The reason most subscription tools cannot answer a genuinely useful question is structural. "Which subscriptions on this card renew in the next thirty days, have gone unused for ninety, cost more than twenty a month, and have a cheaper alternative?" is five reports in a normal schema, and nobody runs five reports. It is one traversal in a graph.
So the graph is not an add-on here. As records are written, five entity types are projected into the platform's graph store — providers, relationships, plans, invoices, and AI agents — along with four named edges that carry the meaning: a relationship is held with a provider, a relationship is on a plan, a plan is offered by a provider, and an invoice is billed for a relationship. The projection is registered from a single configuration at service start and driven by the write path itself, so the graph does not drift away from the tables it mirrors.
Each projected node keeps the identifiers that make isolation enforceable — the workspace it belongs to and, where relevant, the organisation — and carries a human-readable title so a traversal result reads as "Netflix Premium" rather than a UUID. The relationship record is deliberately the hub: the fundamental object in Subscriber Bot is not a payment or an invoice but the recurring relationship itself, and everything else hangs off it.
You can already see the shape of what this unlocks without any model involved. The Insights screen normalises every active subscription to a monthly-equivalent amount so plans on daily, weekly, quarterly, and annual cadences are comparable, buckets dated renewals into a six-month forward forecast, ranks your largest recurring costs, and surfaces a watchlist of low-scoring relationships. Every figure is computed live from your portfolio — there are no illustrative numbers on that page.
The reasoning layer on top is honestly not built. Duplicate and overlap detection over the graph, the semantic matching that would tell you two AI tools do the same job, a scoring model that would compute a relationship's health rather than storing a number someone typed, and a price benchmark corpus that would say whether you are overpaying — all four are designed and ticketed, none are shipped. The health and vendor scores you can see in the app are pass-through fields today: the gauge is real, the model behind it is not. What exists is the substrate and the honest aggregation on top of it, which is the part that is expensive to retrofit.
The interactive traversal UI that would let you walk the graph yourself — asking cross-cutting questions by clicking edges instead of reading aggregations — is not shipped yet. The live app shows the monthly-normalised Insights that the graph makes possible, but this story remains a narrative about the data model and what it enables, not a clickable in-app journey today.
Ready to make this your story?
