Tear down a multimodal feature and assess whether the modality earns its place.

Grayvine turns a set of listing photos into a narrated video walkthrough automatically, no camera crew, no agent voiceover. Ottoline Farr sells homes in a mid-size suburb, and Grayvine is the reason her listings started getting more views this spring. It's also the reason she almost got a fair-housing complaint.

How to answer this, stage by stage

Eight moves. The fourth is a commitment, said plainly, before any hedging gets a chance to creep in.

Let's learn

Say we build a tool that turns a folder of listing photos into a two-minute narrated video, with a voice describing each room as the camera pans across the photo.

Before Grayvine, Ottoline's listings had photos and a text description she wrote herself, maybe 90 seconds of a buyer's attention per listing on the portal.

With Grayvine, average view time on her listings roughly tripled, to just over four minutes, since buyers watched the whole narrated video instead of skimming photos.

The turn: the extra view time was never the real story. The real story is what the narration script says once it runs out of things it can actually see in the photo.

At its worst: a buyer schedules a viewing specifically for a described feature that doesn't exist, and the brokerage's name ends up on a housing-discrimination complaint over a line nobody at the company ever wrote or reviewed.

What I would leave alone: narrating a genuinely visible feature, a bay window, hardwood floors, a fenced yard, all clearly in frame. That part of the script never caused a problem and shouldn't get more restrictive than it needs to.

The lesson: a multimodal feature earning its place isn't about whether it's impressive. It's about whether the new modality can fail in a way a caption never could, and whether anyone built a fence around that failure before it shipped.

Now here is the same thing as a story

The short version above is what you'd say in the interview room. Read this one for how close the actual complaint came.

Ottoline has sold homes for nine years and can tell within a showing or two which buyers are serious.

One of these two errors a buyer forgets by the next listing. The other one gets a lawyer's attention.

The good weeks were genuinely good. Grayvine's videos made even her smallest listings look considered, and two buyers separately told her the walkthrough was why they scheduled a showing at all.

The script-writing step, second in line, is the one nobody was reading before it went to voice.

The trigger was one small den, four photos, no closet visible in any of them. The script called it a "spacious home office with built-in storage." Nobody at Grayvine or the brokerage had written that line. The model had, filling a narration gap the same way it filled every other one.

The fixer-upper, the listing type with the least in frame to describe honestly, is exactly where the model had to invent the most.

A buyer's agent, not Ottoline, caught the line during a routine walkthrough of the video before a showing and flagged it to the brokerage's compliance contact, one day before the listing would have gone out in a wider email blast.

Two months from launch to a near-miss complaint. The script rule didn't exist until month four.

With the script locked to only what's visible in frame, and an agent sign-off step added before the voice renders, the same four-photo den now gets narrated as "a compact den, ideal for a reading corner," with the missing storage claim gone entirely and Ottoline reviewing the eleven-word script in under a minute.

Four guardrails, and the brokerage only had zero of them running before month four.

The old script had permission to fill in whatever made a listing sound complete. The new one only gets to describe what an agent could point to in the actual photo.

I approved letting the model fill narration gaps because every video looked equally finished, and that felt like quality. It took a buyer's agent catching one line, one day before it would have gone wide, to see that "sounds finished" and "is true" were never the same thing, and I'd let the first one stand in for the second.

PICK, in one screenNot a taste debate. PICK is what makes you name which error actually costs something.

Four letters, and cost asymmetry is the one that actually decides the answer.

The recap, one line per letter: position is keep the video with a locked script, impact is a shrugging buyer versus a brokerage's legal exposure, cost asymmetry is cheap-and-visible versus hidden-and-expensive, and kill criteria is a second regulator-facing complaint pulling the feature entirely.

And if you want to be sure it really works, try it somewhere elseSame four letters, a translation service instead of a brokerage.

Waybrook adds AI-generated subtitles and a dubbed voice track to short training videos for companies with non-English-speaking warehouse staff. Delphine Osazee manages localization for a mid-size logistics client.

Mapped onto PICK: position is keep the dubbed audio, but require the AI to flag any phrase it's less than 90% confident it translated correctly, rather than smoothing over the gap with a fluent-sounding guess. Impact: a worker who hears an awkward but accurate phrase loses a little polish. A worker who hears a fluent but wrong safety instruction risks an actual injury. Cost asymmetry: the awkward phrase is cheap, visible, and gets a shrug. The wrong safety instruction is hidden inside fluent-sounding audio and can genuinely hurt someone. Kill criteria: any single mistranslation involving a safety-critical instruction pulls automatic dubbing for that entire training module until a human translator reviews the whole thing.

Swap "in frame" for "confidently translated" and the same four guardrails hold for a dubbed safety video.

Swap the trigger and it still runs.
Speed: an interviewer caps you at sixty seconds. Say "keep the video, lock the script to what's visibly in frame, kill it if a second complaint reaches a regulator," and stop.
Cost: there's no budget this quarter for an agent sign-off step in the product. Say so honestly, and start with a simple flagged-word list that catches invented amenities automatically, even without a full review workflow yet.
The model gets better, for real: if Grayvine's script-writing model gets dramatically more accurate, that's still not a reason to remove the sign-off step, since a rare invented feature in a low-frequency event still carries the same fair-housing cost.

Where people run it wrong.
They judge a multimodal feature on how impressive the demo looks instead of on how the new modality specifically fails.
They treat "it saves the agent time" as the whole argument, without weighing what the hidden error costs someone else.
They fix the model's overall accuracy instead of restricting what it's allowed to claim in the first place.

How to use it live. When someone shows you a flashy multimodal feature, ask yourself one question first: what's the one thing this modality can now claim that the old, boring format never could. That claim is where the real risk lives.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0