The AI meeting co-pilot: a new category for design sales

A typical first meeting between an interior designer and a prospective client runs 45 to 60 minutes. During that time, the client describes their home, their budget (sometimes vaguely), their style preferences (often contradictory), their timeline, which rooms matter most, and a dozen smaller details about flooring preferences and whether they want a kitchen island.

The designer is doing three jobs simultaneously: building rapport, gathering requirements, and mentally estimating whether the project is financially viable. Most of the time, they're scribbling notes on a pad or into their phone, hoping they'll remember what "the client liked but maybe not in that colour" actually referred to.

After the meeting, they sit down and try to reconstruct what happened. They write a follow-up email. They put together a rough quote. They decide whether the lead is worth pursuing or if the budget doesn't match the wishlist. This post-meeting admin often takes 30 to 45 minutes on top of the meeting itself.

The result is that every consultation generates roughly 90 minutes of total time, and a lot of what the client said gets lost or simplified in the handoff from memory to notes to spreadsheet.

A meeting co-pilot is software that runs on the designer's phone or laptop during a client meeting. It listens to the conversation. It processes what's being said. And it surfaces information, prompts, and estimates to the designer in real time. The client doesn't see any of it.

Think of it as a knowledgeable colleague sitting next to you, one who's read every quote your studio has ever sent, knows Singapore market rates for every renovation item, and can whisper suggestions in your ear without the client noticing.

In concrete terms, here's what happens during a session on Open Maison's co-pilot:

The designer opens the co-pilot and acknowledges a consent prompt (the client should be informed that meeting notes are being captured, though the AI cards are only visible to the designer). The system starts recording audio.

Every five seconds, the audio is captured as a standalone WebM file, converted to base64, and sent to the server. There, Gemini 2.5 Flash transcribes the speech into text segments tagged by speaker. The full recording never leaves the server; only the transcription is stored.

Every third audio chunk (roughly every 15 seconds), the context engine runs. It takes the recent transcript, the running conversation context, and the studio's project data, then produces several outputs: AI coaching cards, deal signal updates, quote item estimates, and style keyword detections. Each of these feeds a different part of the designer's display.

The co-pilot surfaces information as cards. Each card has a type, a priority score from 1 to 10, and a brief explanation. The system generates at most three cards per cycle to avoid overwhelming the designer. Here's what each type looks like in practice:

• Whisper cards are coaching prompts. If the client mentions a tight timeline, the card might say: "Timeline concern detected. Consider mentioning your express renovation package." These are the most common card type.
• When the conversation touches a scope area that has common add-ons, an upsell card appears. Say the client wants new kitchen cabinets — the co-pilot surfaces: "Premium pull-out organiser with soft-close, 65% acceptance rate, ~S$2,800." That acceptance rate comes from aggregated data across the studio's past projects.
• Objection cards trigger on negative sentiment. When a client says something like "that sounds expensive" or "I'm not sure about the timeline," the card suggests a response strategy. These tend to be short and direct — a reframe, not a script.
• Alert cards are the highest-priority type. Budget mismatch is the most common trigger: the client describes a full-home renovation but mentions a budget that would barely cover the kitchen.
• Every time the client describes a new scope item, a quote update card appears with a cost estimate. "Kitchen island mentioned, estimated S$3,000–S$5,000." The estimate pulls from a lookup table of 30 Singapore market rates, built from averaged Qanvast platform data.
• Finally, visual cards suggest showing the client a portfolio piece or generating a mood board when style keywords appear. "Client mentioned Scandinavian and warm wood tones. Show portfolio match?" There's a 90-second cooldown between these to prevent them from becoming annoying.

Cards are ranked by a composite score: 40% AI-assigned priority, 25% recency, 20% type urgency (alerts rank highest, quote items lowest), and 15% engagement rate (how often the designer interacts with that card type). This means the most relevant, most recent, most urgent cards always float to the top.

Behind the cards, two other systems run continuously.

The deal scorer tracks six dimensions of how likely the lead is to convert: Budget (weighted at 25%), Trust (20%), Decision-making authority (20%), Timeline (15%), Style alignment (10%), and Scope clarity (10%). Each dimension is scored from 0 to 1 based on what the AI picks up from the conversation. The weighted total maps to a 0-100 score and a tier: Hot (80+), Warm (60-79), Cool (40-59), or Cold (below 40).

When the score crosses a tier boundary, the co-pilot generates a whisper card. For a Hot score: "Customer is ready to commit. Suggest next steps such as timeline confirmation and contract signing." For a Cool score: "Build trust with a portfolio piece. Try showing a similar completed project." There's a two-minute cooldown between these tier-change whispers so the designer isn't bombarded during a fluctuating conversation.

The shadow quote is a running cost estimate built from everything the client mentions. As the client describes each room and each item, the system matches keywords against Singapore market rates and adds line items to an invisible quote. By the end of the meeting, the designer has a rough estimate, grouped by category (Kitchen, Bedroom, Living Room, Bathroom, Common), without having opened a spreadsheet.

The shadow quote also tracks headroom: if the client stated a budget of $80,000 and the running estimate is at $72,000, the designer can see there's roughly $8,000 of room. If the estimate exceeds the budget, an alert card fires immediately.

When the designer ends the session, the system generates a post-meeting brief within 60 seconds. The brief is built by feeding the full transcript, all AI cards generated during the meeting, the shadow quote items, the deal score history, and the customer context into Gemini 2.5 Flash with a structured prompt.

The output includes:

• A narrative summary of what was discussed
• A customer profile (budget, preferences, decision factors)
• The final deal score with a breakdown across all six dimensions
• A quote estimate with per-category subtotals
• Suggested next actions (specific to what came up in the meeting)
• A drafted follow-up message ready to send
• Identified upsell opportunities with estimated values

The designer can save the brief to the lead record with one tap. That action updates the lead's details, attaches the quote estimate, and creates a follow-up task with the suggested timeline. The entire post-meeting admin, which used to take 30-45 minutes, collapses to a few seconds of review.

The co-pilot works well in a specific context, and we should be clear about where it doesn't.

Audio quality matters a lot. The system relies on Gemini's speech-to-text, which works best with clear audio from a single direction. A quiet meeting room with two people is the ideal case. A noisy cafe, a site visit with construction sounds in the background, or a group meeting with four people talking over each other will produce poor transcription and unreliable cards. We've added an anti-hallucination prompt ("Do NOT invent speech. If silent, return empty array.") but the quality of the AI output is directly tied to the quality of the audio input.

Accents are still a challenge too. Gemini handles standard English and Mandarin well. Singlish, with its mixed vocabulary and tonal patterns, is less reliable. Arabic accents in our Dubai testing produce noticeably more transcription errors. We expect this to improve as the underlying models improve, but today it's a real limitation.

What about the cost estimates? They are approximations. The shadow quote pulls from a static lookup table of 30 Singapore renovation item categories with min/max/typical price ranges. These are averaged from Qanvast platform data — useful as ballpark figures during a conversation, but they are not a substitute for a proper quote.

Privacy requires care. Recording a client meeting raises obvious questions. Our implementation requires the designer to acknowledge a consent screen before starting, and we recommend informing the client that notes are being taken with AI assistance. The full audio is transcribed and discarded; only the text is stored. But the ethical responsibility for how this tool is used sits with the studio, not with us.

If the microphone fails, the environment is too noisy, or the designer prefers not to record, the co-pilot can run in manual mode. The context engine runs on timer ticks rather than audio chunks, and the designer gets deal scoring and quote tools without transcription. Less useful, but it's there as a fallback.

We built this tool because we think the interior design consultation is one of the highest-value moments in the entire project lifecycle, and it's also the moment with the least software support. Designers are expected to sell, listen, estimate, and remember, all at once. The co-pilot makes the designer's own judgment better-informed.