Digital Division Analytix
Pages
JCORP AI · How it works

How JCORP AI works

From a question to a cited, audited answer. Every request to JCORP AI passes through the same five stages — classify, retrieve, generate, ground, audit — whether it arrives from GenChat, from a platform dashboard, or from an agent. The diagram below animates one request through the pipeline; the numbered notes under it say what each step does and why it is there.

The five stages · every request, in this order
  1. 1ClassificationWhat is being asked, and what kind of answer it needs.
  2. 2RetrieveFetch the enterprise data and source documents that bear on it.
  3. 3GenerateDraft the answer with a frontier model chosen for the task.
  4. 4GroundTie every claim back to a source, and cite it.
  5. 5AuditRecord who asked, when, and what was returned.
One request through the pipeline
Left to right. The coloured tags say which of the five stages a block performs; the numbers match the notes below.
1User querya question, a task, a file
Classification
2AI Ingestprompt conversion
3Routing
  • Reasoning
  • Deep research
  • Image generation
  • Video generation
  • PPT slides
  • General chat
Frontier LLMs (routed)
Data collection (×3) GroundAudit
4Enterprise Datathe lake, the platforms, the registers
5Groundingwith citations
6Audittime stamp · user
Retrieve
Generate
Source documentswhat was read
Cited outputthe answer, with its sources
Audit logwho · when · what
7LLM / platform connectionthe same pipeline serves GenChat, the dashboards and the agents
Step by step
  1. 1
    User query. A person — or a platform acting for one — asks something: a question in GenChat, a request for a report, a file to be read. Nothing has been interpreted yet; this is the raw ask.
  2. 2
    AI Ingest — prompt conversion. Classification The ask is turned into a structured prompt: what kind of task it is, what it should return, and which constraints apply. Classifying first is what lets the rest of the pipeline treat "summarise this policy" and "draw me a chart" differently instead of sending everything to one model.
  3. 3
    Routing to a frontier model. The classified request is routed to the model and toolchain suited to it: reasoning, deep research, image or video generation, slide-building, or general chat. The models are frontier LLMs behind a common gateway, so a route can be re-pointed without changing anything a user sees. The GenChat log records which route served each message in its agent field.
  4. 4
    Enterprise data. Retrieve Before anything is written, the pipeline fetches what JCorp already knows: source documents, the platforms' data, the registers. Retrieval is deliberately its own stage — a model that answers from memory alone is guessing about JCorp.
  5. 5
    Grounding, with citations. Generate Ground The answer is generated against the retrieved material and every claim is tied back to its source. The output carries citations, so a reader can check the sentence against the document rather than trust the model. An answer that cannot be grounded is flagged, not polished.
  6. 6
    Audit. Audit Who asked, when, which route answered, and what was returned are written to the audit log. That is what makes an AI answer reviewable after the fact — and it is the same telemetry the Usage Analytics pages read.
  7. 7
    The LLM / platform connection. The pipeline is shared. GenChat, the business-analytics dashboards, the Digital Registry and the agents all call the same five stages over the same connection, so a fix to grounding or a new model route lands everywhere at once.
Data collection is marked ×3 because the same collect → ground → audit sequence runs for each of the three things a request can draw on: documents, platform data and prior conversation.
This page describes the design, as drawn by Digital in September 2026. How ready each piece is — the grounding pipeline, the guardrails, the agents — is scored in the readiness model, not here; this page carries no figure.
Newton · JCorp's answer engine

This Is How Newton Works

A recorded Newton chat, start to finish — the pipeline above as a user sees it: a question in, a grounded, cited answer out.
Newton is JCORP AI's answer engine — the assistant that runs the five stages above for a person at the keyboard. The recording is a screen capture without narration; its length and size are shown beside the controls. Use the controls to play, pause, scrub or replay it.