Hello, Artificial Design Intelligence
- Diego Méndez Arce
- Jul 22
- 13 min read
Updated: 2 days ago
Or: How I Learned to Stop Worrying and Love the New I in BIM
On Artificial Design Intelligence, the collapse of the production layer, and why machine authorship should make the "I" more transparent to humans, not less. Diego Andrés Méndez Arce, Arkamos Architecture, Costa Rica.
Sometime in the next few years, an architectural practice will spend twelve months developing a scheme, and the complete structural and MEP response to that scheme will be generated overnight. Not visualized. Generated: sized, routed, coordinated, and documented. The technology already exists in early form. Augmenta has demonstrated automatic routing of 25 miles of electrical containment across a data center overnight [LINK: Augmenta], and its MEP alpha produces a discipline-complete response to an architectural model, piping and fixtures included.
When the two halves of a multidisciplinary team operate at those time scales, something structural has changed. My argument is direct: BIM systems management must become a practice's data-governance and audit infrastructure, the discipline that lets licensed humans inspect, approve, and stand behind machine-authored work. The preparation is available to practices of any size now, because it is a matter of governance discipline rather than software budget.
Computation versus computerization
I first made the core of this argument in 2018, in a licenciatura thesis on programming material systems [LINK: thesis], built on a distinction from Menges and Ahlquist: computerization digitizes processes that are already preconceived and well defined; computation starts from elemental properties and generative rules and derives form as
output. One automates the execution of decisions already made. The other participates in making them.
Read our tool history through that lens and the pattern is uncomfortable. CAD was computerization: the drawing board digitized. BIM, sold as a revolution, was mostly computerization too, richer in what it recorded, but with humans still making the consequential decisions and machines still keeping the minutes. The third shift is different in kind: computation is beginning to arrive at building scale. Higharc, which closed a $95M Series C in June 2026 to scale AI for homebuilding, is not a Revit plugin; it converts a sketch into an enriched, code-compliant model in minutes [LINK: Higharc]. Qonic is architected as a database rather than a file, built for machine reasoning rather than human authoring [LINK: Qonic]. Autodesk itself shipped Revit 2027 with a native MCP server and an agentic assistant that reads and modifies the live model from natural language [LINK: Architosh]. When the incumbent rebuilds its flagship around agents, the direction of travel is clear. AEC Magazine's Martyn Day projects augmentation dominating through 2030 and solver-driven, automated engineering generation going mainstream in the decade after [LINK: AEC Magazine].
Four eras, one table
Notice two things. The symmetry between rows 2 and 4: humans machine-assisted becomes machines human-directed, and each era's "work" becomes the next era's derivative. The drawing became a print of the file; the file became an output of the model; the model is about to become an output of the intent. And then notice how every row of the last column begins with the same two words. Authorship migrates, the work migrates, the derivative migrates. "Human signs" never moves. Hold those two words; the essay returns to them.
I'm not going to defend "Artificial Design Intelligence" as the permanent name. The industry will standardize on whatever the largest marketing department ships. But ADI usefully says what the thing is. The first thing every programmer makes a new machine say is "Hello, World." The AEC industry is beginning to run its first real program, and it would be wise to think carefully about what we make it say. For the years between now and then, the honest name may already be here: Building Intent Modeling. The acronym survives intact; only its center of gravity moves. For twenty years we have said the value of BIM lives in the I, meaning Information: the curated record of what was decided. In the era now arriving, the I also stands for Intent: the reasoning that decides. The model becomes a derivative that machines generate from it.
The I doesn't just change meaning. It changes state.
Here is the part of the argument I haven't seen made elsewhere, so I'll make it carefully.
We built the entire apparatus of BIM, the common data environments, IFC, COBie, LOD definitions, information exchanges, the ISO 19650 machinery of project and asset information models, for one underlying reason: information was expensive to produce and impossible to infer, so it had to be explicitly encoded, stored, transported, and traced across the lifecycle. BIM is, structurally, a logistics system for a scarce commodity.
That scarcity assumption is now failing, from two directions at once. From the analysis side, a decade of semantic enrichment research has demonstrated that AI can infer the information humans laboriously encode: element classifications, room types, and missing compliance data recomputed from bare geometry and context, with current methods explicitly aimed at reducing dependency on manually entered data [LINK: Automation in Construction, semantic enrichment literature]. From the generation side, LLM-based frameworks like Text2BIM produce editable models with semantic information directly from natural language instructions, with rule-based checkers in the loop [LINK: Text2BIM, TU Munich]. In both cases the semantics arrive as an emission of the process, not as a curated artifact.
Software engineers have a name for this pattern: the compiled binary, the materialized view. You don't archive what can be regenerated on demand; you archive the source and rebuild.
Most of what fills a building model, everything that follows from geometry, physics, code, and convention, is turning out to be exactly that: regenerable, a cache. Information is becoming something the process churns out at any given moment, because it lives embedded in the process itself.
But not all of it. What cannot be regenerated, in principle and not merely yet, is the contingent information: that this client rejected the courtyard scheme in March, that the setback was traded against height in a negotiation, that the cheaper cladding was the client's decision made against advice. Those facts aren't inferable from anything, because they aren't consequences. They are decisions. Strip away every piece of information a machine can reconstruct, and what remains is precisely intent and its history.
That is the full form of the thesis: the I changes meaning because it changes state. Inferable information collapses into the process and gets emitted on demand. Contingent information, which is just intent wearing its audit trail, becomes the only thing worth persisting. The two halves are one claim.
One current runs the other way, and it defines rather than defeats the argument. Regulation is demanding more fixed information, not less: the UK's golden thread under the Building Safety Act requires auditable, persistent records, and compliance platforms respond by specifying explicit information requirements rather than inferring them [LINK]. A probabilistically regenerated fact is not an admissible record. So the future is not information-free; it is information-on-demand punctuated by materialization points, moments when the stream must be snapshotted into an authoritative, validated, signed record. Someone has to govern when those snapshots happen, what they must contain, and how machine-emitted information gets validated before it becomes evidence.
The accountability inversion
Which brings us back to the two words in the table that never move, because they point at the constraint the entire ADI conversation keeps avoiding: the law.
In most jurisdictions, a building cannot be permitted, insured, or built without a licensed human taking legal responsibility for its design. The architect or engineer of record signs and seals; professional liability attaches to a person, not a process; and no current legal framework treats "the model generated it" as a defense. Emerging AI regulation points in the same direction: high-risk automated systems are being required to operate under meaningful human oversight, not merely human presence [LINK: EU AI Act, human oversight provisions]. Accountability is, by design, non-delegable. Machines will author; humans will answer. That is the Strangelove clause, and the reason for this essay's subtitle: you get to stop worrying only after you have built the oversight that lets you sign.
Now follow that constraint to its structural consequence, because it inverts the intuition most people carry into this topic. The naive expectation is that as machines take over production, humans need to understand less of what happens inside it. The legal reality forces the opposite: the more authorship machines take, the more transparent the information must become to the humans who sign. I cannot honestly stamp what I cannot audit. Every increase in machine authorship raises, rather than lowers, the required legibility of the I, because the human at the materialization point must be able to see what was generated, from what intent, under what constraints, validated by what checks, before converting it into a record that carries their license.
So the I of the coming era has a dual-legibility requirement that BIM never had. It must be machine-readable so that agents can generate and reason over it, and it must be human-traceable so that a licensed professional can follow the chain from intent to output and sign with their eyes open. A practice that achieves the first without the second has built a liability machine: production at machine speed, accountability at zero visibility. The gap between those two legibilities is precisely where professional risk will concentrate, and closing it is an engineering discipline, not a good intention. Provenance capture, decision logs, validation reports designed for human review at human reading speed, sign-off protocols that specify exactly what a human must have seen: this is the audit infrastructure of machine authorship, and someone in every practice has to own it.
Architecture was never immune
Now the uncomfortable part, and I'll aim it at myself first. In 1964, Bernard Rudofsky's MoMA exhibition Architecture Without Architects made the point politely: most of the built world was created without us, by vernacular builders encoding accumulated judgment into repeatable form. The coming era is the second age of architecture without architects, except this time the vernacular is machine-generated, and the accumulated judgment is a training corpus.
Translation is the precedent worth studying. Machine translation commoditized the production layer of an entire profession in under a decade: the bulk market collapsed into software, and human translators were repriced from authors to post-editors. And yet an apostilled document still requires an official translator's signature. Courts, ministries, and immigration authorities do not accept machine output; they accept a licensed human's certification of it, because when the translation is wrong, liability must attach to someone, and it will never attach to the model. The profession's production was delegable; its accountability was not. So the official translator's job quietly inverted: from producing the translation to auditing and certifying machine-drafted output, staking a license on work a machine performed. Read that sentence again with "architect of record" in it. The inversion is not a forecast. In an adjacent profession, it has already happened.
I write that as a practicing architect: machines will design too, and some of what I currently sell as expertise will be generated on demand some time in the future. The engineering disciplines may tip first because many tasks, such as pipe routing, are closer to constrained optimization than questions of taste. The honest move is to ask what, specifically, cannot be regenerated, and reorganize the practice around it. We already have the answer from the previous section: intent, judgment, and the governance of the records that prove them.
The one-question test
Inside our practice, every investment of time, money, or attention now has to answer a single question:
Does this make our production cheaper, or does it make our judgment more available and more valuable?
If it makes production cheaper, it is directionally exposed: AI will make production cheaper faster than any practice can, and investments in that race depreciate on a short clock. If it makes judgment more available and more valuable, it compounds, because judgment is the layer machines amplify rather than replace. The test produces uncomfortable answers, which is how you know it is working. Four examples of what it changed for us.
1. We rebuilt our model organization standard as a governance system with a machine-checkable spec. A formal naming grammar, a validation expression that runs programmatically against every Archicad model we produce and flags non-canonical structure, a bilingual code table, and a written protocol for how new categories are created, justified, and approved. I won't reproduce the specification; our syntax is proprietary and, frankly, only correct for our workflow. The architecture of the thing is the point: a standard is only real if a script can verify compliance without a human opening the model. If your standards live in a PDF and are enforced by memory, you don't have standards; you have suggestions. Machine-checkable governance is the cheapest AI-readiness investment available to a small practice, because it converts model quality from a cultural aspiration into an auditable property. It is also the prerequisite for everything in this essay: an agent cannot reason reliably over a model whose organizational logic lives in one senior modeler's head.
2. We keep the tool stack composable, and treat that as strategy. Month-to-month commitments, tools swappable without rebuilding workflows, no platform allowed to become load-bearing for the practice's identity, and that includes the Archicad workflow we have spent years refining and have no intention of leaving. Loyalty to a tool is fine; structural dependence on one is a bet the practice shouldn't be making while the authoring paradigm itself is in play. The subscription premium is the price of optionality, and in this period, optionality is worth more than efficiency. The same logic governs hardware: for platforms that are clearly transitional, bridge solutions at minimum viable spec, never capital commitments that solve this year's problem and become next year's overhead.
3. We automate the operations around the judgment, not the judgment. Project tracking, meeting capture, file synchronization: the connective tissue of running a practice is delegated to software so that principal attention concentrates where it cannot be delegated. This looks mundane next to generative design demos. It is not. Every administrative hour removed from the judgment layer is an hour of the only product that will still command a premium when production is abundant.
4. We are writing the judgment down. Briefing methodology, decision frameworks, the reasoning behind typological choices on past projects. Today's tools store outcomes, not rationale; the intent behind decisions lives in the architect's head, opaque to any machine. The next generation of tools treats intent as a first-class object [LINK: AEC Magazine], and practices that have externalized their reasoning will have something to feed those systems. Practices that haven't will be starting from zero, again. In an era where information is emitted rather than stored, documented intent is the only proprietary dataset a practice truly owns.
What BIM systems management becomes
Put the threads together and the role inverts rather than dies.
The BIM manager as most practices run it is a descendant of the CAD manager: templates, standards, library hygiene, training, the enforcement of conventions so humans produce consistent documentation for other humans to read. Every clause of that job description assumes a human author and a human reader, and that assumption is the thing expiring. RIBA's 2025 survey found 59% of UK practices already using AI, up from 41% the year before, with Stages 2 through 4, exactly where BIM management effort concentrates, identified as most exposed [LINK: RIBA AI Report 2025, p. 21].
The successor role is the practice's data-governance authority: the person who defines what machine-readable means locally, who proves compliance by audit rather than by trust, who governs the materialization points where regulation demands that emitted information become fixed, validated record, who engineers the human-traceable audit chain that lets a licensed professional sign machine-authored work honestly, and who curates the intent corpus that everything else derives from. Note what dropped out of that job description entirely: the authoring platform. Nearly everything written about AI in AEC is written as if Revit were the industry, but the governing variable is data discipline, not vendor. A well-governed Archicad model with consistent classification is more AI-ready than a poorly governed Revit model with empty parameters and improvised subcategories. Discipline transfers between platforms. Chaos does too.
None of this is an argument that architecture reduces to constraint satisfaction. Writing the brief, reading a site with human sensibility, carrying a client through the managed uncertainty of a multi-year project, judging whether a building is right for its place and moment: these remain the profession's actual product, and the coming era makes them more visible by stripping away the production scaffolding that has historically obscured them. The uncomfortable part is economic: fees are earned on the production layer, and clients who paid for drawings don't automatically pay for the thinking behind them. That repositioning takes years, which is exactly why it must start before production collapses in cost rather than after.
So: govern the data so machines can reason over it. Keep the stack composable so no platform transition is fatal. Automate the operations so judgment has room. Document the judgment so it becomes an asset instead of a memory. Build the audit chain so a human can sign what machines produce without signing blind. The I in BIM is changing meaning, from Information to Intent, and changing state, from record to emission, while remaining answerable to the humans who put their name on the work. Direction is knowable even when timing is not. Practices that act on direction while others wait for certainty will help write the standards everyone else adopts.
About the Author
My name is Diego Méndez Arce, and I’m a detail-oriented Costa Rican architect with over a decade of experience in residential design and construction. I studied architecture at the Instituto Tecnológico de Costa Rica and the Technische Universität München in Germany, and I’m the founder of Arkamos Architecture Costa Rica.
At Arkamos, I lead BIM systems, standards governance, and computational design across an Archicad-based multidisciplinary workflow. My 2018 thesis on computationally programmed material systems anticipated several of the arguments developed here. We design and build homes with a deep belief that good design nourishes the human spirit and can profoundly improve daily life. Our team approaches every project with a methodical, transparent process to ensure that each home is crafted with precision — and delivered with predictability in scope and quality.
I enjoy sharing what I’ve learned with people unfamiliar with Costa Rica’s real estate architecture and construction landscape. You can explore more of our blog posts on architecture, design, and building in Costa Rica on our blog section. If you have questions or are starting your own project, feel free to send me an email or book a free virtual consultation.
Sources:
Augmenta — automated MEP and electrical routing. augmenta.ai
Méndez Arce, D. (2018). Sistema material programable mediante herramientas de diseño computacional. Licenciatura thesis, Tecnológico de Costa Rica. https://repositoriotec.tec.ac.cr/items/f3440872-f5e5-4ec2-b184-0158dda375e5
Menges, A., & Ahlquist, S. (2011). Computational Design Thinking. Wiley.
Higharc — $95M Series C announcement, June 2026. higharc.com/newsroom
Qonic. qonic.com
Architosh — "Autodesk Revit 2027: big new AI and graphics changes," April 2026.
AEC Magazine — Martyn Day, "The agentic future of BIM."
Automation in Construction — semantic enrichment of BIM models (ScienceDirect).
Text2BIM — Technical University of Munich, open-source repository (GitHub).
GOV.UK — "Keeping information about a higher-risk building: the golden thread."
EU Artificial Intelligence Act, Article 14 — human oversight.
AEC Magazine — "Neural CAD: AI foundational models."
RIBA — AI Report 2025, p. 21.
Rudofsky, B. (1964). Architecture Without Architects. Museum of Modern Art, New York.
Comments