Skip to content
WHO/AFRO HIS Review

Search is only available in production builds. Try building and previewing the site to test it out locally.

12 · AI methodology annex

Owner: Steven Uggowitzer (personal-voice chapter) · Editors: Stephanie Berrada (release-approval) Effort: ~1.5 person-days · Due: v0.5 EOD Fri 2026-09-04 · v0.8 Fri 2026-09-11

What WHO expects here

Genuine transparency about a subject WHO is still forming its own institutional posture on. Not a marketing chapter. Not a hedge chapter either. A substantive account of where AI accelerates the work, where the team still does the work by hand, and what evidence exists that the outputs are reliable.

This chapter carries reputational upside for IQVIA if it lands well and reputational downside if it reads as either evasive or naive.

Ultimate contents

What AI does in this engagement

  • Documentation ingest. Structured extraction from long PDF documents (SCORE, HMN, prior evaluations, AFRO strategy documents) into machine-readable summaries. Every extracted claim traceable to source page.
  • Terminology consistency. afro-terminology (bilingual EN/FR digital-health glossary, publicly published) enforces canonical terminology across deliverables. Coverage report generated per deliverable.
  • KII transcription assist. transcriptapi.com or equivalent for first-pass transcription; human reviewer corrects before coding.
  • Coding acceleration. Two-analyst human coding remains the ground truth; AI-assisted first-pass proposals accelerate the second-analyst review.
  • Discovery-stream tooling. Shadow-IT DNS + CT-log sweep is fully automated (SSLMate CertSpotter); results reviewed by hand before entering the inventory.
  • Diagram + document composition. Mermaid / Typst / Structurizr rendering, driven by human-authored source.

What AI does NOT do in this engagement

  • Score platforms. Scoring is human. Full stop.
  • Draft findings or recommendations without evidence-check. Draft output that goes anywhere near the deliverables is reviewed against source evidence by a named consultant.
  • Author citations. Every citation is looked up by a human before it lands in a deliverable.
  • Speak to stakeholders. No AI-driven outreach. All stakeholder correspondence is human-authored or human-reviewed at correspondence-signature quality.
  • Access private WHO data unattended. Any access to non-public WHO material goes through a named consultant with credentials, not through automated pipelines.

How human review is applied

Every AI-touched artefact carries a marker in its version-control history. The reviewing consultant signs off in the commit message or in a review log entry. Reviewer identity is recoverable from git blame or the review log.

Audit trail

  • Every AI-authored commit uses a distinguished git author identity (afro-his-da-iqvia in Gitea) so the audit-trail is unambiguous.
  • Every deliverable’s terminology coverage report ships with the deliverable and pins the afro-terminology release tag used.
  • Every KII transcript retains the raw first-pass and the human-corrected version so re-review is possible.
  • The AI-usage discipline is auditable at file-level in Gitea, not just claimed in prose.

What we still handle by hand

Long stretches of the analytical work. The scoring rubric authoring. The methodology chapter (Chapter 4). The recommendation authoring. The stakeholder correspondence. The Brazzaville trip logistics. The delicate KII moments where a human hears something between the lines of what was said.

Ethics + guardrails

No AI-processing of KII audio without participant consent. No PII/PHI ingestion into third-party AI APIs. Terminology tooling is public-corpus only (afro-terminology corpus is not participant data). AI-generated draft output does not enter the deliverable pipeline until a named consultant has evidence-checked it.

Evidence to draw on

  • afro-terminology corpus + coverage tooling
  • .claude/memory/reference_gitea_assistant.md for the AI-agent Gitea identity
  • Steven’s personal working practice (documented in the Phase 1 master plan, Annex B)
  • IQVIA firm-level AI-usage policy (Stephanie to confirm)

Complementary effort triggered

Steven-authored, Steven-voice. Substantive, not decorative. Stephanie release-approval review before it goes near WHO. This chapter carries genuine reputational upside.