- 01Vacancies collected
- 02Cleaned, rewritten, published
- 03Candidate replies to the job
- 04A person confirms dates and documents
End of June to October 2026. Sources: Google Search Console, PostHog, the community member count and the ad accounts.
Curve traced from Vlad's report; the labelled points are exact.
20× in 16 weeks, and still about 230 new members a week. 10,187 jobs posted to the groups. Curve traced from Vlad's report; the labelled points are exact.
August's AI videos were optimised for clicks. From September the ads optimised for Braços' own qualified-employer event, and October added a video and landing page per trade.
- 89.8% AI agents
- 7.7% The founder, through an AI session
- 1.7% Code pipelines
- 0.8% A person typing on a phone
About 2,660 people wrote in over four months: roughly 1,900 workers and 290 employers.
About Braços
Braços is a marketplace in Portugal that connects people looking for work with employers who need them fast, across construction, cleaning, logistics, restaurants, healthcare and eight more sectors. The name means “arms” in Portuguese, as in a helping hand.
A small team with no marketing department, and the problem every marketplace has on both sides at once: attention, trust and community.
The brief
Grow both sides of the marketplace without a marketing team. Reach job seekers and small contractors where they already look, give them communities they trust, and answer every message fast.
The question: how much of that can agents run, and where does a person have to stay in charge?
What we did
Vlad built five steps that run as one loop. Your ideal customers already leave a trail of public data; point AI at it.
- Public data. Eight public job boards read every hour in working hours, about 20 Facebook job groups of up to 200,000 members each, and a public WhatsApp jobs channel. 65,516 vacancies collected so far, about 275 new every day. 29% of Facebook job posts were only a photo, so a local vision model reads them and that part of the market isn't lost.
- An AI pipeline. Every post is classified (a job, a person looking for work, or noise) every five minutes, matched against repeats by phone, author and then text (13.5% were repeats), rewritten in its own words with phone, pay and company kept word for word, and published every 60 seconds to the website and the right WhatsApp groups. About $0.50 per 1,000 posts. A stronger model graded the rewrites.
- Owned communities. 130 WhatsApp groups, one per sector and region, so a cleaner in Faro never sees building jobs in Braga. No links in posts: people reply to the job, not to an ad. WhatsApp has no spam tools, so Vlad built one: an AI moderator reads every untrusted member's message first. Clean messages go through, and three clean messages make the member trusted. Spam is deleted and the sender is banned from every group at once.
- Content that search and AI cite. Single job pages answer one question each, which is what AI search picks up. The site ships an llms.txt file, a Markdown copy of every page, “Quick answer” blocks, and lets AI crawlers in.
- An AI inbox for both sides. Two WhatsApp lines, one for workers and one for employers, each with an intake and qualification script. A dispatcher agent works both sides of each order, and every fact is saved to the database in the same turn.
- Paid ads, run by agents. Every ad is made by agents in about 40 minutes: a presenter image, a voice, lip-sync, render. Campaigns are created paused, and a person approves budget, trade and region.
What’s running now
Where people stay in charge
Agents only reply; they never start a chat. Dates, documents and commitments go to a person, and a person approves every ad's budget, trade and region before it runs.