AI Labs
All examples

Transformation Discovery with Claude Structured Output

Extract initiative risks, stakeholders, and systems from unstructured text using schema-constrained generation.

Structured extractionEnterprise

Target outcomes

  • Workshop-to-charter cycle cut from days to hours with HITL
  • Field-level accuracy tracked before ERP/PM sync

Initiative playbook

Typical delivery arc for this pattern in enterprise programs.

  1. 1
    Discovery2 to 4 wks

    Define Zod schemas with PMO for initiative charters; collect 50 anonymized workshop transcripts.

  2. 2
    Pilot6 to 8 wks

    Human review all extractions before backlog import; track field-level accuracy.

  3. 3
    Scaleongoing

    Embed extraction in discovery workshops with HITL approval and ERP/PM tool sync.

Business use case

Problem

Transformation discovery produces workshop notes, interviews, and legacy documents that PMOs manually turn into charters, slowing steering committees and duplicating effort.

Who benefits

  • Program management, structured backlog inputs
  • Consulting / transformation leads, repeatable discovery accelerators
  • Enterprise architecture, systems-in-scope captured consistently

Success metrics

  • Workshop-to-charter cycle from days to hours with human-in-the-loop
  • Field-level accuracy tracked before syncing to Jira/ServiceNow
  • Mandatory human approval before executive readout

Solution

Paste discovery text; Claude returns JSON matching a Zod schema (initiative name, risks, systems, timeline). Powered by AI SDK generateObject for type-safe server responses.

Technical implementation

Stack

  • @ai-sdk/anthropic + ai generateObject
  • Zod schema in lib/demos/extraction-schema.ts
  • Model default: claude-sonnet-4-20250514 (override via ANTHROPIC_MODEL)

Architecture

Workshop notes go in; a schema-validated charter comes out, ready for human edit before it hits the backlog.

How it runs
Drawing the flow…

Implementation highlights

  • Schema encodes risks with severity enum and mitigation text
  • Server validates output against Zod before returning to client
  • UI renders pretty JSON for reviewer edit before export

Outcomes and learnings

  • Schema-first extraction beats prose summaries for prioritization forums
  • Claude handles long workshop transcripts in one pass
  • Always keep human approval before committing to program charters

Where else this applies

Schema-constrained extraction turns messy narrative inputs into systems your organisation already uses, PMO tools, risk registers, architecture repositories.

Initiative chartering

Workshop output becomes structured scope, risks, and timelines for steering forums.

Vendor RFP comparison

Extract requirements fit, gaps, and pricing assumptions into a scoring matrix.

Regulatory gap assessments

Turn interview notes into control gaps and remediation owners.

Architecture discovery

Capture systems, interfaces, and data flows from brown-bag transcripts consistently.

Using this stack elsewhere

Claude via AI SDK `generateObject` excels at long-context workshop notes; the same approach works on Azure or Bedrock with different models if procurement requires it.

Live demo

The demo is the same code path described above, not a simplified mock UI. Add keys in .env.local when you are ready; the narrative and diagrams stand on their own without them.

Business

Dump workshop notes, get a charter-shaped JSON back. Still needs a human before it becomes a Jira epic, we are not pretending otherwise.

Technical

Claude generateObject against a Zod schema; validated on the server before the UI renders it.

Structured discovery extraction

Claude returns a Zod-validated transformation assessment for PMO review.

Live