Article

AI now communicates with your BIM models. Is your as-built drawing ready?

Co-founding partner, business development and technology watch

Share

The number that says it all

According to the Unanet 2026 AEC Inspire report, cited at AU 2026 by the firm Avant Leap, approximately 75% of the AEC firms surveyed are already using AI. However, when asked the following question, only 29% of those same firms expressed a high level of confidence in the data feeding the AI.

Three out of four companies use AI. Fewer than one in three trust what they read. As the speakers summarized: it’s not an AI problem; it’s a problem with how workflows are defined. And for a model built this way, the first piece of data in the workflow is the reading.

At Topo3D, this statistic directly affects us. We deliver point clouds and models that become our clients’ “ground truth.” If AI agents are now going to analyze and correct these models, the quality of what we deliver determines the quality of what they will produce.

A terrestrial laser scanner converts a machine shop into millions of measured points. It is this data—not the model—that serves as the reference: if AI agents query the model, the reliability of their responses depends first and foremost on the quality of the survey data.

MCP is no longer an emerging technology

A year ago, people were still talking about MCP as a promising standard. The data presented at AU 2026 paint a very different picture:

  • Launched in November 2024, the SDK had reached approximately 97 million downloads per month by March 2026;
  • OpenAI, Google, Microsoft, Salesforce, and AWS have adopted it, and it is now governed by the Linux Foundation;
  • The official registry listed 9,652 servers in May 2026;
  • Autodesk is offering an MCP server for Revit as a technical preview, and independent projects are already showcasing Civil 3D, Inventor, and AutoCAD.

As a reminder, the principle is simple. An LLM, on its own, cannot see your model and cannot interact with your software. An “MCP server” exposes an application’s capabilities (reading a model, querying elements, modifying a parameter, generating an export) in a standard format that an AI agent can use. The infrastructure is in place. The real question has become: what do we entrust to the agent, and what do we keep for ourselves?

We’re not just talking about this in theory. Topo3D runs its own MCP servers in-house, which connect our AI agents to our project management tools and automation workflows. That’s what allows us to view AU presentations from a practitioner’s perspective.

Three real-life cases presented at the AU 2026 conference

1. Egis, Grand Paris Express: Compliance Verified in Days Rather Than Months

On the Grand Paris Express (200 km of subway, 68 stations), every additional square meter underground comes at a high cost: deeper excavation, thicker walls. Egis must therefore minimize the size of technical rooms while ensuring they meet hundreds of requirements (dimensions, clearances, equipment, access). Ten years ago, this verification process required six experts working for months, reviewing each Revit model one by one. With their automated workflow—which combines Autodesk Platform Services and AI—Egis reports that the process now takes “days, not months,” and that the system groups together non-conformities that share the same root cause.

2. Nacorm: an agent that learns from corrections made to the project

The classic problem with an automated workflow: it repeats the same error every time it runs, and someone has to correct it manually each time. Nacorm presented an MCP pipeline that saves human corrections as project rules. Out of 60 coordination issues to be classified, the agent’s accuracy rose from 38% to 83% after just five statements such as “anything that goes through the structure goes to structural.” No lines of code: just facts about the building, already known to the team. A new session, with no history, achieved 78% accuracy because the state was stored on the server, not in the conversation. The speaker himself highlighted the limitation: a single project, a single AI model.

3. Autodes: the “Ask → Check → Correct” loop

Phuc Le of Autodesk presented a personal research project built on public APIs: a multi-agent system in Revit. One agent reads the model via MCP, a second checks it against standards written in plain language (YAML files), and a third suggests corrections. Each correction is previewed, approved by a human, and logged before being written to the model. The loop repeats until the model passes validation.

His conclusion is worth noting: the parameters can be automated; the geometry and design intent require a human.

The dividing line: parameters on one side, geometry on the other

This sentence precisely defines the scope of scan-to-BIM.

What the agent does well: naming families according to the client’s standard, filling in missing parameters, verifying that a fire door has the correct resistance rating, and generating the discrepancy log and the draft report. These are deterministic tasks: the correct answer is known in advance and is therefore verifiable.

What the surveyor cannot make up: the actual position of a wall, the deflection of a beam, the lack of plumb in a post, the discrepancy between what was designed and what was built. This information comes from the field: from a georeferenced survey, from a point cloud whose uncertainty is known (see our article on the true accuracy of a point cloud), or from a model validated by someone who knows how to interpret a point cloud.

In other words, the more agents automate the downstream process, the more critical the upstream process becomes. An agent who checks 300 doors in a few seconds will just as quickly identify 300 misaligned doors if the initial model is incorrect.

When superimposed on the point cloud, the model reveals a wall and a series of doors that are offset from their actual positions. An inspector who checks only the parameters will validate each of these doors; only the measured data reveals the discrepancy.

The boundaries are starting to shift: the Prevu3D case

The line between parameters and geometry is not set in stone. Our partner Prevu3D, a Quebec-based digital twin platform, is a good example of this trend. A few weeks ago, its point cloud distribution tool—a Revit connector—was already making data available very quickly within Revit. The benefit at this stage was access to the data, not yet modeling.

Then the boundary shifted for good. About a week ago, using the Prevu3D API, the AI took over: using a model like GPT-6 Astra, it autonomously navigated the scan, segmented and delineated objects, recognized them, built their geometry, and produced IFC files ready for integration into Revit. To achieve this, the model combines two sources with complementary strengths: the mesh derived from the point cloud, which carries the geometric truth and ensures accuracy, and the Gaussian Splatting derived from the photos, rich in visual information, which helps recognize objects. So we’re no longer talking about a week between capture and modeling in Revit, but 24 hours—and, within a few months, possibly just a few hours from the moment the data is captured, aligned, meshed, classified, and converted by AI into a geometric asset. AI is now tackling the geometry itself.

The mesh provides the measured geometry; Gaussian Splatting provides the visual characteristics that help identify objects. Together, they enable the AI to segment the equipment and generate its CAD model, which must then be verified.

Prevu3D doesn’t take humans out of the loop, however. Eventually, the platform will allow users to visualize the CAD data generated for each object or for the entire site, and to monitor the automated production process as it progresses. Heat maps will quickly highlight problem areas, and tools will automatically compare the produced object to the reference CAD model.

That is exactly the approach we advocate: the more geometry AI generates, the more we need to validate its output using measurements and checks that are independent of the AI itself. See the example of Prevu3D.

Verify what comes in… and what goes out

Reliable input data alone is not enough. We also need to verify what the AI produces, and that is often the hardest part.

A human error often looks like a mistake: an outlier, a blank cell, a missing unit. An AI’s error, on the other hand, looks correct. The table is complete, the report is well-written, and the references seem consistent. Language models are designed to produce plausible information, not true information. The more polished the result appears, the more tempting it is to believe it without verifying it—and the harder it is to spot the error.

That is why, at Topo3D, we have developed our own mechanisms for verifying AI-generated output. They are based on the fundamental principle of surveying: you do not validate a survey against itself; you compare it to independent control points. The same applies to AI: it cannot be the sole judge of its own work. Everything it produces is subjected to independent checks (deterministic tools, field measurements, human review) before it leaves our offices.

The most common pitfall is a subtle one. When reconstructing the geometry, AI tends to model standardized elements: a straight beam, a plumb column, a catalog-style structural member. But the actual building isn’t always like that. A beam has sagged, a column is leaning, a room has been modified on-site. The generated model is clean and consistent, but it describes the building as it should be, not as it is. The operator must therefore quickly identify the differences between the point cloud—which represents the measured reality—and the rendering generated by the AI, understand where the discrepancies lie, and then decide whether to correct the model, accept the discrepancy if it falls within the agreed-upon tolerance, or return to the site. This is the purpose of deviation heat maps.

The AI modeled a straight beam and plumb columns, just like in a catalog. The point cloud shows the reality: the beam has sagged, and one column is leaning. The areas in red indicate where the model deviates most from the point cloud; the targets and the benchmark are control points measured in the field, independent of the AI. The operator can see at a glance where the discrepancies are and can decide whether to correct the model, accept the discrepancy if it falls within tolerance, or return to the field.

The checkpoints we recommend:

  1. An independent source—never AI alone. Each deviation he reports is recalculated using a deterministic tool (a script or measurement software) and compared with control points measured in the field.
  2. Each value must be linked to its source. A value in the deliverable must point to the model element, the point in the cloud, or the source file. Without a source, the value will not be displayed.
  3. Automated plausibility checks. Physical constraints (wall thickness, elevation range), balanced totals, and consistent units and coordinate systems. These tests catch the “plausible” errors that a hurried eye might overlook.
  4. A sample verified by hand. A randomly selected portion of the results is double-checked by a technician. If even a single error is found, the sample size is increased.
  5. A collection of well-known cases. Before modifying a workflow, we run it on data for which we already know the answer, including test cases. If it fails, it does not go into production.
  6. The right to say, “I don’t know.” The agent must leave a field undefined rather than fill it in. Any value that is inferred rather than measured is identified as such.
  7. A documented human signature. A designated professional approves the final deliverable, along with a record of what was verified, how it was verified, and by whom.

Ultimately, the goal is simple: verified, traceable, and defensible data. AI speeds up production. Verification, however, cannot be delegated.

Rules Learned the Hard Way

Avant Leap, which runs MCP servers in production, has shared its best practices—many of which were learned the hard way. Several of these same safeguards can be found in Phuc Le’s project:

  • Only a few tools at a time. Avant Leap recommends a maximum of six tools visible to the agent while they are working. Some servers on the market display more than 170. Seven connected servers already take up a third of a model’s working memory, even before the first question is asked. The more tools the agent sees, the more mistakes they make and the more expensive they become.
  • Preview, then confirm, for every entry. No changes are made to the model without a human reviewing them first.
  • Audit log (write-only). We need to be able to reconstruct what the agent did and what the human approved.
  • A deterministic core, agent-based orchestration. The function chain performs a series of proven functions; it does not replace the calculation.
  • No arbitrary code execution. A tool that cannot be audited or tested will not go into production.

The cost, however, is often underestimated. An agent doesn’t respond just once: they plan, launch a tool, read the result, and start over—sometimes as many as ten times for a single task. According to Gartner (cited at AU), a task assigned to an agent thus consumes 5 to 30 times more computing resources than a simple question asked of an AI assistant. Agents should therefore be reserved for repetitive and high-volume tasks, where the time saved outweighs the cost of automation.

Above all, we must ensure that the agent performs the task in the most direct way possible. An agent who lacks the right tools goes in circles: they waver between similar tools, make unnecessary calls, or process 400 items one by one—and every detour comes at a cost. Practitioners interviewed at AU suggest simple ways to avoid this waste: display only the tools relevant to the current task, clearly describe each tool (what it does, when to use it, when not to use it), provide a single operation that processes an entire batch rather than one call per item, and delegate calculations to proven scripts rather than the model. A well-designed workflow costs less and makes fewer errors.

How this affects your statements: the “ready for agents” deliverable

If your BIM teams are about to link agents to their models, here’s what you should require from the as-built survey and the accompanying as-built model:

  1. Explicit georeferencing. A documented coordinate system, control network, and control points, so that the agent can compare objects that are actually in the same location.
  2. A stated uncertainty. An 8-mm discrepancy can have different implications depending on the accuracy of the measurement. Without this information, the inspector cannot distinguish between a nonconformity and measurement noise.
  3. Standardized naming conventions and parameters. Product families and settings must comply with the customer’s standards upon delivery. That’s the first thing agents read.
  4. A level of detail (LOD) defined per element. The agent must know what was modeled from the point cloud and what was inferred.
  5. Traceability of validation. Who approved what, on what date, and using what method.
As part of a monitoring network, each data point has a known position and a reported uncertainty, represented here by the red ellipses. This is what allows an operator to distinguish a true nonconformity from mere measurement noise.

Risks and Common Mistakes

  • Confusing fluidity with accuracy. A well-worded answer isn’t necessarily correct.
  • Automate the downstream process without improving the reliability of the upstream process. That’s the difference between 75% and 29%.
  • Plugging in too many tools. Performance is declining, and the bill is going up.
  • Believing a result simply because it is well presented. The more polished an AI-generated output appears, the more it deserves to be verified.
  • Aim for complete autonomy. In 2026, the value lies in supervised assistance, not in autopilot.
  • Connect AI to project templates that do not have a data policy. Hosting, access rights, and contractual terms must be settled first.

Recommendations

  1. Start with a workflow that already exists, occurs frequently, and involves errors that can be corrected: parameter verification, deviation log, draft report.
  2. Write your standards in clear, straightforward language. The Nacorm and Autodesk cases show that they become direct guidelines for the agent.
  3. Maintain explicit and documented human approval for all high-stakes deliverables. The agent prepares the document, and the professional signs it.
  4. Require a “field-ready” survey: georeferenced, with stated uncertainty, standardized, and traceable.
  5. Validate the output as well as the input: independent recalculation, traceability of each value, manual sampling.

Conclusion

AU 2026 confirmed it: AI agents working in Revit and other CAD/BIM software have moved from the lab to production. They effectively automate deterministic tasks, such as parameters, naming conventions, checks, and reports. They are even beginning to generate geometry. However, they are entirely dependent on the quality of the source data provided to them, and their output must be verified. The firms that will derive the most value from this will not be those that integrate the most tools, but those that start with data they trust.

Are you preparing to integrate AI agents into your BIM or quality control workflows? Let’s talk about your survey data: Topo3D delivers point clouds and as-built models that are georeferenced, documented, and ready to be used by your tools, and we’re already running our own MCP servers. Let’s discuss a specific case.

FAQ

What is the Model Context Protocol (MCP)?

An open standard that allows an AI agent to use software functions—such as reading a Revit model or modifying a parameter—without requiring custom integration for each application.

Can an AI agent create a model from a point cloud?

It’s getting started. Platforms like Prevu3D are already generating CAD models from mesh data and images, and AI models can now produce IFCs on their own. But the resulting geometry must be verified against measured data, and the design intent still requires human input. Today, the agent excels primarily in handling parameters, naming conventions, and validations.

Is this reliable for engineering purposes?

Only with safeguards in place: a preview before each change, human approval, and an audit log. The agent prepares the document, and the professional signs it.

How can we detect errors in AI if they seem credible?

By never allowing AI to be the sole judge: recalculation using an independent tool, traceability of each value back to its source, automatic plausibility checks, manually verified sampling, and a signature from a responsible professional.

Where should I start?

Through a process that already exists, occurs frequently, and in which errors can be corrected: parameter verification, deviation log, draft report.

Why is the quality of the data even more important with AI?

Because a surveyor can verify data very quickly—even if it is incorrect. A georeferenced survey with a stated uncertainty and standardized measurements is essential for its results to be usable.

Sources

  • AS2698 — MCP on the Autodesk Platform: Three Production Servers Across the Full Stack (E. Galicia Tovar, J. L. Baquero Rivera, Avant Leap), Autodesk University 2026. Cites, among others, the Unanet 2026 AEC Inspire Report, Gartner (March 2026), Epoch AI, and Chroma.
  • AS2183 — Using MCP to Build Self-Improving AEC Data Workflows (N. Sadeg, Nacorm), AU 2026.
  • AS1569 — Multi-AI-Agent BIM Orchestration: Query · Check · Fix Loops for Revit (Phuc Le, Autodesk), AU 2026.
  • BLD3171 — Design Requirements Verification Using BIM Automation and AI Agents (A. Hersant, Egis), Fall 2026.
  • BES4067 — Agents of Change: Agentic AI and MCP Workflows (M. Gustafson, Qnect; A. Sundal, HGA), Fall 2026.
  • Prevu3D — LinkedIn post: https://lnkd.in/p/gxiJSXta

About the author

Co-founding partner, business development and technology watch

A co-founding partner of Topo3D, Jonathan has worked in 3D scanning and reality capture for more than 15 years: terrestrial, mobile and drone LiDAR, modelling and infrastructure inspection. He holds a patent for a point cloud registration technique, applied in particular to the Turcot Interchange scan. As head of technology watch, he follows new sensors and AI tools and evolves Topo3D’s methods accordingly. He holds a bachelor’s degree in environmental design from UQAM.