Email metadata. Who wrote to whom, when, and whether they replied. Note metadata, not content: the sender, recipient and timestamp carry nearly all the relationship signal, and the body of the message carries nearly none of it.
Calendar. Stronger than email, because a meeting costs both people time. A recurring thirty-minute slot with an external attendee is one of the most reliable signals of a real working relationship there is.
CRM activity. Who owns what, who has been contacted, which accounts went quiet. Thinner than the first two, and the most likely to be out of date.
Public professional history. Where people worked, when, and overlapping with whom. This is the source that reaches beyond your own company's interactions, and the only one that finds a relationship formed two jobs ago.
Not all contact is a relationship. A good scorer weighs:
The scores are not comparable across evidence types without care, which is where most naive implementations go wrong. A public comment on someone's post and a six-year working relationship are both "a connection", and treating them as equal puts the comment first.
The graph is not the product. The query is: given this target account, who can get us in?
The answer has to name three things to be actionable - the person at the target worth reaching, the colleague who knows them, and why that colleague is willing to ask. An output that stops at "you have a connection at Acme" is a dead end; the rep still has no idea what to do next.
The honest operational notes: initial enrichment of a company's network takes hours, not seconds. Scoring is cheap; collecting is not. And coverage is never total - people without a public professional footprint stay invisible however good the software is.
Email and calendar are sensitive. Reputable implementations read metadata rather than message bodies, run inside the customer's own tenant with their consent, and let individuals opt out. If a vendor cannot tell you exactly which fields it reads, that is the question to keep asking.