I've spent 10+ years in corporate sales. Red Bull. Pepsi. Then founded and scaled a marketplace startup.
And I hated every minute of the selling part.
Not because it didn't work, it did. But it always felt like a tax on my time. Every day was the same impossible choice: build the product or go-to-market. The pipeline doesn't fill itself. But the moment you stop building, you fall behind. I wanted a tool I could just turn on and leave alone. It didn't exist, so my co-founder Aleksei and I built it.
This is the story of building Wayy.ai, a self-driving LinkedIn sales agent, and the hard lessons from 100,000 outreach messages.
Version 1 was a copy of something that already existed
Our first mistake was classic: we did market research, ran customer interviews, and built exactly what customers said they wanted. The result was a LinkedIn sequencer with manual list uploads, campaign controls, and a dashboard full of metrics, a slightly worse version of tools that had been around for years with bigger teams and more funding. Nobody switched.
So we scrapped everything and started over.
The turning point: we stopped asking "what do customers want?" and started asking "what does a truly self-driving sales system look like?" We built for ourselves, for someone who wants to be a passenger, not a driver.
The self-driving car metaphor that changed everything
The most useful mental model we found: we're building a self-driving car, not a race car.
Our early users split into two groups. The first, "sales professionals", wanted steering wheels, gear shifts, and a cockpit full of controls. They wanted to tune every step, see every micro-metric, approve every decision. We spent months trying to satisfy them with manual overrides and custom dashboards.
Then we realized: these are exactly the wrong customers.
A self-driving car doesn't need a steering wheel. The passenger needs a comfortable seat and a display that says "arriving at destination in 12 minutes." That's it. We fired the race drivers and focused entirely on passengers - founders and solopreneurs who said: "here's my product, go find me customers" and trusted us to do it.
That single ICP clarity unlocked everything. Instead of building a control panel, we started building a decision-making engine.
What we learned the hard way: 7 lessons
1. Fire wrong customers - especially the paying ones
This is the most painful lesson. Early paying users are psychologically hard to fire. But customers who want full control will drag your roadmap toward the wrong product. Our right customers gave us upfront green light: "make decisions, I trust you." That trust let us build the actual core - autonomous decision-making, instead of endless settings pages.
2. Start from vision, not from research
The first version came from interviews. It was mediocre by design. The second version came from a question: what's the most ambitious thing we could build? A fully autonomous sales agent that anyone, no sales background, no technical skills, can turn on and get results from. We built that instead. We love the product now. Customers love it. It's much more fun to build something genuinely hard.
3. Define your three core values and cut everything else
With a small team and limited runway, you can't build everything. We picked three non-negotiables:
Everything else - conversion dashboards, A/B test controls, detailed targeting UI - is noise for our user. Cut.
4. Build only what you can't buy
We started building everything ourselves: the LinkedIn automation layer, the dataset, the agentic orchestration, the multi-agent infrastructure. Progress was painfully slow and our attention was constantly split.
We switched to a simple rule: only build what doesn't exist on the market. We started using third-party services for data, infrastructure, and automation where good enough options existed. It added costs, but it forced us to identify what's actually unique about Wayy - our decision-making algorithms and UX flow. That's the only thing we build ourselves now. We can optimize costs later, once we've proven the concept works.
5. Cut until it hurts (then cut more)
The genius solution is usually the simplest one. Elon Musk's first principles of engineering: if you're not occasionally adding things back in, you're not cutting aggressively enough.
For a self-driving product, this means asking: does a non-sales, non-technical solopreneur need to know what A/B test won? What their accept rate is? What a conversion funnel looks like? No. They need to know: "Wayy found you 3 interested prospects this week." Everything else is noise. Cut it.
The specific failures (these cost us months)
Finding the right customer took way longer than expected
We discovered three customer segments the hard way:
Once we stopped marketing to anyone else, revenue became consistent.
Free users are a trap
We accepted free users early on. They don't care. They sugarcoat feedback. They point you in the wrong direction because they have no skin in the game. The moment users pay, they give you the hard feedback that actually improves the product. Free users are a false signal.
Pay-per-result was the right idea at the wrong time
We ran a pay-per-result experiment early. The market reception was our best announcement response ever - people genuinely want to pay for outcomes, not software subscriptions. But we weren't ready: the tracking infrastructure wasn't mature, and we couldn't sustain the model financially.
We pulled back to a simple 14-day trial + $49/mo plan. New cohort converted to paid immediately. We'll return to pay-per-result once the infrastructure can support it.
Browser extension vs. cloud
We built on a browser extension first. Critical mistake for our audience. Non-technical solopreneurs struggle to keep extensions connected - some rarely use a desktop at all. We had a persistent problem: large chunks of our user base had disconnected extensions and the agent simply wasn't running.
Moving to cloud fixed this completely. The agent runs whether the user's laptop is open or not. It's also more honest to the "self-driving" promise - a self-driving car doesn't need you in the driver's seat.
The death spiral problem in LLM-driven targeting
This one was subtle and almost killed campaign performance without us noticing. When left to optimize on its own, our LLM-driven targeting would narrow its search criteria over time - getting more and more specific until the candidate pool hit zero and the campaign stopped sending anything.
We had to build explicit self-correction mechanisms: constant A/B experiments running at every level (prompts, personas, targeting parameters, message sequences), with automatic resets when any parameter reaches a dead end. This self-optimization loop is now the most critical part of the system. Without it, the agent quietly fails.
Trust is fragile - one wrong connection destroys it
I remember the first time I used Tesla's self-driving mode. It's uncomfortable. You want to grab the wheel. And I know people who disabled it permanently after the car made one mistake.
Same dynamic for software. The three mistakes that caused immediate churn for us:
We added multiple verification layers before every connection request and multiple checks before every message. It slowed things down slightly but churn from trust failures dropped dramatically. Trust takes months to build and one bad message to destroy.
One question we get constantly: is it safe to use on LinkedIn?
LinkedIn is aggressive about automation. I get why people ask.
Our take is simple: if you create value for every stakeholder, the model is sustainable. So we deliberately stay below LinkedIn's own recommended limits for daily actions. We never do anything a regular user wouldn't do themselves on the platform - we just do it more consistently and efficiently than a typical human would.
The way we see it, Wayy actually helps LinkedIn. Our users become more active on the platform, which drives more paid LinkedIn subscriptions and reactivates accounts that hadn't been used in months. That's a win for LinkedIn.
And because of our rigorous vetting process before every connection request - we only reach out when a profile is highly likely to need the product or service being offered - the prospect on the other end gets a relevant message instead of spam. That's a win for them too.
Value for our user. Value for the prospect. Value for LinkedIn as a platform. That's the only model that works long-term, and that's the one we're building.
The results after 100,000 connections
The baseline for a human SDR doing cold LinkedIn outreach is 0.8%-2% connection-to-interested-reply conversion. We just crossed that threshold as our average.
But some categories are well above it:
Industry | Conversion Rate |
Marketing consulting | 10.0% |
Computer training services | 8.5% |
Independent artists / writers / performers | 5.1% |
Application software | 4.9% |
Online SaaS products | 4.7% |
Outpatient care services | 4.2% |
Telemarketing services | 3.9% |
Elderly / disability care services | 3.9% |
Data processing and hosting | 3.8% |
App software publishing | 3.7% |
We also tested Wayy for fundraising. In two weeks, 250 connection requests got 11 interested replies and 1 actual angel check from a highly relevant investor.
Currently running Wayy for BD outreach to find community managers and influencers. Already booked two meetings with community leaders with 100k+ audiences.
Where we're going
The goal is to expand beyond sales - toward a full AI business co-founder. Something that handles sales, BD, fundraising, and eventually more, so that anyone - no MBA, no sales background - can run a company.
We're early. We're a small team. But the agent is working and we're keeping you posted.
Happy to answer anything about the technical architecture, the LLM self-correction approach, the targeting logic, or the business model experiments. - Leo and Aleksei
wayy.ai
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。