Discovery deep-dive — Data Platform Agents

What it does One ingestion event in → an updated, defensible engagement record out.

Ingest anything

A Teams meeting transcript, a forwarded client email, or a filled-in questionnaire. The agent normalizes each into structured, redacted text.

Build the catalog

Every requirement, decision and open question is extracted with a verbatim citation back to the source, and written into a structured catalog & decision log.

Surface the hard bits

Contradictions across meetings, missing owners, and unanswered questions are flagged — then turned into a draft next-meeting agenda and follow-up emails for the consultant.

Why it matters

Goal

Reduce consultant hours in discovery (target ~40%) and make the offering cheaper.

Defensibility

Every requirement cites its source; every approved decision carries real approval evidence. Wins the “who approved that?” argument 18 months later.

Human-in-the-loop

The agent flags, the human decides. It never auto-resolves a contradiction, never sends an email, never invents an approver.

Client-facing

The engagement repo is itself a client deliverable — internal-grade structure, client-presentable surface (Teams + SharePoint).

The headline run (real, live gpt-5.4)

Two meetings ingested for a sample asset-management engagement (“Meridian”). Numbers below are from an actual end-to-end run.

22

cited items extracted across two meetings — 0 hallucinated citations

4

decisions tracked as ADRs; the decision-log index matched the ADRs 1:1

2

genuine contradictions surfaced as open questions — neither auto-resolved

Jump to the scenario walkthrough to step through it.

The pipeline One ingestion event runs through seven stages against a single engagement repo. Click a stage.

Inputs & outputs Three ways data comes in; one coherent engagement record comes out.

What the agent produces (the engagement repo)

Output	What it is	Written by

Trace one requirement, end to end

The same item as it moves from raw transcript → cited extraction → catalog → (later) refined.

Trace one decision, end to end

The agents Each one enforces its own guardrail in code, not just in the prompt. All merged to main.

Code-enforced guarantees The trust story — each is a structural invariant with unit tests, not a prompt instruction.

Generated from real end-to-end runs of the discovery agent on gpt-5.4 (Azure AI Foundry, uksouth).