Cost data for a building does not live anywhere. It lives in an IFC file, and a spreadsheet, and a PDF quote from a subcontractor, and someone's handwritten note, and a paragraph of contract text. Different units, different resolutions, different levels of confidence.
An estimator reconciles all of that by hand and then produces a single number. That number then travels through the business as though it were a fact.
The research team quantified both halves of that problem before building anything, which is the part I respect most. Interviews with twelve cost engineers and analysis of three real projects put manual reconciliation at an average of 14.3 hours per project, and cost overruns at an average of 15.8% against the initial deterministic estimate.
One agent per job
Rather than one model doing everything, the system assigns each part of the estimating job to an agent, coordinated by an orchestrator.
- 1
Orchestrator
Decomposes the task, allocates work, routes exceptions.
- 2
Three ingestion agents
One each for text, tabular and geometric sources. The formats never meet until they have been read.
- 3
Integration and alignment
Entity resolution and semantic conflicts. Metres against feet. Concrete grade 30 against C30.
- 4
Uncertainty quantification
Finds missing fields, infers distribution types (triangular, beta-PERT, lognormal) and flags its own assumptions.
- 5
Monte Carlo
Ten thousand iterations produce a cost distribution rather than a point.
Agents communicate over a publish-subscribe bus with shared memory that persists between runs. The stack is Python 3.11, AutoGen 0.2 for the agent roles, LangChain calling GPT-4 Turbo, IfcOpenShell for the IFC, pandas with pdfplumber and Tesseract for the documents, and NumPy with SciPy for the simulation.
Two temperature settings, and this is a nice detail: 0.2 for parsing, where you want determinism, and 0.7 for uncertainty reasoning, where you want the model to consider alternatives.
How it was measured
Three real buildings, a concert hall, an exhibition veranda and a teaching building, with between 98 and 386 cost elements each. Compared against three baselines: deterministic BIM using Revit and CostX, manual probabilistic estimation in @RISK, and a non-agentic single-prompt GPT-4 Turbo. Three runs each, paired t-tests, five-fold cross-validation, and sensitivity checks on temperature and sample count.
That last sentence is why this paper is worth your time. Most AI-in-construction papers compare against nothing.
Note where the single-prompt LLM lands. Bottom. Handing the whole problem to one model in one call performed worse than a spreadsheet-based deterministic method. The agent architecture is doing the work, not the language model.
The finding that matters most
Accuracy is the headline, but calibration is the real result. The framework achieved 86% coverage on nominal 90% prediction intervals, against 74% for manual estimation and 62% for the LLM-only approach, with intervals 38 to 46% narrower.
Coverage is the question of whether your confidence interval means what it says. An estimate that claims 90% confidence and is right 62% of the time is worse than useless, because people make decisions on it. The LLM-only baseline was confidently wrong. The agent system was closer to honest.
A narrower interval is only an improvement if the interval was telling the truth in the first place.
The ablation isolates why. Removing the integration and alignment agent damaged entity resolution most, which ran at 86.5%. Removing the uncertainty quantification agent damaged calibration most. Those are two different failure modes and the architecture separates them cleanly.
What a developer would actually do with this
- Evaluate more options. At 4.2 minutes an estimate you can price six design variants in the time one used to take, which changes what gets considered rather than just what gets costed.
- Put contingency where the risk is. Tornado diagrams from the simulation show which inputs drive the spread, so contingency stops being a flat percentage applied out of habit.
- Defend the number. Agent logs give an audit trail, which matters when a board asks why the estimate moved.
- Improve with repetition. Shared memory persists across runs, so a firm estimating similar building types repeatedly gets compounding benefit.
The caveats worth naming out loud
Three projects, all in Hong Kong and mainland China. Irish and UK cost structures, labour rates, procurement routes and contract forms are different enough that the MAPE figure will not transfer unchanged. It runs in batch rather than real time. And it depends on GPT-4, which is a genuine data privacy problem for anything commercially sensitive, and a reason to look hard at the small-model argument in the compliance research.