Run a clash detection on a federated model of a real project and the software returns a number that stops the room: forty thousand clashes. Everyone knows the number is meaningless, because most of those clashes are duplicates, irrelevant overlaps, or geometry that will resolve itself, and the coordinator's real job is not finding clashes, the software does that, but figuring out which of the forty thousand actually matter and in what order to solve them. This is exactly the work AI is suited to, triaging a huge list down to the few hundred real, consequential clashes and grouping them by system, by area, and by which trades need to be in the room. But it is also work where AI's judgment about what matters is a proposal, not a ruling, because the priority of a clash depends on cost, sequence, and stakeholder consequence that the coordinator understands and the model does not. This lesson shows you how to use AI to turn forty thousand clashes into a workable coordination agenda while keeping the priority judgment where it belongs.

The Forty-Thousand-Clash Problem

The clash detection report is the canonical example of data that is technically complete and practically useless. The software federates the models, the structural, the mechanical, the electrical, the plumbing, the fire protection, and reports every place two elements occupy the same space, which on a complex building is tens of thousands of clashes. The number is meaningless because it conflates clashes that matter with the vast majority that do not: duplicates where the same conflict is reported many times, irrelevant overlaps like a pipe passing through a non-structural element it is designed to penetrate, soft clashes within tolerance, and geometry that resolves itself as the design develops. The coordinator who tried to work the list top to bottom would spend weeks dispositioning noise.

So the real coordination work was never finding clashes; it is triage, separating the few hundred clashes that represent real, consequential conflicts from the tens of thousands that are noise, and then sequencing the real ones so the coordination meetings solve the right problems in the right order. This triage is judgment-heavy and tedious in equal measure, which is why it has always been the coordinator's bottleneck: the tedium of working through forty thousand items and the judgment of knowing which matter. AI addresses both halves, the tedium by processing the entire list quickly, and the judgment by proposing which clashes are real and which are noise, which is truly valuable because it attacks the actual bottleneck rather than the already-solved problem of detecting clashes. The question is how much of the judgment half AI can own, and the answer is that it proposes the triage and the coordinator rules on it.

What AI Truly Does With the Clash List

AI's contribution to clash coordination is real and specific. It filters the noise, identifying and grouping the duplicate clashes so one real conflict reported forty times becomes one item, and flagging the irrelevant overlaps and within-tolerance soft clashes that do not need coordination. It clusters the real clashes by system interaction, by physical area, and by the trades involved, so instead of a flat list of thousands the coordinator gets a structured set of coordination issues, this cluster of clashes in the northeast mechanical room involves ductwork, sprinkler, and structure, which is the form a coordination meeting can actually work from. And it can propose a priority order, surfacing the clashes that look most consequential first.

This transforms the artifact from a meaningless number into a workable agenda, which is the genuine value: the coordinator starts not from forty thousand undifferentiated items but from a few hundred clustered, prioritized, real conflicts organized by the systems and trades that have to resolve them. The clustering by trade is especially valuable because coordination is fundamentally about getting the right people in the room to resolve a conflict, and grouping the clashes by which trades are involved directly produces the meeting structure. But every part of this, the noise filtering, the clustering, the priority order, is the AI's proposal based on the geometry and patterns it can see, and the coordinator's job shifts from doing the triage to ruling on it, confirming the noise is really noise, the clusters are right, and the priority reflects what actually matters, which is where the AI's view and the project's reality can diverge.

AI turns forty thousand clashes into a few hundred clustered, prioritized, real conflicts organized by system, area, and trade, which is a workable coordination agenda rather than a meaningless number. But the noise filtering, clustering, and priority are proposals based on geometry, and the coordinator rules on them, because what matters depends on cost, sequence, and consequence the model does not contain.

Why Priority Is a Judgment the Model Cannot Make

The heart of the lesson is that the priority of a clash, which one matters most and must be solved first, depends on factors the model does not contain. The geometry tells the AI that two elements overlap, but it does not tell the AI what it costs to resolve the clash one way versus another, which trade has the sequence priority because they install first or have the least flexibility, which clash sits on the critical path because the area is about to be built, or which resolution has downstream consequences for other systems. A clash that looks minor geometrically can be the most consequential one in the model because resolving it wrong forces an expensive rework, and a clash that looks major can be trivial because one trade has easy flexibility to move.

This is why the priority is a judgment the coordinator makes, not a property the AI can read off the geometry. The coordinator knows that the mechanical clash in the area scheduled to be built next month outranks the larger clash in an area six months out, that the clash involving the trade with no flexibility outranks the one where any party can easily move, that the clash whose obvious resolution would violate a code clearance or a maintenance access requirement is more consequential than its geometry suggests. The AI can propose a priority based on clash size, count, and system, useful as a starting point, but the real priority weaves in cost, sequence, stakeholder flexibility, and downstream consequence, which is exactly the coordinator's expertise and exactly what the model omits. So the AI's priority order is a first cut the coordinator re-ranks against the project's reality, and that re-ranking is the coordination judgment that AI accelerates the inputs to but cannot replace.

The Stakeholder Dimension: Coordination Is About People

There is a dimension of clash coordination that pure geometry entirely misses and that the lesson's title points at: coordination is fundamentally a stakeholder problem, about which trades have to give way and which meetings have to happen, and the priority of a clash is partly a function of its stakeholder consequence, not just its physical one. A clash between two trades who coordinate easily and have a cooperative relationship is lower priority than a clash of the same geometry between trades where the resolution is contested, because the second will take more meetings and more management to resolve and risks becoming a dispute if not handled early. The coordinator who knows the project's human dynamics prioritizes partly on that basis.

The AI sees none of this, because it is in the people and the contracts, not the model, so the stakeholder dimension is a layer of priority judgment that sits entirely with the coordinator. The clash that involves a sole-source trade, a long-lead system, a contentious scope boundary between two subs, or a design element the owner cares about carries a stakeholder weight that its geometry does not show, and the coordinator factors that into the order and the meeting structure. This is why the AI's clustering by trade is a starting point that the coordinator adjusts: the AI groups by which trades are geometrically involved, but the coordinator sequences the resolution by which stakeholder conflicts are hardest and most consequential, getting the contested ones into coordination early when there is time to resolve them rather than late when they become change-order disputes. The stakeholder dimension is the part of coordination that is irreducibly human, and it is precisely the part the AI's geometric view cannot touch, which is why the priority and the meeting plan remain the coordinator's to set.

Trusting the Federated Model Itself

A subtler point underlies the whole exercise: the clash detection is only as good as the federated model, and the AI's triage inherits whatever is wrong with the inputs. If a trade's model is out of date, missing elements, or modeled at the wrong level of detail, the clash report reflects that, reporting clashes that do not exist in the real design or missing clashes that do, and the AI triaging that report cannot know the model is wrong, because it sees only the geometry it is given. So a confident AI triage of a flawed federated model produces a confident, well-organized agenda of partly-wrong conflicts, which can send the coordination meetings chasing clashes that will evaporate when the stale model is updated or missing the real conflict that the incomplete model never showed.

This means part of the coordinator's verification is upstream of the triage: confirming the federated model is current, complete, and at the right level of detail before trusting any clash report, AI-triaged or not. The AI's organization of the clashes is valuable only to the extent the underlying model is trustworthy, so the coordinator checks that the latest models are loaded, that no trade is missing, and that the level of detail is appropriate for the coordination phase, because triaging a stale model efficiently just produces efficient waste. The discipline is that the AI accelerates the triage of whatever model it is given, and the coordinator owns both the judgment about the triage and the judgment about whether the model underneath it is worth triaging, the second being a check the AI cannot perform because it cannot see beyond the geometry to whether that geometry is current and complete. A good coordinator verifies the model before the meeting, not in it, because the worst coordination meeting is the one that disposition clashes that were never real, burning the trades' patience on phantom conflicts while the genuine ones wait, and that meeting is exactly what an efficient triage of an unverified model delivers.

The Applied Problem: From Forty Thousand to a Coordination Agenda

Here is the exercise. Take a real or representative federated IFC model clash report with a large clash count, use AI to triage it down to the real, consequential clashes clustered by system, area, and trade, and produce the coordination agenda a coordinator would actually run the meetings from, with the priority order re-ranked against the project's cost, sequence, and stakeholder reality. Run the workflow: AI filters the noise and clusters the real clashes, proposes a priority; you confirm the model is current and complete, rule on the noise filtering and clustering, and re-rank the priority against what actually matters on the project.

Produce two things. First, the coordination agenda: the real clashes clustered into coordination issues by area and trade, sequenced by the priority you set, in the form that structures the actual coordination meetings, who needs to be in which meeting to resolve which cluster in what order. Second, the priority-rationale record: for the clashes where you re-ranked the AI's order, why, the cost, sequence, stakeholder, or downstream-consequence reason the real priority differs from the geometric one, because that record is both the coordination logic and a demonstration of what the geometry could not show. Pay particular attention to the clashes where the stakeholder consequence diverges most from the geometry, the contested scope boundaries and inflexible trades, because those are where the coordinator's judgment adds the most over the AI's geometric priority.

The deliverable is the coordination agenda and the priority-rationale record, and the lasting product is a clash-coordination workflow that uses AI to turn the meaningless forty-thousand number into a workable set of clustered, real conflicts while the coordinator owns the priority judgment that weaves in the cost, sequence, and stakeholder reality the model cannot contain. This is the coordination core of the BIM chapter, and it is a clean instance of the level's pattern: AI does the tedious processing and proposes the structure, the human rules on it with the project knowledge the model omits. The coordinator who masters this runs coordination meetings that solve the right problems in the right order, having spent their time on the priority judgment rather than the triage tedium, achieved because the AI clustered and proposed and the coordinator ranked against reality, which is the only way forty thousand clashes become a coordination plan rather than a number nobody can act on.

Key Takeaways

• The clash detection number (forty thousand) is meaningless because it conflates the few hundred real, consequential conflicts with tens of thousands of duplicates, irrelevant overlaps, within-tolerance soft clashes, and self-resolving geometry. The coordinator's real job was never finding clashes but triaging and sequencing them.
• AI attacks the actual bottleneck: it filters the noise (collapsing duplicates, flagging irrelevant and soft clashes), clusters the real clashes by system, area, and trade, and proposes a priority order, turning the meaningless number into a workable coordination agenda organized by the trades that must resolve each cluster.
• Every part of this, the noise filtering, the clustering, the priority, is the AI's proposal based on geometry, so the coordinator's job shifts from doing the triage to ruling on it: confirming the noise is noise, the clusters are right, and the priority reflects what actually matters.
• Priority is a judgment the model cannot make, because it depends on cost to resolve, sequence (which area is built next), stakeholder flexibility (which trade can easily move), and downstream consequence, none of which are in the geometry. A geometrically minor clash can be the most consequential, and a major one trivial.
• Coordination is fundamentally a stakeholder problem: a clash between contentious trades or involving a sole-source or long-lead scope carries a stakeholder weight its geometry does not show, so the coordinator sequences the contested conflicts into coordination early, before they become change-order disputes. The AI sees none of this.
• The triage is only as good as the federated model: a stale, incomplete, or wrong-LOD model produces a confident, well-organized agenda of partly-wrong conflicts, so part of the verification is upstream, confirming the model is current and complete before trusting any clash report. A good coordinator verifies the model before the meeting, not in it.
• The artifact: triage a real federated IFC clash report down to clustered real conflicts, produce the coordination agenda re-ranked against cost, sequence, and stakeholder reality, and document the priority rationale where your order differs from the geometric one, demonstrating what the model could not show.