How Architecture Firms Should Build Their AI Stack

Most architecture firms are not short on AI tools. They are short on structure.

One person uses a general chatbot for meeting notes. Another tries image generation for concept work. A third experiments with a specification assistant. Someone else builds a spreadsheet full of prompts. The experiments may be useful, but they rarely add up to a durable capability for the firm.

Architecture firms need an AI stack.

That does not mean a large enterprise system or a complex technology program. It means a clear set of layers that define how the firm uses models, agents, project knowledge, product knowledge, and review. Without that structure, AI stays personal and inconsistent. With it, AI becomes part of how the studio works.

The stack problem

The problem with many AI deployments is that they start at the tool level.

A firm buys a tool for writing, another for meetings, another for images, another for search, and another for product data. Each tool has its own interface, its own memory, its own permissions, and its own assumptions about what an architecture firm needs.

The result is fragmentation. Project context is copied into multiple systems. Product decisions are recreated from scratch. Staff develop private workflows that do not transfer. The firm cannot easily govern what is happening, and it cannot build compound value from prior work.

An AI stack should solve that by separating capabilities into layers.

The six layers

The first layer is model access.

Firms need a way to use current models safely and flexibly. The specific model will change over time. The stack should make it possible to route work to the right model without rebuilding the whole workflow each time the market changes.

The second layer is the agent harness.

This is the operating layer for agents. It defines how requests are routed, what context is available, what tools can be used, how work is logged, and where human review enters the process. We have argued that architecture firms need an agent harness because practice work is too contextual for isolated prompt interfaces.

The third layer is domain agents and skills.

A project dossier agent is different from a product researcher. A decision log skill is different from a marketing copy workflow. Architecture firms should define agents around recurring practice jobs: project intake, meeting follow-up, precedent research, material research, specification preparation, consultant coordination, and client presentation support.

The fourth layer is project memory.

AI becomes more useful when it can work from the firm’s actual project context: briefs, decisions, meeting notes, drawings, correspondence, standards, and unresolved issues. This memory should be structured enough for agents to use and reviewable enough for teams to trust.

The fifth layer is product and material memory.

This is where Norma fits. A conventional material library helps people find products. An AI material library helps agents work with product memory: what was specified, what was rejected, what performed well, which vendor responded, what lead times mattered, and which alternatives were acceptable.

The sixth layer is governance and review.

Architecture firms cannot treat AI output as finished work. The stack needs permission boundaries, source traces, review states, and clear ownership. The goal is not to slow teams down. It is to make AI output easier to trust, revise, and approve.

Why the harness matters

The agent harness is the layer that turns AI from a collection of tools into a firm capability.

Without a harness, each workflow depends on a person remembering the right prompt, finding the right files, pasting the right context, and judging the output from scratch. With a harness, the firm can define repeatable patterns.

For example, a project intake workflow can always gather the same categories of context. A material research workflow can always check Norma. A meeting follow-up workflow can always create action items, unresolved decisions, and references back to the relevant project.

This is the role Architecture Studio is designed to play. It gives the firm a place to run agents, connect skills, and keep work tied to the practice context instead of scattered across disconnected interfaces.

Buy, build, own

Architecture firms do not need to build every layer themselves. They should buy model access, use strong platforms where they make sense, and adopt purpose-built tools for architecture workflows.

But firms should be careful about where their knowledge lives.

Project history, material decisions, product relationships, standards, and office methods are strategic assets. If those assets are trapped inside a single vendor workflow, the firm loses leverage. A strong AI stack should make firm knowledge more portable, more structured, and more useful over time.

That is why the distinction between a tool and a layer matters. A tool completes a task. A layer organizes capability.

A starter stack

A practical starter stack for a small or mid-sized firm can be simple.

Start with approved model access and clear internal guidelines for what can be shared. Add Architecture Studio as the agent harness for repeatable workflows. Define a small number of agents around real pain points: project dossier creation, meeting follow-up, product research, and specification support.

Then connect project memory. Do not try to structure every document in the firm at once. Begin with active projects, current meeting notes, client briefs, and decision logs.

Next, build product and material memory in Norma. Capture the products the firm actually uses, the vendors it trusts, the substitutions it has accepted, and the product data it repeatedly needs.

Finally, add review states. Make it clear when output is draft, reviewed, client-ready, or approved. The value of AI increases when the firm can see where work stands.

The stack does not have to be complicated. It has to be intentional.

Architecture firms that build this way will get more value from each new model and each new agent because the underlying structure is already in place.