back
workers.ai
EXTENDED version
About the company
Workers.AI built Brazil's first AI SDR platform, launched in 2024, before AI agents became the crowded default. Autonomous agents that qualify sales leads, with human-in-the-loop supervision, a knowledge base, and granular behavior configuration. Bootstrapped, and at break-even within a year.
—
Tags
I joined Workers.AI as Product Designer and Product Manager and turned a working prototype into Brazil's first AI SDR platform, built before agents became the crowded default they are now. When I came in there was a POC but no product: no backlog, no MVP, no flows, no specs. For a long stretch I carried both the product design and the product management, working as PM and PO while the company found its footing. This is the long version, with the full mechanics of what we built and how each part earned a company's trust.
The short version of the problem: by 2023, language models could do work that used to take a team, but sales still ran on human SDRs and basic chatbots, and both were breaking. The deeper problem wasn't capability. It was trust. No company hands its sales pipeline to an AI on day one, and most AI sales tools ignored that entirely.
Context and problem
The SDR role, the sales development rep who prospects and qualifies leads before they reach a closer, had become a structural weak point for most companies.
A junior SDR cost R$6,000 to R$9,000 a month before benefits, tools, and overhead. They took around 90 days to reach full productivity, and turnover ran high, by the founder's read of the market, around 40%. So companies spent months and real money getting someone productive, only to lose them and start over, and every departure took undocumented knowledge out the door with it. Performance also varied wildly from person to person, so results were inconsistent even on a good team.
Language models could clearly help. But the tools that existed weren't the answer. Most were chatbots: they answered questions but didn't do the work, didn't understand business context, couldn't follow a sales cadence reliably, and had no governance, no limits, no defined objectives. Worse, their interfaces were built for engineers, not for the sales managers who would actually run them.
Underneath all of it sat the real barrier, the one that shaped the whole product. Companies don't trust AI overnight, especially not with their pipeline, the thing that feeds their revenue. Any tool that ignored that distrust would fail no matter how capable it was.
The pain landed on four groups. Sales and operations managers dealt with constant turnover that destroyed continuity, and with execution that varied enormously between reps even under well-defined playbooks. Salespeople spent most of their day on repetitive work, emails, follow-ups, CRM updates, fighting tools that were supposed to help, and carrying a quiet fear that AI would take their jobs. Strategic leadership watched commercial operations eat a large share of budget while growth stayed capped by people: to double sales you had to double the team. And operational users, the RevOps analysts and tool managers, had no intuitive way to teach an AI about products, objections, or tone, and no governance layer to set rules and limits.
How I got to the product
The work started with research into a moving target: what a sales development process actually looks like, and how much it varies. There is no single SDR. Every company qualifies differently, talks differently, sells differently, so the first job was understanding that variability rather than pretending it away. Most of that understanding came from interviews, and it came fast, because the company itself had strong commercial people who knew the work cold. I learned the pattern from them.
Out of that research, one design method emerged, and it ended up shaping every major decision: take a process the user already understands from the human world, and translate it to the agent. I saw the pattern, and we built on it three times.
The first was the core interface. Technological literacy among these teams varied a lot, depending on the product they sold and the tools they already used, so an interface full of AI machinery would have lost half of them. But there was one mental model everyone in sales already had: hiring, training, and managing people. So the product spoke that language instead of the language of models.
The second was supervision. It wasn't in the first plan. It surfaced once I understood how uneasy users felt handing their pipeline to an AI, and that unease pointed directly at the answer: give them the control a manager expects over a new hire, graded rather than all at once.
The third was onboarding. Once the product worked, we needed people to adopt it, and again the move was to look at what was already familiar. Onboarding a human SDR, teaching them the product, the objections, the tone, was a process these teams knew by heart, so we mapped that same ritual onto setting up an agent. Onboarding came after we had a working product, which is the normal order of things, but the method for designing it was the one I'd used from the start.
The through line was consistent: the technology was new, so we anchored it in processes that weren't.
WHAT I BUILT
The product had to do two things at once: run real qualification work autonomously, and stay legible and controllable enough that a skeptical manager would let it. We shipped it incrementally, starting with a focused MVP, lead qualification, basic personalization, and the knowledge base, then layering in granular configuration, supervision, and omnichannel as trust and usage grew. Here is each piece and how it worked.
Design it like managing a team, not debugging a model
I could have exposed the machinery, models, parameters, tokens, embeddings. Technically honest, impressive to an engineer, and useless to the person actually buying and running the tool. Sales managers don't think in models. They think in hiring, training, and managing people.
So the entire product borrowed that mental model. You hire a Worker. You train it on your business. You set its goals. You manage it, and you can step in when you need to. Nothing on screen asked the manager to think like an engineer, and that translation, more than any single feature, is what made adoption fast.
The virtual SDR agent
The core of the product was an agent that ran full qualification conversations toward a clear goal: book a meeting, confirm interest, or send materials. It operated around the clock, with no breaks or ramp-up. It classified where each lead sat in the funnel, cold, engaged, qualified, or not interested, and followed the cadence it was configured with: first touch, follow-up timing, number of attempts, channel priority. It handed the human rep only the opportunities that were genuinely qualified. The point wasn't to add another chatbot. It was to do the prospecting work itself and let salespeople focus on closing.
Build trust in layers
This was the keystone of the whole design, and the reason skeptical companies said yes. Trust in AI isn't granted on day one, so I designed supervision as a ladder rather than an on-off switch, with three levels.
At level one, total approval, every message the agent wrote went to a queue for a human to review before it sent. At level two, selective approval, only certain agents, high-value leads, or flagged messages needed sign-off, and the rest flowed automatically. At level three, full automation, the agent ran on its own and the manager simply monitored, stepping in only when they chose to.
Most clients started at level one. After a week or two of watching the quality hold, they moved to level two, and some reached level three within a month. The ladder let them build confidence at their own pace instead of making a single leap of faith. And throughout, managers could see why the agent made each decision, not just what it did. That visible reasoning is what turned skepticism into permission.
Make it say "I don't know"
An AI that invents answers under pressure is worse than no AI in sales. One confidently wrong claim about pricing, a roadmap, or contract terms can kill a deal outright. So I designed the knowledge base to ground every answer in the company's own materials, and to admit the gap when the answer wasn't there.
The knowledge base ingested documents, product pages, blog posts, FAQs, case studies, pitch decks, objection handlers, and structured them for retrieval during a live conversation, fast enough that responses felt instant. When a question fell inside what it had learned, the agent answered with context and consistency. When it fell outside, the agent said it didn't know instead of guessing. We trained the system on more than 1,500 high-performance sales emails collected from early clients, so it learned from what actually worked in the market, not from generic text. It sounded like someone who had studied the material, because it had.
Configuration, so the agent didn't sound generic
Generic AI sounds generic, and prospects disengage the moment they feel it. Most tools offered either rigid templates or nothing. We built a granular configuration layer in between, one that let a manager shape exactly how the agent behaved without touching anything technical.
On personality and tone, they set communication style, formal, casual, or technical, the level of formality from "Dear Mr. Silva" to "hey, just following up," the level of detail, and the industry language the agent used, so it spoke like SaaS in SaaS and like construction in construction. On objectives and limits, they defined what counted as a qualified lead ready to hand off, which topics the agent must not touch without approval, pricing, roadmap, contractual terms, and when to escalate to a human, on a complex objection, a custom request, or anything outside its scope. The result was an agent that operated with autonomy but inside clear guardrails: real independence for the manager, and conversations that felt human for the prospect.
Onboarding, borrowed from what they already knew
Once the product worked, adoption was the next problem. Rather than invent a setup flow from scratch, we modeled it on something every sales team already does: onboarding a new SDR. Teaching an agent its product, its objections, its tone, and its goals followed the same steps a manager would walk a new hire through, in the same order. The familiarity did the heavy lifting. People weren't learning an AI tool, they were onboarding a new rep, and they already knew how to do that.
Omnichannel, because leads don't think in channels
A lead sends an email, then replies on WhatsApp, then checks their phone later. Tools that treat each channel in isolation lose the thread, and the agent ends up asking a lead to repeat what they already said. We built the agent to operate across email and WhatsApp with a unified history, so when a lead switched channels the agent still knew the full context. Which meant higher response rates and stronger engagement simply because the conversation held together.
Managing expectations, honestly
People project too much onto AI. They assume it can do everything, and they fear it will replace them, and both distortions get in the way of adoption. So I kept the product honest about its own limits. It was explicit about what the agent could do, qualify leads, follow cadences, respond in context, what it couldn't, negotiate complex contracts, handle sensitive complaints, grasp deep nuance, and what was still coming. No "AI solves everything." That candor was part of why clients trusted the parts that did work.
How we worked
We were a small, remote, distributed team, so process wasn't a nicety, it was how the product got built at all. It started as just me and a developer, and as things grew I helped structure the company, hiring a PO and a front-end developer, and I managed through the developer turnover we went through, keeping continuity while the team churned. Everything ran documentation-first: complete specs before any code, so engineers knew what to build and QA knew what to test, with no "I thought you meant something else." Communication lived in Discord, tickets and docs in Jira and Confluence. We ran two-week sprints with asynchronous stand-ups, each person posting updates in writing rather than gathering for a daily call, and kept only one mandatory sync, the retro every two weeks. Feedback loops were tight: we met with beta clients weekly to review real conversations, which messages worked, which felt off, where the agent got stuck, and those insights went straight into the next sprint instead of waiting for a quarterly review
Results
LIDE Paraná, a major business association, put the agent on its member base and booked meetings across it in the first week, with no human SDR touching the keyboard. Across clients, the pattern repeated: reps stopped prospecting and moved to closing, receiving qualified opportunities instead of noise, and sales cycles shortened because the leads that reached a human were actually ready. Nobody fired their team. They kept their best people and pointed them at the work that closes, while the agent absorbed the repetitive qualification that used to burn out SDRs and walk out the door with them.
By the numbers, clients contracted around 100 agents, at roughly 60% the cost of a junior SDR, and the company, bootstrapped, reached break-even in under a year. For a first-of-its-kind product in a market that didn't exist yet, that was the real proof: companies paid for it and kept it.
What I learned
Trust in AI isn't granted, it's earned in layers. The instinct with a capable model is to show it off and let it run. What actually made companies adopt was the opposite: visible reasoning, a supervision ladder they controlled, and an agent honest enough to say "I don't know." Years later, that oversight is exactly what the industry is scrambling to add. The design job was to make people comfortable handing it something they cared about, one step at a time.
FACT SHEET
Role: Product Designer and Product Manager
Period: 2024 to 2025
Context: Brazil's first AI SDR platform. Bootstrapped, reached break-even in under a year.
Scope: Product design and product management from prototype to shipped product. Discovery, MVP definition, agent design, human-in-the-loop supervision, knowledge base, configuration, onboarding, omnichannel, remote process setup. Helped structure the team, hiring a PO and a front-end developer and managing through developer turnover.
Team: Started as me and one developer; grew as I helped hire a PO and a front-end developer. Commercial team grew separately, outside my scope.
Ways of working: Remote-first, documentation-first, asynchronous, two-week sprints.
Domains and keywords: applied AI, AI agents, agentic sales, human-in-the-loop, SDR automation, lead qualification, knowledge base, RAG-style grounding, governance and guardrails, omnichannel, zero to one.
Outcomes: Brazil's first AI SDR; 1,500+ sales emails as proprietary training data; ~100 agents contracted by clients; ~60% cheaper than a junior SDR; break-even in under a year.
