ConTechLabs explores AI-native construction software and agent workflows for architecture, engineering, and construction (AEC) teams. I run it as a venture under RMG Associates. The scope runs from the first feasibility question to the building in operation.

What it is

The lab explores, tests, and builds AI-native applications that understand real construction processes, workflows, and decisions. Today the work is research. There are six experiments, starting with a working construction ontology, plus a software track called the Software Factory. The agent workflows shown on the site are proposed. They are not live agent runs.

Why it exists

On most projects, the problem is not a lack of data. Teams have drawings, specs, schedules, submittals, RFIs, field reports, and email. But these records sit in different systems, and the links between them often live only in people's heads. A changed equipment selection can affect the roof curb, structural supports, electrical rough-in, controls, and a commissioning test. Someone has to notice in time. That is our starting point. AI needs more than data. It needs context.

Twelve phases, one lifecycle

The home page follows the same 12 phases as the construction ontology:

  • Strategy & need
  • Site & feasibility
  • Team, funding & delivery
  • Design & engineering
  • Approvals, rights & permits
  • Estimate, tender & award
  • Procure, fabricate & mobilize
  • Construct & coordinate
  • Inspect, test & commission
  • Handover & commercial closeout
  • Operate, maintain & learn
  • Adapt, renew & recover

At each phase the lab maps the requirements, resources, responsibilities, and dependencies, then asks where AI could help. Work can overlap and iterate; the phases do not prescribe a contractual sequence.

The construction ontology

The starting point is a construction ontology. It is a structured map of the things, processes, and relationships that make a project work. It covers six groups:

  • The physical: sites, spaces, systems, assets, and materials
  • The work: activities, workflows, dependencies, and milestones
  • The people: teams, responsibilities, authority, and decisions
  • The resources: labor, equipment, suppliers, capacity, and logistics
  • The obligations: requirements, contracts, approvals, and commitments
  • The evidence: documents, observations, costs, dates, and revisions

Every connection should keep its source and revision. A link between an air handler and its submittal is only useful if you can see which document says so, which revision, and whether anyone reviewed it. The model is at version 0.2. It is aimed mainly at US projects and is open for practitioner review. It references standards such as IFC, ISO 19650, and COBie without claiming conformance.

Valid and early signal

The lab's main test for any AI finding is whether it gives valid and early signal. Valid means the finding rests on current, identifiable evidence about the affected work. It keeps confirmed facts, reported information, and AI inferences apart. Early means it arrives before a decision or commitment becomes hard or costly to change, with time for the team to respond. A useful finding says what changed, what supports it, what is still uncertain, what work is affected, and when a response is needed.

The experiments apply this to specific problems: tracing blockers to their causes, understanding and shortening lead times, finding avoidable cost and rework, keeping project communication connected, and capturing field knowledge before it is lost. Each says how value would be judged, including unsupported alerts and missed concerns. No time or cost improvement is assumed from adding AI.

People stay in charge

Agents can find the relevant records, do the groundwork, and prepare a draft or proposed action. Architects, engineers, and construction managers keep responsibility for technical review, approvals, and issued direction. An agent carries out only permitted actions. Where the workflow requires it, the agent gets the responsible person's approval first. A generated answer, an executed action, and a human approval are different events, and the software should show each one.

How to take part

The lab asks for practitioner input. Design, engineering, construction, and operations professionals can join the ConTechLabs Slack workspace to review the ontology and experiments, challenge assumptions, and bring project examples. The working model of the ontology can be downloaded as JSON from the ontology page.