Construction sequencing
Compare installation sequences against access, predecessor work, crew availability, and equipment requirements.
Experiment 06 / Research concept
Capture field knowledge before it is lost
Research conceptConstructability / Field experience / Operations / Learning
Use experience from completed work to improve the next decision.
Project teams continually develop practical knowledge: why an installation sequence worked, what caused repeated review comments, which procurement assumption failed, or what made a completed system difficult to maintain.
That knowledge may remain with the people involved or become separated from the evidence and conditions that explain it. Another team can encounter the same issue without knowing that someone has already investigated it.
We’re exploring how AI agents could capture experience as work proceeds, connect it to project records, and bring supported findings into subsequent decisions. Each application would generate further evidence about where the lesson works, where it needs adjustment, and where it should no longer be used.
Before teams commit to the next design, purchase, or installation, agents could also use that experience to prepare and test alternative approaches through early simulations.
The questionCan AI use project experience to anticipate problems, test better ways to perform the work, and improve subsequent decisions through measured results?
A principle across the lab
Valid and early signal.
Valid: establish the basis.
Use identifiable, current evidence relevant to the affected work. Keep confirmed facts, reported information, and AI inferences distinguishable. Identify missing or conflicting information.
Early: preserve time to act.
Identify a concern before a decision or commitment becomes difficult or costly to change, with enough time for the team to investigate and respond.
Prepare a useful response.
Explain what changed, the supporting evidence, what remains uncertain, the affected work, and when a response is needed to preserve the available options.
For this experiment, bring relevant experience into the next design review, purchase, fabrication release, installation, or maintenance decision while the team still has time to evaluate it. Explain both the supporting evidence and the differences that could limit its use.
01 / A working learning loop
Let each result inform the next attempt.
AI would perform much of the capture, comparison, review preparation, and follow-through. Practitioners would judge applicability and authorize changes. The measured result would inform the next proposal.
Capture experience
Record what happened and retain its supporting evidence.
Establish applicability
Compare the earlier conditions with the current work.
Simulate alternatives
Test proposed approaches using appropriate models and tools.
Review and authorize
Practitioners evaluate the results and decide what can proceed.
Execute and measure
Record the actual work, conditions, and outcomes.
Refine the knowledge and analysis
Use the result to update the lesson, assumptions, and subsequent review.
02 / The experiment
What we would test
Capture experience while the context is available.
Identify opportunities for learning from design reviews, resolved RFIs, first installations, mock-ups, inspections, procurement outcomes, commissioning, and maintenance work.
Agents assemble the relevant records and prepare a short account of what was expected, what occurred, what changed, and the observed result. A targeted question to the people involved could capture reasoning that the documents do not explain.
Capture successful methods, unsuccessful attempts, and unresolved outcomes. Do this throughout delivery, while participants can still explain the circumstances.
Establish what the evidence supports.
Connect each proposed lesson to drawings, photographs, measurements, test results, production records, correspondence, and practitioner observations.
Distinguish the observed result from the explanation offered for it. A faster installation following a sequence change may also reflect different access, crew experience, or site conditions.
Agents identify missing evidence, competing explanations, and questions for review. The responsible practitioners assess what can reasonably be concluded.
Record where the lesson applies.
Describe the conditions that matter: building or infrastructure type, system, materials, equipment, location, design requirements, installation method, access, and operating conditions.
Retain the source revisions, reviewer, supporting evidence, known exceptions, and unresolved questions.
The ontology would connect that experience to the relevant assets, work packages, requirements, and decisions. This gives agents a basis for comparing situations beyond matching similar words in documents.
Bring experience forward before the next commitment.
When a team approaches a related design review, purchase, fabrication release, installation, or maintenance task, agents look for applicable prior experience.
They explain the connection, identify important differences, and prepare a specific recommendation or question.
A finding might help a designer check maintenance access before issuing a package, a procurement team clarify a supplier commitment before ordering, or a superintendent assess a proven sequence before mobilizing crews.
Each finding should arrive while the team still has time to investigate and respond.
Apply, observe, and refine.
Practitioners assess whether the earlier approach fits the current work and authorize any necessary changes.
Agents prepare the proposed method, expected benefit, assumptions, and evidence needed to assess the result. Depending on the work, that evidence could come from a calculation, simulation, mock-up, first installation, inspection, commissioning test, or operating record.
After application, agents compare the expected and observed outcomes. Did the change help? What additional work did it create? Which conditions explain the result?
The lesson is then strengthened, narrowed, revised, or withdrawn. An unsuccessful application becomes useful evidence for the next review.
Improve the AI’s own work.
The same feedback loop should apply to the agents.
Record which findings were useful, which arrived too late, which were dismissed as inapplicable, and which consequential issues the system missed. Use that feedback to propose improvements to retrieval, applicability checks, review questions, and response preparation.
Test proposed changes against retained project cases, including cases where a lesson should not apply. Preserve the previous system version so changes can be compared and reversed if performance deteriorates.
For this experiment, recursive learning means that the result of one application informs the next attempt, with each revision checked against project evidence. The initial implementation could use reviewed project memory and versioned workflows. Model retraining would be a separately evaluated option.
03 / Test before committing
Use prior experience to test the next plan.
Before a design is released, an order is placed, or installation begins, agents could use relevant project experience to prepare scenarios for evaluation.
Fabrication and installation
Check proposed assemblies against dimensions, tolerances, delivery routes, lifting arrangements, and connection access.
Procurement
Examine how alternative selections, review durations, supplier commitments, and delivery dates affect planned construction.
Commissioning and operations
Test whether proposed arrangements support the required operating modes, maintenance access, and replacement procedures.
Agents would identify applicable lessons, assemble the current project information, and prepare alternative scenarios. Appropriate engineering, scheduling, or simulation tools would perform the analysis. Practitioners would assess the assumptions, results, and proposed changes.
Where information is incomplete, the analysis could identify which measurement, supplier confirmation, mock-up, or field check would most help resolve the decision.
Compare predictions with actual results.
After the selected approach is used, agents would compare the predicted outcome with installation records, inspections, tests, and operating evidence.
That comparison would improve both the lesson and the assumptions used in later simulations. For example, did the revised sequence preserve the expected installation access? Did the assumed production rate hold? Which site conditions explain the differences?
Each record would distinguish a simulated result, a field-tested approach, and an outcome demonstrated on completed work.
04 / An illustrative scenario
An installation lesson could help the next floor.
During the first corridor installation, the trade teams find that installing the main duct first restricts access needed to complete nearby pipe connections.
They test a revised sequence in the next comparable area. Installation records and crew observations indicate an improvement, and the team documents the access conditions and remaining limitations.
Could the next team test a better sequence before work starts?
Before overhead work begins on another floor, agents identify the similar arrangement and retrieve the earlier installation records.
They compare the current drawings, access, ceiling heights, and fabrication status with the earlier work. Agents prepare alternative duct-and-pipe installation sequences for evaluation using the current layout and schedule. The superintendent and trade foremen receive the proposed sequence, its predicted access and schedule consequences, the supporting evidence, and the differences that need review.
After the accepted sequence is used, agents compare its actual performance with the prediction. If a different layout reduces its usefulness, that limitation becomes part of the lesson and informs later simulations.
The knowledge improves through application and review, and remains connected to the conditions that explain it.
05 / Put the learning to use
Carry learning into standard practice.
Where repeated, comparable evidence supports an improvement, agents could prepare proposed updates to design checklists, standard details, procurement questions, installation plans, commissioning procedures, or training.
The responsible organization reviews and issues those changes. Project-specific requirements, contracts, and technical approvals continue to govern the work.
Knowledge reuse also follows ownership and access permissions. Experience from one project should be shared across projects only where that use is permitted.
The supporting record would retain the applicable conditions, source evidence, review history, and any later findings that narrow or overturn the lesson.
06 / Evidence of value
What would establish value
Did experience from earlier work improve the next decision, and did the new result make the knowledge more useful?
Assess whether early simulations identified a practical improvement before commitment, and compare their predictions with actual results.
- Less effort required to capture and retrieve useful experience.
- Relevant findings reaching teams before affected decisions or work.
- Fewer repeated problems and better use of successful methods.
- Clear evidence about where an earlier solution applies and where it does not.
- Documented differences between predicted and observed results, with revisions to the underlying assumptions.
- Improvements that persist in reviewed procedures, details, and training.
- Better agent performance without losing previously demonstrated capability.
- A record of inappropriate recommendations, missed opportunities, and lessons revised or withdrawn.