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