Our Thesis on AI

Most firms approach AI with tool questions: What platform should we buy? Which model is best? Who gets access? Those questions feel like progress, but they lead to the same place -- a license nobody uses after week three, a pilot that never leaves one enthusiast's desktop.

A tool sits outside the work and helps one person go faster. A layer runs through the work: it changes how knowledge is captured, how information flows between people, how output gets reviewed, and how production actually happens. The questions that matter are layer questions.

Becoming a layer requires three things, and each one is a firm capability rather than a purchase: a defined workflow, trusted context, and human judgment. The rest of this guide takes each in turn, then shows what they look like inside real AEC workflows.

A clever prompt that produced one good render is a moment. It becomes a workflow when it has all five parts:

• Trigger -- the event that starts it (an RFP lands, a submittal arrives, a site walk ends).
• Sequence -- the ordered steps, of which the AI step is only one.
• Reviewer -- the named person who validates before anything moves forward.
• Output -- a defined deliverable, not "whatever the model said."
• Record -- where the inputs, prompts, and rationale are saved so the next project inherits them.

Worked example: a library renovation concept study

On a public library renovation pursuit, the workflow looked like this: the RFP was the trigger. The sequence ran from a stakeholder survey, to prompts written from the survey findings, to first AI renders, to style transfer against site reference photos. The architect and design team reviewed. The output was a set of concept directions for the client. The record was the survey, the prompts, and the design rationale -- saved, so the approach is reusable on the next pursuit.

The prompt was not the workflow. The prompt was one step inside the workflow. The AI worked because the workflow was clear.

Worked example: an agent-orchestrated project assistant

A second pattern, built by designer Serjoscha During, pushes further: a client email triggers an AI assistant (Claude, connected to the firm's tools through MCP) that sets up the project page in Notion, runs web research, posts insights to a Miro board for the designer to react to, and moves concept sketches into Rhino for modeling and rendering. The human stays in the loop at every review point -- rearranging the board, redirecting the model, making the calls. The assistant handles the connective tissue between tools; the designer supplies direction and judgment.

The models are a commodity. Your twenty years of project history, client knowledge, standards, and lessons learned are not. The problem is that in most firms this knowledge exists but isn't readable:

• Project history -- scattered folders
• Standards -- people's heads and personal folders
• Proposals -- multiple systems
• Client knowledge -- the principal's memory, maybe an ERP
• CA knowledge -- Procore and email threads
• Lessons learned -- never captured

The goal is not more data. It is connected context: one structured layer holding project history, firm standards, client knowledge, and lessons learned. Data-native does not mean complicated. It means three things:

• Connected. Your data does not live in seven disconnected places.
• Structured. Your knowledge has meaning, labels, relationships, and owners.
• Accessible. Your tools, teams, and AI can reach the right information when needed.

Before your firm becomes AI-native, it has to become data-native.

The old shape of professional work was search, think, produce -- hours at each step. The new shape is ask, validate, edit -- the asking takes seconds, and the value concentrates in validation. The time is not in the making anymore. It is in the judging.

That matters because AI output can look right and still be wrong:

• Specifications that looked complete -- written for the wrong jurisdiction.
• A schematic design that looked viable -- and quietly killed the passive strategy.
• A code compliance check where the numbers worked -- but the occupancy classification was wrong.

In each case the math was right and the logic wasn't. AI doesn't know what it doesn't know. Your senior people do -- that is the value of them, and it is why review has to be a named step in every workflow rather than an assumption.

The honest question for a firm is not "what can AI do?" It is: what is AI about to reveal about your firm?

Put together, the layer sits between your business objectives and your day-to-day work. Business objectives -- win rate, cost, revenue, client outcomes -- sit at the top. The AI layer sits underneath them, powered by the three capabilities: real workflows, trusted context, and human judgment. All of it rests on a data foundation that is connected, structured, and accessible. And the whole stack serves both kinds of work: process workflows (business development, project setup, project management, knowledge management, firm operations) and project workflows (the phases from concept through construction administration).

None of the boxes in that stack is a tool you can buy. Each is a capability you build -- which is also why the layer compounds. A firm that records its workflows, structures its knowledge, and trains its reviewers gets more out of every new model release. A firm that buys seats gets a better autocomplete.

Proposals and RFP responses

The context that makes this work is a structured library of past proposals, project data, and bios -- the classic case where the knowledge exists but isn't readable.

Early design options

This is the library example from Section 02. The AI step is generative, but the record step is what turns a one-off into firm capability.

Construction document QA

Judgment matters most here: a checker that looks right and is wrong is worse than no checker. Encode firm-specific rules incrementally and keep a named reviewer on every run.

Submittal and RFI administration

The record step quietly builds the CA knowledge base most firms never capture -- the same knowledge that currently retires with senior staff.

Field reporting

Low risk, high frequency, and the output is exactly the kind of structured project history the trusted-context layer needs. A good first workflow for skeptical teams.

Two developments are worth watching because they extend the layer rather than replace it.

Agent orchestration. The MCP-powered assistant in Section 02 is an early example of a pattern that is arriving fast: AI that operates across your tools -- email, project management, whiteboards, modeling software -- rather than inside one of them. This raises the stakes on all three capabilities. An agent without a defined workflow is chaos at machine speed; an agent with one is a genuine force multiplier.

Spatial intelligence. World Labs -- the company founded by Fei-Fei Li -- is building world models: AI that generates and reasons about explorable, persistent 3D environments rather than flat images. Its first product, Marble, turns text, images, or video into navigable 3D worlds that can be exported and edited. For AEC the implication is direct: the render-and-style-transfer step in today's concept workflows becomes a walkable environment a client can move through, and eventually a model that understands space, adjacency, and site the way current models understand text. Firms whose precedent imagery, site photography, and design rationale are structured and accessible -- data-native firms -- will be the ones able to feed these models something worth generating from.

Neither development changes the argument of this guide. Both make it more urgent: the better the models get, the more the difference between firms comes down to workflow, context, and judgment.

That gap between expectation and readiness is where wasted AI spend lives. The way to close it is to measure readiness before buying anything, across four dimensions -- the 4 E's: Exploration (are you solving the right problem?), Efficiency (are your workflows ready?), Effort (is your team ready to change?), and Expense (is the investment tied to a business outcome you can measure?).

The three capabilities map directly onto them: trusted context feeds Exploration, defined workflow feeds Efficiency, and human judgment feeds Effort. Your weakest E is your starting point. A low readiness score is not a failure -- it is a map of your biggest quick wins.

AI will not make every firm smarter. It will make every firm more visible.

The practical sequence: pick one workflow where the hours and pain are real. Draw its five parts. Fix the context it needs. Name the reviewer. Run it on one live project, record everything, and only then scale. That is how a tool becomes a layer -- one workflow at a time.