MeltPlan

The repeated positioning as 'AI-Native Planning Software' (rather than 'AI-powered' or 'AI-enabled') implies AI is intended as the fundamental control plane for planning, not an add-on — meaning the data model, UX, and core scheduling algorithms are likely designed around learned models rather than traditional deterministic schedulers.

Given the construction context (AEC) the product almost certainly needs to fuse multimodal inputs — BIM/IFC models, schedules (MS Project/Primavera), drawings, specs, site photos, IoT/sensor feeds, and human text — implying a nontrivial data normalization/ontology layer that maps 3D elements to tasks and resources.

A practical way to be 'AI-native' for planning is to represent plans as graphs (tasks, dependencies, resources, spatial adjacencies). This suggests use of graph-based ML (GNNs) or graph-structured optimization combined with learned priors to generate and revise schedules — an architecture that blends symbolic constraint solving with learned heuristics.

To move beyond static schedules they must be building a continuous, closed-loop system: ingest live site state (progress photos, sensor telemetry, crew reports), perform probabilistic forecasting of task durations, and replan — effectively a real-time digital twin for sequencing. That closed-loop real-time replanning is technically much harder than batch optimization and is a distinctive implementation signal if true.

Hidden complexity they must solve but do not advertise: robustly parsing IFC/BIM at scale (different levels of detail/degeneracies), canonicalizing vendor-specific schedule formats, mapping ambiguous text in RFPs/specs to constraints, and building reliable uncertainty models for durations/resource availability in noisy field environments.