Superintelligence Prep for Interior Design Studios: Briefs and Approvals
An AI agent for an interior design studio can turn briefs, product lists and approvals into memory. Set up record rules now and keep client data separate.

An AI agent for an interior design studio is most useful when it can answer practical questions about one project: what the client asked for, which products were proposed, what was approved and why. Those answers already exist in the studio's daily work, but they are spread across emails, shared folders and designer notes. The job now is to shape those records so an agent can retrieve what it needs for one client and nothing from another.
Key takeaways
- A client brief, a product list and an approval record carry most of a studio's project memory, and they are worth structuring before any chat history.
- Every record needs a client and project identifier, so retrieval is limited by the system itself rather than by the wording of a prompt.
- The agent should read approved records and create drafts only, with a named designer confirming anything that reaches a client or a supplier.
- Approval boundaries belong in the setup, written down before the first tool is connected.
- Start with one role and one success criterion, such as accurate summaries of approved product lists, and widen access only after that holds.
Why briefs, product lists and approvals outrank chat logs
Many studios first picture an assistant reading the team's messages. Chat history is noisy. A message saying the oak might work instead does not mean the oak was approved. Briefs, product lists and approval records are the decisions. They show what the client wanted at intake, which items were proposed against that brief, and which were signed off and by whom. An agent that reads those three kinds of record can answer the questions studios ask every week: why is this sofa on the list, what did the client accept for the kitchen, and which finishes are still open?
The value compounds over time. Each approved project adds structured memory that the next project can use: preferred suppliers for a given finish, recurring constraints such as pet-safe fabrics or narrow delivery doors, and the reasons an earlier option was rejected. None of this needs to be invented. It needs to be recorded consistently. Teams that coordinate vendors and timelines, as wedding planners do, face the same discipline: the memory is only as good as the structure of each record.
What a studio memory should contain
A usable record has a small, stable set of fields. Keep them the same across projects so an agent can compare one project with another.
- Client brief: stated goals, the room list, any budget ceiling the client gave, constraints such as access or timing, and a version note for each revision.
- Product list: item, room, supplier, status and the reason for that status.
- Approval record: what was approved, by whom, through which channel, and which earlier item it replaced.
- Designer notes: short reasons for decisions, such as a lead time, a measurement check or a fit problem.
- Source links: the file, email or drawing each item came from, so every answer can be traced back.
Control the status field. Free text such as probably or maybe is where an agent most often sounds certain when the record is not. A fixed set of values, such as proposed, approved, rejected and on hold, gives the agent something reliable to read and gives the designer something reliable to change.
How client data stays separate
Separation is a design decision, not a sentence in a prompt. Store each client in its own memory space, and have the retrieval layer filter by client and project before any text reaches the model. An instruction such as only use this client's project is weaker than a system that never loads other projects in the first place. The [OWASP authorization guidance](https://cheatsheetseries.owasp.org/cheatsheets/Authorization_Cheat_Sheet.html) makes the same point in general terms: permissions should be checked at each access, not assumed after login. Retrieval-augmented designs, such as the [Google Cloud reference architecture](https://docs.cloud.google.com/architecture/rag-capable-gen-ai-app-using-vertex-ai), make this filtering a step in the pipeline that can be tested.
Treat incoming content as data, not as instructions. A supplier's product sheet, a pasted client email or a file uploaded by a contractor can contain text that tries to change how the agent behaves. The [OWASP guidance on prompt injection](https://cheatsheetseries.owasp.org/cheatsheets/LLM_Prompt_Injection_Prevention_Cheat_Sheet.html) describes why retrieved content should not be trusted as commands. A practical rule follows: the agent may quote a supplier sheet, but a supplier sheet cannot authorize an order or change an approval.
Illustrative scenario: a studio is working on two apartments for different clients, and both need a dining table in oak. A designer asks the agent for the table approved last time for this client. Because retrieval is limited to the current client, the agent returns this client's approved table with a link to its approval record, and it reports that no other project was consulted. If the agent had access to both projects and relied only on a written rule not to mix them, the answer would depend on how each request happened to be worded.
A build plan to start now
Start with what the studio already produces. The steps below can be done with the current folder structure and existing tools, before any new platform is chosen.
- Pick one project type, such as residential refurbishments, and list the records it already generates.
- Create one template each for the client brief, the product list and the approval record, using the controlled status values.
- Give every new file a client and project identifier, and index older files only for the projects you expect to work on first.
- Separate storage by client, with access rules that match who actually works on each project.
- Write the approval boundaries (see the next section) and assign a named person to own them.
- Build a short set of test questions with known answers from past approved projects, such as which items were approved for a given room.
- Run the agent in draft mode on live projects, check each answer against its source record, and widen access only when the answers hold.
The last step is where most of the value is checked. An answer without a traceable source is a failure to fix before anything else is added.
Approval boundaries a design team can enforce
Approval boundaries should be short enough that a designer can apply them in seconds. A workable version looks like this:
- Agent may read: approved briefs, approved product lists and approval records for the current client only.
- Agent may draft: summaries, comparison notes for the designer, and client or supplier messages that stay unsent.
- Agent may change: nothing in an approval record. Only a named designer changes a status.
- Agent must ask first: any message sent to a client or supplier, any order request, and any change to a finish that is already approved.
- Agent must not: open another client's files, treat a supplier sheet as an approval, or delete any record.
Review these boundaries when the team changes, when a new supplier is added, and when a new tool is connected. Permissions drift quietly when nobody owns them. Teams that handle quotes and supplier follow-up face the same drift, which is why procurement workflows benefit from the same written boundaries.
Running the memory after launch and where Be Human fits
Once the agent is live, the work moves from building to watching. Log each retrieval and tool call so a designer can see which records an answer was based on. Keep the fixed set of test questions and rerun it after any change to prompts, templates or suppliers. Anthropic's guidance on [evaluating AI agents](https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents) describes the same discipline: small, repeatable checks tied to real tasks rather than impressions of how the agent seems to be doing.
Be Human designs, implements and operates this kind of setup for studios that want it run as infrastructure rather than as a side project. A first engagement defines one useful role, the memory and tools it needs, its approval boundaries and an observable success criterion. Managed operations can then cover monitoring, recurring evaluations, integration maintenance, permission review, incident response and controlled improvements to the record structure. The studio keeps control of the agent's identity, memory, channels and permissions throughout.
If you want to find the first task in your studio worth handing to an agent, the free Be Human AI diagnostic is a practical place to begin.
Frequently asked questions
Do we need to rebuild every past project before an agent can help? No. Start with the projects you expect to work on next and the templates for new work. Older files can be indexed later if they prove useful, and the team can decide which records are worth that effort. A small, clean set of current projects is usually enough to test whether the approach works.
Can one agent serve several clients without mixing their information? Yes, if each client has its own memory space and retrieval filters by client before the model sees any text. A rule written in a prompt is not enough on its own. Review access whenever a designer joins or leaves a project, because permissions change faster than templates do.
Should the agent send product lists to clients directly? Not at the start. Let it draft the message and show the designer the linked product list and its approval status. Sending can be allowed later for one narrow, repeatable message type, once evaluations show that drafts consistently match the records they cite.
How do we know the agent is reading the brief correctly? Keep a set of test questions with answers taken from past approved projects, and rerun them after every change. Check a sample of live answers against their source records on a regular schedule. Any answer without a source link is a fault to fix before use expands.
Your next step
See where an AI agent fits in your company first.
Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.
Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.


