Every firm has that one person who built a custom GPT on a Saturday, named it something like FirmBot 3000 and shared the link with the whole office. By Wednesday, three people had tried it. By Friday, it was a ghost town.
Here's the short answer: a custom GPT is a configured version of ChatGPT with its own instructions, knowledge files and settings, built for one job. An architecture or interior design firm can build a useful one, like a proposal assistant or a firm standards assistant, in an afternoon if it starts narrow.
Narrow is the secret. Here are the other six.
1. One GPT, one job
The firm assistant that answers code questions, drafts emails, summarizes meetings and reviews specs ends up doing all of it a little badly.
The custom GPTs that get used do one thing well. Use this test to pick the job: it repeats every week, it's mostly text, and you have good past examples. A custom GPT for proposals checks all three boxes, which is why we'll use it as the running example.
2. Check your plan and data settings before you build
Before anyone uploads a single file, confirm which ChatGPT plan your firm uses, who can create and share GPTs, and how your data settings are configured. The ChatGPT Team vs Enterprise question matters here, because sharing options and admin controls differ by plan. Check OpenAI's current help docs on the day you build, since these details change.
3. Your knowledge files are the real product
Custom GPT knowledge files are where your firm's expertise lives. For a Proposal Assistant, upload:
- Your proposal style guide: tone, words you use and words you never use
- Two or three current templates
- Five to ten of your strongest past narratives and project descriptions
Before uploading, strip client names, fees and anything confidential. Clean, well-labeled files beat a giant pile of PDFs every single time.
Name files so the GPT knows what they are, like STYLE_GUIDE_2026.docx or EXAMPLE_Healthcare_Narrative_01.docx. Remove old versions completely; a GPT doesn't know which of two conflicting templates is current. Keep examples you'd proudly submit today, not the ones from 2014 that were fine at the time.
4. Write instructions like you're onboarding a new hire
Vague custom GPT instructions produce vague output. Be specific about the role, the task, the format and the limits. For example: You are the proposal assistant for [Firm Name], a commercial architecture firm. Draft project descriptions and narrative sections using the style guide and past examples in your knowledge files. Match our tone: clear, confident, no jargon. Keep each project description under 150 words. Never invent project data, square footage or awards. If information is missing, write MISSING and ask for it. Never write the win theme; flag where the principal should add it.
That last line keeps the part that wins work in human hands.
5. Test it with real work and a scoring sheet
Run five real tasks through it before launch. Use actual work from the last month, not easy softballs: one long project description, one short one, a tricky healthcare pursuit, a renovation and something with missing information. Score each output from 1 to 5 on accuracy, tone and how much editing it needed. If most outputs need heavy rewrites, fix the knowledge files or instructions before anyone else sees it. First impressions decide whether your team ever opens it again.
A good bar: if four out of five drafts need only light edits, you're ready. If the GPT invents facts even once, tighten the instructions and test again. A proposal assistant that makes up an award is not a time saver. It's a liability with a friendly interface.
6. Give it an owner or watch it die
Rollout is where most custom GPTs quietly fail. Assign one owner who updates the knowledge files every quarter, collects feedback in one channel and retires outdated examples. Then show the team exactly when to use it, ideally in a five-minute demo on a live proposal.
Once that works, consider three more narrow builds: a spec section checker, a client email drafter in your firm's voice, and a meeting-minutes-to-action-items assistant.
7. Know when you've outgrown it
Custom GPTs are a fantastic start. But they have limits. When the assistant needs live project data, different permissions per team or a record of every answer, you have outgrown a custom GPT. That is where a firm-specific AI system starts.
That's the jump from a helpful tool to AI for architecture firms that runs like infrastructure: connected to your systems, governed and maintained.
The bottom line
Learning how to create a custom GPT takes an afternoon. Building one your team actually uses takes focus: one job, clean files, clear instructions, real testing and an owner. Get those right and your Saturday project becomes a Monday habit.
Ready for an assistant that knows your firm?
DesignHub Solutions has supported AEC firms since 1996, and today we run AI operations inside a 400+ person design firm. Whether you need help building your first custom GPT or you're ready for firm-specific AI, we can help you get it right the first time.
Book your free 30-minute AI Readiness Call with DesignHub Solutions. Call: 1 (305) 290 1002. Visit: www.designhub.solutions.
Frequently Asked Questions
How do I create a custom GPT for my firm?
Pick one narrow job, prepare clean knowledge files, write specific instructions, configure sharing and data settings, test with five real tasks and assign an owner before rolling it out.
Do I need a paid ChatGPT plan to create a custom GPT?
Creating and sharing custom GPTs depends on your ChatGPT plan, and the options change over time. Check OpenAI's current help docs for your plan before you start.
Is our firm's data used for training?
It depends on your plan and data settings. Business plans typically offer more data controls, but confirm the current terms in OpenAI's documentation and set a firm rule before uploading anything.
Can the whole team use the same custom GPT?
Yes, if your plan allows sharing within your workspace. Assign an owner, keep the knowledge files current and collect feedback in one place.