Every morning Signál puts at least five companies on your desk that are worth phoning today. This is the story of a design that didn't pan out, and a list of features I refused to put into it, because I know what data protection allows, what European AI regulation says and what simply isn't done to people.
Every morning Signál puts at least five companies on your desk that are worth phoning today. This is the story of a design that didn't pan out, and a list of features I refused to put into it, because I know what data protection allows, what European AI regulation says and what simply isn't done to people.
The client's brief went roughly like this: I need to simplify sales somehow. That was all. No idea what it should do, what it should look like, where exactly the problem is.
Don't take that as a reproach, this is what most of the briefs that come to me look like. A company knows something is pressing on it. Naming it and proposing what to do about it is exactly the work I'm paid for.
The whole system I then came up with myself. What it should watch, where to get the data from, what a salesperson's morning should look like, how companies are ranked, what will not be done in it and why.
I even built a basic demo so you could see how you would actually work in it. In the end the deal fell through and the whole thing stayed lying in a drawer.
When I opened it again after a few weeks, an uncomfortable realisation hit me: the design is better than the job it was created for. And above all it hit me that this is exactly what we are missing too.
So I threw out everything to do with the client, the name, contacts, numbers, the whole business case, and the rest we at Transformuj.ai finished into a product called Signál.
It's a sales radar. Every morning it puts at least five companies on your desk that are worth phoning today. For each it writes the reason why this one, and the contact of the person who decides on it.
But what's interesting about it is not what it can do. It's the list of things I refused to put into the design, even though technically they could be done and some of them a customer would happily order.
It's not that you are missing a salesperson
A company of fifty people has a salesperson. Usually one, sometimes two. What it doesn't have is everything that stands behind one in a large company. An analyst who follows the market. A person who prepares lists. A tool that tells you in the morning where something moved. This machinery doesn't exist in a smaller company, so the salesperson builds it themselves. Between meetings, in the evening, or not at all.
So Signál doesn't replace the salesperson. It replaces what is missing behind them. It won't call for you, arrange a meeting or close a deal. It makes sure your person knows in the morning whom to call and why, instead of spending three hours finding it out themselves.
But we have a CRM
We have our own CRM system, which we use for our sales and build for clients. Signál is its offshoot. Not because we wanted one more product, but because everyday work showed where the hole is.
A CRM is memory and skeleton. It holds who you talked to, when, what you promised each other and where the deal stands. For what you already know, it is irreplaceable and Signál doesn't replace it.
Signál is eyes and brain. It looks outward at companies that aren't in your CRM because you haven't heard of them yet, and decides which are worth hearing about today.
A skeleton without eyes won't take you anywhere. Eyes without a skeleton have nowhere to store what they saw.
The difference is where the data flows in from. Into a CRM someone has to type it. If nobody does, the CRM stays silent, and silent convincingly, because an empty field looks exactly the same as a field nobody needed. Signál takes data from outside, so it can't get stuck on the salesperson not having had time to fill something in.
What I deliberately left out of Signál
When I design a system that works with data about companies and people, every feature I consider goes through three sieves before the first line of code is written.
The first is legal. What the terms of service allow for the sources I would take data from, and where things stand with database rights.
The second is regulatory. What data protection and European AI regulation say about the feature.
The third is ethical, and that is the most important. Whether the thing is fine towards the people it affects, even if it passes both previous sieves. Because legal and right are not synonyms.
I have to build these sieves myself. The client doesn't deal with them and doesn't have to, it's my craft, not theirs.
And let it be clear right away: not once was I stopped by not knowing how. For most of them I knew exactly how. That's precisely why I know where the problem is. Whoever doesn't know how usually doesn't know either what they would break by it.
Tracking people on LinkedIn
I had this in the design first and I was thrilled about it. Watching who changed position, who started as a new director, who writes about expansion.
Then I went through the conditions under which it could actually be done, and threw it out. Technically it can be solved and tools for it exist, which is also why everybody advertises this feature. But LinkedIn has no public interface for this that could be used honestly, and automated harvesting of profiles is against the terms of service. Publicly visible doesn't mean freely usable.
So I take changes in management from the commercial register, where they are official and dated.
Connecting to job advertising portals
Hiring is the strongest thing you can read from data. When a company is looking for twelve people for operations within a month, something is going on there and it's going on now.
I checked the terms and dropped it. It can be built in one afternoon and works reliably until someone notices. These portals have no public interface, and automated harvesting of them without a contract goes against the terms of use and against database rights. So Signál builds on the open data of the Ministry of Labour.
Scoring people
In the first version of the design people were scored too. When I turned the design into a product, I threw it out and today Signál gives a score exclusively to companies. It isn't a technical limitation, it's my decision.
Systems that profile specific natural persons fall in European regulation into the strictest category, with everything that comes with it.
I decided not to fall into this category at all. And a grade on a person is useless to a salesperson anyway. They need to know whether the company has a reason to deal with something right now, and who to talk to about it.
Reading emotions
If the system estimated from an email or a call transcript what mood the other side is in, you could adjust the tone and timing of the next message accordingly.
It could be built and it would work. Still, it isn't there and won't be. Recognising emotions in the workplace and in recruitment is outright banned in Europe. But even if the ban fell tomorrow, I won't build it. There is a difference between reading public data about a company and guessing the inner state of a specific person so that you can push them harder.
Automatic sending
I could build this feature in an afternoon and I know it would sell best of all. So Signál can prepare an email, but it can't send it on its own and never will.
Here I'm stubborn even when the customer wants the opposite. A system that sends out messages by itself in the company's name will one day write something it shouldn't have, and the owner will hear about it from an angry recipient. The tool prepares the draft, a person clicks to send.
Promises of instant data
The last thing I forbade myself not in the code, but in what I say about the product. I don't claim that Signál watches the market in real time. Registers are updated in batches, typically within days, and no clever add-on will change that.
What may be added one day
Most of those things can be built. Just not like this, not quietly and not without asking.
A contract can be signed with a job portal and its data then used entirely legally. It isn't unavailable, it just isn't free and automatic. A person can be tracked when that person gives consent.
The difference lies in a single thing: whether I take it without asking, or we agree.
And even consent is only a first step, not a ticket. Consent doesn't erase the obligations that follow from regulation.
With reading emotions, though, it isn't about whether someone gives consent. It's about what you do to the other person with it. No signature will convince me there.
Why every product should have a list of what it won't do
Tools in this category die from having everything anyone can think of added to them, until a slow colossus emerges that does twenty things by halves.
That's why for every product we build I also write a list of features that won't be in it. Not as an alibi, but as part of the brief. When you know what you're not building, you can build the rest properly. And every feature you don't build is also a feature you don't have to maintain, fix and explain for years.
Where the data flows from
Everything Signál builds on is public and official. No bought bundle of contacts of unknown origin.
You can click through every signal to its source. That is the whole trick for trust. The salesperson doesn't have to believe the score, they can check the reason.
But the list of registers is the easier half. Anyone can download the data. A tool is what the layer above them turns it into.
What runs in it even though nobody talks about it
A signal has a half-life. Hiring that a company launched four months ago isn't an opportunity, it's history. Each type of event has its own window in which it means something, and outside it its weight falls.
One signal on its own means nothing. Meaning arises from combinations and above all from order. A company that is hiring is one thing. A company that is hiring a month after winning a public contract and replacing a person in management is a different situation and you talk to it differently.
Thresholds are calculated relatively. Five open positions at a company of thirty people is an earthquake. At a company of three hundred it's a Tuesday.
Segmentation doesn't rest on the sector. The activity code from the register is an administrative classification, not a description of how the company behaves. I don't define the ideal customer from a desk, it's derived backwards from the deals you have already closed.
Exclusion rules are as important as the selection ones. Half of the work of a good radar is not offering you a company that isn't worth touching.
And a score without feedback is just a number. When two of five companies from today get in touch and three don't, that's information the system has to take and use to recalculate tomorrow morning.
On top of all this doesn't run one model, but several. Different steps have different demands, so different models are used for them. Extracting structure from a register entry is cheap mechanical work. Judging whether three unrelated events together make an opportunity is, on the other hand, a task for the best thing available at the moment.
The limit of what such a tool can do is not fixed. Things I had to reject a year ago as unfeasible are routine today. What doesn't shift by itself is the legal and ethical line. What can't be done technically today may be possible tomorrow. What I rejected because it touches people in a way I don't want will be rejected tomorrow too.
And then there is the question of where those models actually run. With sales data it's a question you should ask before the question of price. It isn't just a list of companies. It's also whom you approach, why him in particular and what you write internally about it, in other words a fairly accurate picture of your sales strategy.
That's why we have a partnership with Kosmik Compute, a Czech company that runs models on its own GPUs in Prague.
What the hours it eats cost
This is a model calculation, not a measured result for a specific client.
When a salesperson watches the market themselves, it's usually about eight hours a week. At a rate of 20 EUR an hour and forty-eight working weeks that is 7 680 EUR a year.
But that amount is the smaller part of the problem. Worse is that those eight hours are time when your salesperson isn't calling or sitting with a customer. A machine can do the research for them. It can't do the meeting.
When not to buy Signál
It makes no sense if you sell to end consumers. In public registers you learn nothing about your customers, because they simply aren't there.
It makes no sense if your market is so small that you know it by heart. If there are thirty companies in the Czech Republic that could buy from you, you don't need a radar. You need a phone and patience.
And above all it makes no sense if nobody is there to approach those companies. Five companies a day is twenty-five a week. If nobody touches them, Signál is just another list nobody looks at. In other words, a dead folder, and every company already has enough of those.
What to take from this even if you'll never have Signál
When someone offers you a system that watches the market and looks for customers, ask them two things.
First, where the data comes from and whether you can click through to its origin. If the answer is public sources and you learn no more, that's a bad answer.
Second, what the system deliberately doesn't do. Whoever has no answer to this question hasn't thought about it, or doesn't want to say. Either way costs you the same.
Who watches the market at your place? The managing director between meetings, or nobody?
The article was first published on LinkedIn. Original article on LinkedIn
