

Ingestion normally works from IFC rather than a proprietary format, so the pipeline handles both IFC2x3 and IFC4.

• Elements are traversed by type, and quantities come from base quantity property sets where the exporter wrote them, or from geometry evaluation where it did not.
• Elements are classified against a work breakdown structure, usually mediated by a classification standard such as Uniclass, OmniClass or MasterFormat.
• Activities are then attached with crew compositions and output rates, and everything the model failed to supply appears in an explicit gap report rather than being silently defaulted.

The scheduling engine

Scheduling logic belongs in a deterministic engine. That means a real critical path implementation with forward and backward passes, finish-to-start, start-to-start and finish-to-finish relationships, lags, calendars and shift patterns, resource assignment and levelling. Scenario work adds constrained optimisation, typically a mixed integer or constraint programming formulation for sequencing under crew and equipment limits, and probabilistic analysis where duration ranges are simulated to show a distribution of completion dates rather than a single figure.

Agent architecture and grounding

The agent layer orchestrates tools: model query, quantity takeoff, rate lookup, schedule computation, scenario comparison and report generation. Each tool has a typed contract, and the model chooses and parameterises calls rather than inventing results. Retrieval over specifications, method statements and historical productivity data grounds the reasoning, and every recommendation carries provenance back to the element identifiers and rates that produced it. An evaluation harness built on held-out data is essential, because plausible-sounding schedule advice is easy to generate and hard to falsify by reading it.

Integrations, telemetry and exports

Round-tripping matters. Export to XER and to Microsoft Project XML preserves calendars, resource assignments, activity codes and user-defined fields, and the safe pattern is to write changes as a new revision rather than overwrite a baseline. Where site telemetry is in scope, sensor and equipment data arrives over MQTT or a vendor API, lands in object storage, and updates actual start and finish dates and productivity estimates.

What This Delivers

Planners can ask what happens if a crew is reduced, a material slips or a sequence is reversed, and get an answer computed by a scheduler, not asserted by a model. Commercial teams see which assumptions and rates drove a date, and site teams receive programmes that respect how the building actually goes together. Because every change is versioned, attributed and traceable back to model elements, the output can be reviewed and reversed rather than taken on faith.

Technologies and Tools

• IFC2x3 and IFC4 ingestion with geometry-based quantity evaluation
• Uniclass, OmniClass or MasterFormat classification mapped to a work breakdown structure
• A deterministic critical path engine with calendars, lags, resource levelling and probabilistic duration simulation
• Mixed integer and constraint programming solvers for constrained sequencing
• Tool-calling agent orchestration with typed contracts and grounded retrieval
• XER and Microsoft Project XML export, MQTT and vendor APIs, object storage