Fast or Custom, AI Gets You There - The Modern Agency Build Workflow

The Real-World Case Study: AI or Die Webinar Landing Page

Bogdan anchors the entire "custom flow" section around a real project: a webinar landing page built for Breezy itself, created in collaboration with Vito from Atarim. The session title "AI or Die" is reflected in the concept behind the page, which used a space/asteroid visual metaphor to communicate that agencies and businesses that fail to adapt to AI will go the way of the dinosaurs.

The project started with a standard marketing brief from Breezy's internal marketing team.

The brief contained the usual elements: the webinar title, the date and time, speaker information, and a small amount of supporting copy. This brief was compiled into a Google Docs document.

The finished landing page lives at breezy.io/AI-or-die-webinar and took a total of three to four hours from brief to published, including all design iteration, import, tweaks, animations, and final content entry.

The Custom Flow: Figma AI to Breezy

Step 1 - Conceptual Direction and Reference Gathering

Before opening any tool, the Breezy team made a creative decision about the design direction. They settled on a dark, outer-space aesthetic tied to the "AI or Die" theme: the idea that an asteroid is coming for businesses that refuse to adapt to AI, a deliberate reference to the extinction of the dinosaurs. This direction was deliberately a stretch concept, chosen to push the design capabilities of Figma AI.

To support this direction, the team researched visual inspiration on Dribbble and Behance, finding designs from the space exploration genre. These reference images were saved as screenshots for use inside Figma AI. Bogdan is careful to note that nothing was copied from these references; they were used purely as visual direction material.

Step 2 - Prompting Figma AI (Figma Make)

The team used Figma Make, the prompt-based AI interface inside Figma. The tool supports multiple underlying AI models: Bogdan confirmed that Claude, ChatGPT, and Gemini are all available. He selected Claude Sonnet, describing it as the best balance between speed and cost-effectiveness for this type of task. He notes that Claude Opus may produce marginally better results but did not find it meaningfully different for this use case, and suggests leaving the model on default is a reasonable choice.

The prompt they entered was intentionally brief. It included:

• An instruction to create a landing page for an upcoming webinar
• The title of the webinar
• A link to the Google Docs brief (which Figma AI was able to access)
• Some thematic direction for the design aesthetic Alongside the prompt, the team attached the Dribbble and Behance reference screenshots as image attachments within Figma Make. This combination of a text prompt, a linked document, and reference images gave the model sufficient context to begin generating design options.

Step 3 - Iterative Design Refinement in Figma

Figma AI does not produce a single result and stop. The process is conversational: the model generates a design, the user evaluates it, and further prompts refine or redirect the output.

Bogdan emphasises that this back-and-forth is expected and is part of the workflow, not a sign that something is wrong.

The early outputs from Figma AI in this project were predominantly minimalistic and light in tone, which did not align with the dark, atmospheric aesthetic the team wanted. Bogdan describes these as "too white" for the intended feel. This was resolved through prompting: the team explicitly directed Figma to shift toward darker tones and a more dramatic visual style. The tool responded to this instruction and iterated toward a design the team was satisfied with.

The team went through multiple candidates before landing on the final design. The chosen design incorporated the outer space concept, matched the dark palette they wanted, and reflected the content from the marketing brief. Speaker images and portraits were not included in the initial AI output because Figma AI did not have access to those image files.

The team added them manually in a separate step that took approximately twenty to thirty minutes of direct design work inside Figma.

Bogdan estimates that total design time inside Figma, inclusive of prompting, iteration, and manual additions, was around twenty to thirty minutes.

Step 4 - Exporting from Figma

Once the design was finalised inside Figma, the team needed to move it into Breezy for web publishing. Figma Make can export designs in multiple formats including HTML, but for this project the team chose to export as a JPEG file. The reasoning was practical: they wanted to test the performance of Breezy's AI image-to-website feature, which accepts a single image and generates a fully editable web project from it.

Step 5 - Importing the JPEG into Breezy AI

The JPEG export was uploaded into Breezy's AI generation tool. At the time of recording, this feature was not publicly available. Bogdan describes it as being in internal testing only, "a bit buggy," and notes that publishing it involves resolving a range of edge cases that can cause problems when the AI interprets complex design details from a flat image.

Breezy's AI processed the image and generated an editable project inside the Breezy Builder interface. The result was described as approximately 70-80% complete. Specific elements that did not transfer correctly included the logo (which was replaced with placeholder text rather than the actual logo asset) and interactive elements such as a speaker card flip animation and a FAQ accordion, neither of which could be inferred from a static image.

The parts that did work well included the overall layout, colour scheme, typography choices, the general section structure, and most of the static text content. The AI interpreted the page hierarchy correctly and produced a draggable, editable Breezy project that matched the visual structure of the Figma export reasonably closely.

Step 6 - Manual Finishing in Breezy

The remaining 20-30% of the work was completed inside the Breezy Builder editor. This included: Replacing placeholder logo text with the actual logo asset Building the speaker card flip interaction manually, since the static image could not communicate interactivity Populating the FAQ accordion with the correct text content (sourced from the marketing brief via copy-paste, approximately ten minutes of work) Adjusting padding, positioning, and spacing throughout the page Adding CSS animations to bring the page to life Bogdan is explicit that this finishing work is not a failure of the AI but an expected and appropriate part of the workflow. The value the AI provides is in eliminating the time-consuming structural and creative groundwork, not in replacing the need for all human input.

The Fast Flow: Google My Business to Published Website

Concept and Use Case The fast flow is designed for situations where an agency or freelancer has minimal input from a client. Where the custom flow requires a detailed brief and a creative direction, the fast flow needs little more than a business name or a Google My Business profile link. It is particularly suited to small-to-medium business websites: local service providers, retail outlets, hospitality businesses, and similar clients who already have a populated Google Business profile.

How It Works Inside Breezy's new project flow, there is an option to generate a website from a Google My Business profile. The user provides the business name (or a direct link), and Breezy queries the Google Business directory to retrieve: Business name and category Phone number and address Customer reviews (real text from actual reviewers) Gallery images uploaded by the business or its customers Breezy then uses this data alongside ChatGPT (which Bogdan identifies as the model powering the text generation in this part of the product) to generate copy, section text, and page structure. The tool allows the user to configure the pages they want (e.g., homepage, inventory, finance) before clicking to build.

Breezy supports 30 languages for website generation in this flow, making it viable for agencies serving non-English-speaking markets.

Live Demo: Gravity Autos To demonstrate the fast flow live, Bogdan used a real business as his example: Gravity Autos, a luxury car dealership located in Hollywood, Florida. He had set this up prior to the session and walked through the results in real time.

The generated website included:

• A hero section with AI-generated copy describing Gravity Autos as specialising in "premium exotic cars," which accurately matched the business category.
• Multiple content sections populated with relevant business information.
• Gallery images pulled directly from the Google Business listing.
• A customer testimonials section using real review text from genuine customers, though the reviewer avatar images were not real, being AI-generated or stock images.
• A contact section with the correct address pulled from Google.

Bogdan also demonstrated the ability to add new sections mid-session. As an example, he added a pricing section on the fly. The AI generated the section with text and layout, though he acknowledged the pricing figures were not accurate; the system produced a low illustrative price that did not reflect the actual cost of luxury vehicles, which he suggested might be interpreted as a rental price. He was clear that this section would need to be manually corrected before publishing, which is again consistent with the 70-80% framing.

Moving into the Editor Once the fast-flow generation is complete, the user can take the generated site into the full Breezy Builder editor for fine-tuning. From there the workflow is identical to any other Breezy project: drag-and-drop editing, section management, colour adjustments, text editing, and page management through a simple menu system.

The generated sites include responsive layouts as standard. Bogdan showed both the mobile and tablet views of the Gravity Autos site, describing the responsive output as requiring "no need for custom or too many tweaks."

Bogdan's Philosophy on AI and Agency Work

The closing Q&A between Bogdan and the host (Andrew) produced the most substantive philosophical content of the session and is worth covering in depth because it frames the practical demonstrations in a broader professional context.

Honest Ambivalence About AI Bogdan opens his answer to the "biggest tip" question with a confession that he is "not the hugest fan of AI." He clarifies that he believes it is the future and loves the technology, but he is concerned about the pattern of agencies and individuals building their entire workflow around AI at this stage of its development. His concern is that the current capabilities, while impressive, still produce outputs that require meaningful human correction, and that treating AI as a complete replacement for craft judgment is premature and potentially damaging to the quality of work.

The Risk of Homogenisation The most pointed argument Bogdan makes is about the risk of every website looking the same. His reasoning is straightforward: AI models are trained on the same datasets, they observe the same high-performing design patterns, and they converge on the same solutions. If every agency uses AI to generate designs without applying distinctive human creative judgment, the result will be a web full of websites that look and feel indistinguishable from one another. They may convert well because AI is good at optimising for known conversion patterns, but they will lack identity, personality, and brand differentiation.

He connects this directly to agency competitive advantage: the distinctive design style that an agency builds its reputation on is at risk if that agency delegates all design decisions to an AI.

The 80/20 Rule Bogdan's recommended operating model is to use AI for the structural and generative heavy lifting, targeting a point where the output is 70-80% of the way to completion, and then

apply human craft, judgment, and creative refinement for the remaining 20-30%. He frames this as "giving AI a lot of stuff to do, looking at how it performs, and then taking the last 20-25% into manual work to give it the human touch and personal vibe."

This is not presented as a temporary compromise while AI matures; it is presented as the appropriate and philosophically sound division of labour between machines and human practitioners right now.

The Bespoke Suit Analogy To make the case for preserving human craft, Bogdan invokes the analogy of bespoke tailoring. A custom suit, made to measure by a skilled tailor, commands significantly more value than a machine-produced garment available on every high street. As AI-generated websites become ubiquitous, the handcrafted, bespoke website, made by an agency that brings taste, client understanding, and creative distinctiveness, will carry increasing premium value.

Andrew extends this analogy: a client who commissions an agency is fundamentally placing their trust in that agency's judgment. The AI gets you 70% of the way there efficiently, but the client is paying for the agency's expertise in directing, refining, and completing the remaining work. That relationship and that trust are not things AI can replicate.

Tool and Pricing Notes

Breezy Builder Visual website builder with drag-and-drop editing AI generation available from Google My Business (live in production at time of recording) AI generation from image/JPEG upload (in internal testing at time of recording, not publicly available) Mobile and tablet responsive layouts generated by default Supports 30 languages for AI-generated website copy Pricing (as described by host Andrew): approximately $400 per year for un-hosted plans covering roughly 100 sites; hosted plans priced higher but provide an all-in-one solution Breezy is an event sponsor Figma Make (Figma AI) Prompt-based AI design tool inside Figma Supports Claude, ChatGPT (GPT models), and Gemini as underlying models Claude Sonnet was used for this project; Bogdan describes it as a good balance of speed and cost-effectiveness

Accepts image attachments as reference/inspiration material alongside text prompts Can access linked Google Docs documents as context Exports designs as HTML or image formats (JPEG used in this project) ChatGPT Identified by Bogdan as the model powering text generation in Breezy's fast flow (Google My Business integration)

Session Structure Summary

• Introduction by host: brief, host introduces Bogdan and his role at Breezy.
• Custom flow demo: the majority of the session, covering the Figma AI workflow, JPEG export, Breezy AI import, and manual finishing.
• Fast flow demo: mid-to-late session, covering the Google My Business integration and a live Gravity Autos demo.
• Q&A with host: closing segment, tips for agencies and philosophy on AI and human craft.

About the speaker Bogdan Condurache Co-Founder and Chief Product Officer, Breezy Builder

Bogdan Condurache is co-founder and Chief Product Officer of Breezy Builder, a website building tool. He focuses on how AI changes the practical workflow of building sites, from fast templated builds to fully custom work.