The problem Velho sells AI automation to established companies, but its buyers (CEOs, CFOs, COOs) are pitched constantly and ignore generic AI messages. The hard part is not writing the message: it is knowing which company has the problem right now, and proving it.
What we built Two halves that share one database (Twenty CRM): - IN · Velho CRM: type a company name, Y-tunnus or website. In about 90 seconds Velho reads the company's own website, its open job ads, its latest financial statements (finder.fi) and the Finnish trade register, then saves a score, verified evidence, who to contact and a suggested first message. - OUT · The Worst Monday Awards: people nominate the repetitive work that ruins their Mondays, see what it costs and try an automation demo. Every nomination (with consent) lands in the CRM as a lead with data no scraper can get: the process in the sufferer's own words, hours per week, team size.
How it works, and what's AI about it - Signals, not lists: companies hiring people for invoicing, order handling or data entry right now are paying for manual work today. That is the reason to talk now. - A fixed source waterfall first (fast, the same every run), then an AI agent reads it and picks evidence. You choose the engine: DeepSeek on OpenCode, or GPT-5.6 Terra/Luna on Codex, with the same tools served over MCP. - No hallucinations by design: every quote is checked word for word against the cached source, or rejected. The message may only state what a quote says; anything inferred must be asked as a question. - A decision model (TypeSafe Jev via Vercel AI Gateway) answers atomic questions (pain, automatable, urgency, size); code combines them with hard rules. Each company gets a next step: build a prototype, send a message, watch, or too small. - Everything is visible live inside the CRM: every step, source and finding as the agent works. - Velho never contacts anyone by itself: a person approves every message.
Real example: Normet Oy, researched live in 95 s with GPT-5.6 Terra. €179M revenue, 522 employees, a new sales-coordinator ad describing ERP order entry and invoice checks. Result: score 89, next step PROTOTYPE, with the brief "forward an incoming order -> validated ERP-ready order data".
Why it matters: 38% of Finnish companies already use AI (2nd in the EU, Statistics Finland 2025), so 62% still don't. And only about 5% of B2B buyers are in the market at any moment (the 95:5 rule, LinkedIn B2B Institute): Velho's app finds that 5% through public signals.
What's next - A pilot with Velho on 200 companies, with a control group to measure meetings booked. - LinkedIn Ads to the same account list (Matched Audiences), so buyers have seen Velho before the first message. - Auto-generated prototypes on the prospect's own data.
Code: https://github.com/JordiSama1/hackaton-op
