How Mai Faces Leverages AI in Talent Agency Discovery in 2026

It is Monday at 10:04 AM. A production company in Bangkok is casting a lifestyle brand video for a Japanese skincare launch. The casting director opens three WhatsApp threads with three local agencies, requests "fresh, approachable, late-twenties, Thai-Japanese or ambiguous Asian, warm smile, confident but not polished." Agency one sends 40 headshots in a PDF, 90 percent of which miss the brief on vibe. Agency two sends a Google Drive with 200 photos and no filtering. Agency three replies with "we are preparing the shortlist, Monday EOD." The shoot is Thursday. The director spends six hours scrolling PDFs and ends up on Instagram DMing talent directly, which the agencies are contractually prohibited from allowing.

That is the shape of casting in 2026. The brief is a vibe, the matching function is a human agent flipping through portfolio PDFs, and the production company burns a day of pre-production because the agency cannot translate "warm but confident" into a tagged query. The gap between the brief and the shortlist is the entire pre-production bottleneck, and the agency bullpen cannot staff it fast enough to close. The booker is reading briefs with a human memory against a 1,200-person roster.

The queue nobody staffed

Every talent agency in Bangkok runs the same roster workflow. 400 to 1,500 models and actors, each one with a headshot folder, a few lifestyle shots, a measurement sheet, a rate card, and a credits list. The agency's booker reads a brief, flips through the roster from memory, pulls 30 to 60 names into a PDF, and emails the client. The booker's taste is the entire matching function, and the booker is one person who sleeps eight hours a night. The agency's revenue depends on the booker touching every booking at a 15 percent commission.

The brief is a vibe. "Warm but confident, approachable but not generic, late-twenties, Thai-Japanese or ambiguous Asian, works in a kitchen as much as on a yacht." The booker's job is to read that sentence and compress the roster of 1,200 names into 30 that match. The booker does the work in their head, pulls from the 40 people they remember best, and sends a shortlist that misses the brief on half the names because the booker's memory is the entire taxonomy. The brand catches the miss on the second lookbook and starts DMing talent on Instagram.

The function that reads a vibe-based brief, matches against a tagged roster of 1,200 names on 40 attributes (demographic, feature, style, vibe tags, availability, past-brand fit), surfaces the top 20 inside two minutes, and auto-generates a lookbook the client can approve is open. The agency cannot staff it because the agency revenue is the booker's commission on bookings the booker personally touches. The booker's seat is the function and the function is the bottleneck. The agency bullpen will not grow fast enough to close the gap.

What an AI talent agency function looks like

Mai Faces runs the matching function that the booker bullpen cannot staff at that commission structure. The platform sits between production companies and a tagged roster of Bangkok talent, runs the match on AI vibe-tagging against the brief, auto-generates a lookbook, and routes the booking inside the platform. The casting director sees 20 matched faces inside two minutes of posting the brief instead of waiting four days for a PDF. The function runs on six parts, each one on cadence against a growing tagged roster.

1. Talent intake. Every signed talent submits headshots, lifestyle shots, voice samples, measurements, availability calendar, and past-brand credits. The platform tags against 40 attributes using a vision model (feature, style, vibe, demographic, aesthetic, voice, body type).
2. Brief parser. The casting director posts a brief in natural language with reference images. A model stack reads the brief, extracts the vibe tags, the demographic target, the style target, and converts it into a match query.
3. Vibe match engine. The query runs against the tagged roster with vector similarity on both the written vibe tags and the reference image features. The engine surfaces the top 20 names ranked on fit.
4. Auto-lookbook. The platform assembles a lookbook from the top 20, with each talent's three best lifestyle shots selected on fit to the brief, measurements, availability, and rate card.
5. Booking flow. The director picks three to five, routes the brief to the talent, confirms the shoot date and the usage rights, and locks the booking with payment escrow.
6. Credit log. The shoot ships, the credits log updates, the talent's reliability score updates on both sides. The next brief reads a fresher roster.

The booker used to run parts two, three, and four from memory. The function runs all six against a tagged roster and the booker's seat moves to roster growth, talent development, and the taste layer on new intake. The margin on each booking improves because the booker is no longer the match engine. The volume climbs because the lookbook is a two-minute surface instead of a four-day email thread.

Why vibe-matching is harder than demographic matching

Every talent platform launched in the last five years ran a demographic filter. Age, height, hair color, ethnicity. The filter worked for 30 percent of briefs and failed for 70 percent, because the briefs that come from brands are vibe briefs. "Warm but not generic." "Confident but approachable." "Fresh but experienced." The demographic filter cannot match a vibe, and the booker's memory was the entire matching function because nothing automated could read the brief correctly.

The vibe is a vector in a feature space, not a filter on a dropdown. Modern vision models read a reference image and encode the vibe into an embedding. The platform tags every talent's lifestyle shots into the same embedding space. The match becomes a nearest-neighbor query against the tagged shots, which means the brief "warm but confident" lands against the ten faces the model has already encoded as closest to that vibe.

The compound is on the roster side. Every new talent that joins the platform adds 10 to 30 tagged images to the embedding space. The match quality improves as the roster grows, which means the brand's second brief lands against a sharper match than the first, which means the brand books more campaigns through the platform, which means the talent earns more and refers more talent. The flywheel is the roster growth times the embedding density, and the embedding density is what the booker's memory could not scale.

The second flywheel is the auto-lookbook. The lookbook used to be a half-day of a booker's time per brief. The auto-lookbook is a template filled against the match engine's output and the talent's three best shots scored on vibe fit. The lookbook ships inside the browser tab, the director approves inside the browser tab, and the booker's time moves to the roster development function.

The unit economics against the booker bullpen

A Bangkok talent agency with three bookers running 60 briefs a month at 15 percent commission on $4,000 average bookings runs $108,000 a quarter in commission revenue against a three-booker payroll of $120,000 a quarter. The margin is negative on commission alone and the agency survives on retainers and talent-development fees. Each booker touches 20 briefs a month and the taste quality drops on brief 15 because the booker is tired. The 15 percent commission is a cap on volume, not a floor on quality.

An agency running the AI matching function on the same roster ships 180 briefs a month on the same three-booker payroll. The bookers' time moves from match to taste review on new intake, dispute resolution, and roster development. The platform fee flows at a flat rate on every booking, the commission flows at a lower rate because the match function is automated, and the brands book more because the lookbook ships inside two minutes instead of four days. The margin inverts because the volume climbs three times and the per-booking cost drops.

Read the case studies for the shape of a function that compressed a booker's four-day brief turnaround into a two-minute lookbook, and the services page for how the same operating pattern runs against the content function and the reporting function on B2B. The process page walks the 14-day sprint cadence that stands up the match engine and the lookbook template before the first brief lands.

What this maps to for every taste-driven marketplace

Talent agencies are one of eight taste-driven marketplaces where AI vibe-matching is replacing the human booker's memory. Casting directories, modeling agencies, voiceover rosters, music licensing catalogs, stock imagery, interior design portfolios, art galleries, wedding vendor directories. Each one runs a taste-based brief against a static database and the match function is a human's memory. The eight share the same operating problem and the same operating fix.

The playbook holds across the eight. Build the tag spec. Build the vision model intake. Build the brief parser against the tag spec. Build the match engine on vector similarity. Build the auto-lookbook template. Build the booking flow with escrow. The function compresses the booker's memory into an embedding space and the booker's time moves to the taste layer on the roster, which is the one part the model cannot replace.

The three questions to run against your taste-driven marketplace

If your business runs on a human agent flipping through a static database to match a vibe-based brief, three questions sort whether an AI vibe-matching function fits the shape. The checklist is honest about the roster depth, the brief flow, and the lookbook latency.

Can you tag your roster with vision embeddings? If every talent, product, or venue in your roster has at least 10 photos, the vision model can encode them. The embedding step is the first week of the sprint and the match quality depends on the depth of the image library per entry. The sprint stalls on roster depth, not on model quality.

Can your clients articulate a brief with reference images? If the current flow is "the client calls and describes the vibe over the phone," the brief parser has nothing visual to anchor against. The sprint starts with teaching the client to post briefs with three to five reference images, and the match quality jumps the moment the references land in the query.

Can you auto-generate a lookbook or shortlist inside the browser tab? If the current turnaround is "the booker emails a PDF in two days," the function is still bottlenecked on the human. The lookbook template is the second week of the sprint and the booker's seat moves from match to taste review. The latency between brief and lookbook is the single biggest buyer-side signal.

The vertical that answers all three the fastest is the one where the booker's memory stops being the matching function. Everyone else is paying three bookers to flip through PDFs while the production company DMs talent on Instagram the night before the shoot.