[{"data":1,"prerenderedAt":4},["ShallowReactive",2],{"article-body-beyond-enterprise-search":3},"\nEnterprise search is having its second youth. Assistants promise\nanswers across every system, vendors publish self-reported hours-saved\nstatistics as their headline benefit, and the category is sprinting\ntoward agents. The market has largely accepted time saved as the\nmetric that matters.\n\nTime saved is a fine metric for finding documents. It is the wrong\nmetric for the job the enterprise actually has; producing an answer a\nperson can act on and later defend to a colleague, a board, an\nauditor, or a regulator. That job has requirements a search box, and\nmost of what is currently sold on top of one, does not meet.\n\n## The four tests of an actionable answer\n\nWhether an answer can be acted on and defended comes down to four\nproperties, each checkable in a procurement meeting:\n\n- **Cited, verifiably.** Every claim in the answer traces to sources\n  the asker can open. Not \"based on your documents\" as a mood, but\n  this figure from this filing, this clause from this contract. An\n  uncited answer is a rumour with good grammar.\n- **Computed inside the asker's permissions.** The retrieval runs\n  under the viewer's entitlements at the data layer, not a service\n  account's broad access with filtering bolted on afterwards. If you\n  cannot see it at the source, you cannot see it, or feel its shape\n  through a count, in the answer. Almost no vendor publishes evidence\n  on this property, and it is the one that decides whether the\n  assistant is a tool or an exfiltration risk.\n- **Resident where the data lives.** For regulated estates, the\n  models, the index and every query stay inside the organisation's\n  infrastructure. No external APIs, no third-party processing of the\n  crown jewels as a side effect of asking a question.\n- **Stable on refresh.** The same question, from the same asker,\n  against the same corpus returns the same answer. The substance is\n  deterministic and the rendered answer is kept, so what anyone was\n  told can be produced again and cited in a decision record. Different\n  askers may rightly receive different answers; that is permissions\n  working, not instability.\n\n> The market publishes hours saved. Nobody publishes leak tests,\n> citation coverage, or answer stability. The absence is the\n> opportunity, and the risk.\n\n## From answers to knowledge to action\n\nMeeting the four tests gets you trustworthy answers. The remaining\ndistance to \"beyond search\" is structural, and it is where a knowledge\ngraph earns its place. Documents mention entities. Suppliers,\nobligations, sites, people, controls. A graph resolves those mentions\ninto distinct entities and the relationships between them, so the\nsystem can answer questions no document contains. Every obligation\ntouching this supplier, every site affected by this regulation change,\nevery contract that breaks if this term shifts.\n\nOnce answers are entity-shaped and permission-scoped, acting on them\nbecomes safe. Drafting the case, routing the exception to its owner,\nwith the citation chain attached. This is where the market's\nrush to agents runs ahead of its foundations; an agent is only as\ndefensible as the answer it acted on. Verifiable answers first,\ngraph-shaped knowledge second, action third, each inheriting the\nguarantees of the layer beneath.\n\nThis sequence is Prophesee's Nexus engine. Cited answers over a\nknowledge graph, permissions enforced at the data layer, fully\non-premises, from 100,000 to 50 million documents. Ask it something\nyour search box cannot answer. [Start here](/contact).\n",1786984935415]