Products
The recommendation is the product
Signal hands over one recommendation to approach one company now, with the evidence and its sources attached. That unit is what makes everything after it checkable.
Signal is the sensing product on Atlas OS. It searches the world outside the company for attributes that match a goal the company is pursuing, most often the attributes that identify a relevant customer, and what leaves it is a qualified opportunity: one company, the signal that makes a conversation worth having now, the events that signal was built from with the sources they were read in, the people who can be reached and a proposed next step. Signal does not hand over a list, and that single choice decides how the rest of the product is built.
Aimed at a research goal instead of a commercial one, the same mechanism returns condensed findings rather than accounts. The target changes. The shape of the output does not.
A list asserts, a card argues
A lead list ends in rows. A row states that a company is relevant, rarely says why now, and almost never says where that conclusion came from.
A person receiving that row has two options: accept it, or rebuild the reasoning. In practice the second happens for a while and then neither does. The failure here is not accuracy. A system can be right in every row and still fall out of use, because correct and checkable are different properties and only one of them survives contact with a working week.
So Signal treats inspectability as an engineering requirement rather than a reporting feature. That decision has consequences in what the system stores, in the order the pipeline runs and in what the system is allowed to say.
Every sentence on a card has an address
A card carries the signal, meaning the change at that organisation that makes a conversation worth having now. It carries the events the signal was built from, each with the source it was read in. It carries the company context, the people who can be reached, a proposed next step and a priority class.
There is no score on it. A score reads as a measurement, and what it would be measuring is a judgement the reader cannot recompute. A class says how the case ranks and stops there, which is the most a number on a recommendation can honestly do.
One rule governs the language on the card. Every factual statement about the organisation traces to a stored field or to an event with a source. A sentence that cannot be traced is flagged before a person sees it rather than after.
The rule has a price. Generated text gets plainer, and the flourishes that make a first message feel personally written are exactly the ones most likely to be invented. Atlas pays that price deliberately. One fabricated detail in one first message costs a team the confidence it had in the whole system, and confidence of that kind does not come back on a second attempt.
The card is built to be judged in under a minute: accept, reject, correct. That constraint, and not an aesthetic preference, is why it is one page rather than a dossier.
A rejected card names its own failure
Because the unit is a recommendation rather than a list, rejecting it carries information.
A rejected list says nothing: nobody can tell which of two thousand rows was wrong, or why. A rejected card identifies its own weak joint. The event type did not matter. The evidence was stale. The fit reasoning skipped something obvious. The contact held the wrong role. Those corrections are captured and become the material the qualification logic is measured against, so the learning loop is a property of the unit rather than a feature added to it later.
For the same reason the terminal record stays deliberately coarse. Signal records whether a qualified opportunity was handed over, whether it was accepted and what came of it, at exactly the depth needed to answer which signals produced revenue. It records no stages, no probabilities and no forecast, and it mirrors what it holds into the CRM the company already runs.
That is a decision against something specific. A full pipeline surface, with deals, stage management and a report builder, was considered and rejected. It is a twenty year old product category with hundreds of engineers behind it, and every hour spent approaching parity is an hour not spent on the part that does not exist elsewhere. Signal keeps the data model complete and the surface narrow.
A claim nobody can inspect is worth less than a claim nobody made.
Autonomy widens from a fixed floor
Signal researches, enriches, scores and watches on its own. Anything that reaches a person passes a gate, and a person opens it. That is the floor the product runs on, and the rule that decides where the gate sits belongs to the customer, per segment and per contact.
An autonomous system is defined by its exceptions, so some of those are fixed rather than configurable. A defined set of actions cannot be delegated by any rule a customer writes. Price commitments. Discounts. Contractual statements. Complaints. Data protection requests. Anything that mentions a lawyer. They route to a person however far the rest has been opened up, and that list belongs in the code rather than in a policy document, because the policy document is the artefact that gets suspended in the week a target is missed.
Above the floor, autonomy widens by class rather than by instance. Once a class of approach has been right often enough to be measured, approval can move from every instance to the edge of the class: exceptions escalate, and what sits inside runs. The threshold for that is set per intent, because the certainty at which a meeting request can run inside a class is not the certainty at which a pricing question can, and a single global threshold would have to be tuned to the most dangerous case, which would make it useless for the ordinary ones.
Permission to contact a person is held per purpose and per channel rather than as one flag, and a withdrawal is stored as a record rather than as a deletion. That is what makes the specific question answerable afterwards: was this contact permitted at the moment of sending, given what was known then.
None of this is a compliance chapter attached to the side of the product. It is the mechanism that lets autonomy widen at all. A system that can show what it did, on which evidence, under which rule, can be trusted with more. A system that cannot has to be supervised at the same depth forever, which is the ceiling most automation quietly reaches.
The objection worth taking seriously
The strongest argument against all of this is that it assumes a behaviour sales organisations do not reliably show. People want volume in the system and meetings in the calendar. A recommendation that asks to be read is overhead, and overhead loses to a longer list.
Part of that holds, and pretending otherwise would be dishonest. In an organisation measured on activity, a system producing fewer and better started conversations looks worse on the dashboard for a quarter before it looks better. What Signal claims is narrower than the objection assumes: when reading costs a minute and the card already contains the evidence a person would otherwise have gone to find, the review replaces work rather than adding to it. That is a testable claim and Atlas treats it as one.
What changes if this holds
The useful question about a sensing system stops being how many opportunities it produced. It becomes how many of its recommendations survive a minute of human attention, and how fast that share rises as corrections accumulate.
That number is available from the first week, and it is the number autonomy is built on top of. Every point it rises is a decision the system can be allowed to carry and a person who no longer has to carry it. Atlas OS understands the company and Signal senses what is outside it, and the loop between them starts with a recommendation somebody could check.
The product behind it
The product this piece is about, with the pipeline, the card and the authority model on one page.