Our First AI Employee: A WhatsApp Reader

Some of the best sales opportunities never touch a form or an ad. They surface in a group chat, as one line — “does anyone know someone who can build this?” — and they go to whoever answers first.

In early 2023, that was happening around us constantly. Across the WhatsApp groups we were part of, three to four people a week were asking for exactly the kind of work we do. The catch: those messages were buried in the everyday noise of the same groups — somewhere between 200 and 1,000 messages a day, across three to five active groups. Miss the message, miss the deal.

A person doing a background process’s job

So we did what most teams do: we watched. Someone kept the groups open and skimmed them through the day, scanning hundreds of messages so a buying signal wouldn’t slip past. It cost one to two hours every single day — not in one block, but bled out in constant checking and re-checking, plus the attention tax of being the thing that can’t afford to look away.

A human being was manually scanning an incoming stream on a loop, watching for one pattern. That is a job description for software — and, by spring 2023, for AI.

The first version of Metatron

The first version of Metatron was small and pointed, written in Go. It connected to the groups through a WhatsApp API, read every incoming message, and put each one in front of GPT-3.5 with a single question to answer: is this someone looking to buy what we sell — yes or no?

When the answer was yes, it pinged us on Discord — the tool we were collaborating on at the time — in under ten seconds from the moment the message was posted in the group. No dashboard, no app, no inbox to check. A message landed; a human picked up the lead while the conversation was still warm.

Teaching it what a sale looks like

We didn’t write rules or keyword lists — buying intent doesn’t live in keywords. We taught by example. The prompt was a lineup of real messages from the groups’ history, each labelled with the verdict we’d have given: genuine opportunities marked yes, and — just as important — deceptive negatives marked no: messages that look like a lead but aren’t.

Our favourite specimen: “can anyone help us with ads for a website?” It mentions a website. It sounds like our customer. It isn’t — it’s an ads request. Near-misses like that are exactly what a naive keyword filter gets wrong, and exactly what a language model gets right once you show it the difference.

Then we ran it and watched where it slipped. At the start it over-flagged — quite a few false positives, almost no false negatives. That asymmetry was a choice, not an accident: we tuned it to catch more rather than less, because a wasted glance costs seconds and a missed buyer costs a contract. Every time it misjudged a message, we added that exact case to the examples. It sharpened the way a new colleague does — corrected on the specific cases it got wrong — until it was nearly perfect, erring only on messages so ambiguous a human couldn’t have judged them with certainty either. Adjusting it never meant touching code. Plain language, one corrected example at a time.

The boring virtues: reliability and cost

Metatron v1 was so reliable we could forget it existed. It ran by itself, day after day. In its entire life, the only outages it ever had were the months we forgot to pay the OpenAI API bill.

And the economics were almost embarrassing: the whole system — reading up to a thousand messages a day, judging every one, pinging us in real time — cost under $10 a month in API fees. On the other side of the scale: one to two human hours a day, plus the certainty that nothing important was slipping past.

The result

• One to two hours a day, back — the constant manual scanning simply stopped.
• 5–6 real opportunities surfaced by the system — buying signals we’d have likely scrolled past, delivered to Discord while they were still answerable.
• Near-perfect judgement after tuning — deliberately biased to overcatch, wrong only where a human would have hesitated too.
• Under $10/month to run, with pings in under ten seconds.

The lesson we still build on

What this project settled for us, and what we’ve applied to every AI system since: AI models have strengths and weaknesses, and a system only behaves predictably when someone understands both. GPT-3.5 was brilliant at reading intent and useless if you assumed it was perfect. The difference between a flaky demo and a dependable employee was not the model — it was knowing exactly where it would fail, and designing the examples, the bias, and the correction loop around that.

That knowledge has to live somewhere: inside your company, or bought from people who have it. It is the single thing that separates AI that quietly does a job from AI that embarrasses you in production.

What it became

That small reader was employee number one of something much bigger. Metatron kept growing, and today it is essential to how we sell: our sales assistant, and the interface between the salesperson and the CRM that holds the company’s entire sales memory. It started by refusing to let a single lead drown in a thousand messages a day — and it never stopped earning its place.