Putting AI to work in the architecture office

At a glance TL;DR

• Current agents can operate architectural software and help build the small automations a practice needs.
• The strongest experiments finish a substantial job: prepare documentation, reconcile project information or remove a recurring production burden.
• Measure the effort removed across the team, including checking and repair. More generated material can mean more work for everyone else.

A building model can contain the information needed for a revision report and still leave someone spending a day copying it into Excel. The information exists. The software exists. A skilled person supplies the connection between them.

In an August 2026 customer account published by Maestro, architecture and engineering firm MOREgroup described replacing precisely this job. A custom Revit tool collected comments from revision clouds, associated them with revisions and sheets, and removed what the customer estimated was eight hours of manual entry. The same account describes an AI-assisted sheet-view renumbering tool built in about half an hour and subsequently adapted in response to staff feedback. MOREgroup’s account

This is a useful place to start thinking about AI in an architecture office: the substantial work that sits between knowing what needs to happen and having it done. Much of it involves finding information, transferring it, applying an established decision repeatedly and checking that the results agree. Agents are beginning to take on that work inside the applications practices already use.

The opportunity deserves an ambitious response. A practice can now attempt automations it previously lacked the time, specialist knowledge or development budget to pursue. The sensible unit of ambition is a completed job.

That gives us a way to distinguish progress from a growing pile of output. A polished coordination report that someone must decode, verify and turn into actions may simply move work onto the next person. A longer email can be worse than the short one it replaced. An AI-written specification assembled without resolving its choices can distribute uncertainty throughout a project.

BetterUp and Stanford Social Media Lab call this failure “workslop”: plausible-looking material that transfers effort to its recipient. Their 2025 survey of 1,150 US desk workers found respondents reporting substantial cleanup work. It is a useful warning about how productivity can be miscounted. The person generating the document finishes quickly; the people receiving it inherit the thinking. Workslop research

Architectural software offers a more tangible test. Has a drawing set advanced? Have inconsistent values been reconciled? Can a colleague use the result without reconstructing what happened?

The technology available now makes those questions practical. An agent combines a model’s ability to interpret a request with access to tools that carry out actions. That access might come through an assistant built into an application, a specialist add-in, or a connector such as MCP—the Model Context Protocol, through which an AI system can call tools provided by other software. It can also write a script for a job the application does not already package conveniently.

In Revit 2027, Autodesk’s Assistant documentation describes a sequence that creates plans, applies templates, tags rooms, repeats the work across levels, assembles sheets and exports a PDF. That is enough to suggest a meaningful office experiment: prepare a defined package of documentation from an existing model using the practice’s standards. Autodesk’s documented workflow

Take a set of room drawings. Establish one satisfactory example, identify the rooms in scope and give the agent the relevant templates and naming conventions. Then attempt the repeated production together: views, annotations, sheet placement and output. The package provides a clear result to inspect, while the example makes the intended standard concrete.

The interesting saving is the repeated assembly, including the attention spent moving between operations. Crop regions still need to make sense. Annotations must describe the right things. A sheet can exist and remain a poor drawing. Reviewing the actual package reveals whether the agent has removed production effort or created a different kind of tidying job.

Another substantial opportunity lies in the exchanges between models and spreadsheets. Door information, room data and equipment schedules repeatedly leave one system, acquire changes somewhere else and need to return. This is where confident AI prose is particularly unhelpful. The job is to reconcile actual records.

BIMLOGIQ’s Argus documentation describes an agent working inside Revit with model data and attached files. Its schedule tools include an export intended for editing that preserves element identifiers, followed by reimport of changed values. The distinction matters: a spreadsheet that looks right is not necessarily connected to the right model elements. Argus, schedule exchange documentation

A useful experiment would take an agreed door-data update, compare it with the model and prepare changes against matched elements. Missing identifiers, duplicate matches and contradictory values become specific questions for the team. Once those are resolved, apply the agreed changes and read the resulting model values back. This workflow needs to be configured and tested; importing a schedule alone does not perform the reconciliation.

Its value is easy to recognise. People stop carrying hundreds of settled values between applications and concentrate on the few that require a decision. A discrepancy attached to an identifiable door is useful. A general paragraph saying door information should be coordinated adds little.

There is also work worth attempting earlier in design. McNeel’s Rhino MCP platform connects agents to Rhino and Grasshopper so they can inspect files, create geometry, run commands and build or edit scripts and definitions. That makes agent assistance available within editable design work. McNeel’s documentation

For example, a designer could ask an agent to help build a configurable shading study around existing façade geometry, then adjust the parameters directly. The useful outcome is a working study whose assumptions and geometry can be inspected. Producing twenty attractive variations is less valuable if nobody can tell what changed or develop the preferred option. Direct modelling remains part of the process: sometimes moving the geometry is the fastest way to express the idea.

These experiments also offer a route to a more capable office over time. Once a recurring operation works, save it as a reusable command or script. A task does not need a fresh conversation every time it runs. AI can help develop and repair the automation; established code can then carry out the repeatable part. The MOREgroup example illustrates how much value can sit in a small tool fitted to an actual frustration.

For a small practice, this suggests a practical starting point. Choose a recurring job that consumes meaningful time and has a result the team knows how to judge. Put someone who understands that job alongside the agent, with the files, standards and software access needed to attempt it. Use a recoverable project copy and approved data access, then push the experiment far enough to discover whether the whole job can be completed.

Count setup, supervision, checking and correction as part of the work. Keep the experiment if it releases useful capacity after those costs; revise or abandon it if the burden merely moves. A successful result should also survive being used by someone other than its creator.

The return can be fewer late hours assembling drawings, less time recovering information and more room to pursue a design question properly. Those are consequential changes in a practice. Agents give us a way to attempt them now, with the software and projects already in front of us. The measure is how much further the work gets—and how much less effort everyone else must supply to finish it.

Sources

1. Maestro: MOREgroup cuts Revit automation delivery time, 26 August 2026 Retrieved 29 October 2026
2. BetterUp Labs and Stanford Social Media Lab: Workslop Retrieved 29 October 2026
3. Autodesk Assistant in Revit Retrieved 29 October 2026
4. Autodesk: How to enable Autodesk Assistant in Revit, 14 October 2026 Retrieved 29 October 2026
5. BIMLOGIQ: What is Argus? Retrieved 29 October 2026
6. BIMLOGIQ: Exporting and Importing Revit Schedules Retrieved 29 October 2026
7. McNeel: Rhino MCP Platform Retrieved 29 October 2026