Every community team we meet wants the same thing: know which conversations matter, and route them to someone who can act. Most attempts fail at a step earlier than they expect, not at the model, but at knowing who is speaking.
The problem is identity, not intelligence
A Discord handle, a support ticket email, a Shopify customer record, and a CRM contact are, in most businesses, four unrelated rows. A language model reading the Discord message can classify intent perfectly and still produce nothing useful, because there is no account to attach the finding to.
That is why so many community-AI pilots stall after the demo. The demo scores intent on a message. Production requires the sentence 'this paying account is at risk', and that sentence needs an identity join the pilot never built.
What an intent graph actually is
An intent graph has three layers. Nodes are resolved people, each one a cluster of platform handles that we believe belong to a single human, carrying a confidence score rather than a silent guess. Edges are observed behaviours: asked about a plan limit, reported a bug, referred someone, went quiet after being active weekly for six months.
The third layer is the one teams skip: decay. An intent signal from yesterday and one from March are not the same fact. Without time weighting, the graph slowly turns into an undifferentiated pile of history and every alert starts looking equally urgent.
Confidence beats completeness
We do not try to resolve every identity. We label each link high, medium, or low confidence, and only high-confidence links are allowed to trigger an outbound action such as a CRM write or a sales notification. Medium-confidence links surface as suggestions a human confirms once, permanently.
This is a deliberately conservative posture. A community agent that is wrong about who someone is destroys trust faster than one that stays quiet, and the recovery cost is measured in relationships rather than tickets.
How we build it in an audit
In a fixed-price audit we inventory every surface where your audience already identifies itself, measure how much overlap is recoverable, and produce a deduplicated audience count with confidence bands. That number is usually the first honest headcount the business has ever had.
The build phase then wires the highest-value path only, typically one signal type to one destination, so value lands before scope grows.
The takeaway
Identity resolution is not preparation for the community agent. It is the community agent's substrate: no resolved graph, no defensible action.