Technical Reporting:�AI-generated zero draft content�Scaffolding, not substitute

27.2.26

Jules Colomer, Leader Portfolio Performance and Results

1

Purpose & Contents

• Introduce the AI-generated zero draft Technical Report content
• Take a tour of the sausage factory
• Describe the delivery pathway & support structure
• Talk about human ownership guardrails
• Contribute to the broader conversation about safe and effective use of AI in workflow
• Answer your questions

​

​

​

​

2

An evolving Performance & Results Management context: More juice, less squeeze

Questions are: More interconnected (e.g. $ and results). X-Integrated partnership (Centers/Independent/System), Domain/Entity/Geog specific. Requirements are: More (dis-)aggregable. Quicker. Simpler. More transparent. Higher Quality. Repeatable. Tunable. Made for AI.

3

AI tool use in Performance and Results Management

Goals:

Efficiency at scale without sacrificing rigor:

AI reduces duplication, accelerates evidence aggregation, and supports traceability. Tools integrate directly into CGIAR workflows, dashboards, and reporting processes.

Credibility through human-in-the-loop validation:

Unlike many general-purpose AIs, PPT’s AI outputs remain transparent, verifiable, and linked to underlying sources. Human-expert in the loop quality assurance and safeguards maintain accuracy.

Status:

We have moved beyond isolated AI pilots and are now demonstrating an integrated, human-in-the-loop architecture for accelerating evidence generation, validation, and synthesis across CGIAR.

Objective of providing AI-generated zero draft

The zero draft removes the heavy lifting of evidence compilation so teams can focus on interpretation, narrative, and decision-relevant insight.

• Remove the blank-page problem�Teams start from a populated, structured document rather than an empty template. This reduces a time-consuming phase of reporting.
• Cut evidence-hunting and compilation time�Key data are pre-assembled from PRMS for rapid x-checking. Less time is spent searching, copying, and reconciling information.
• Shift effort toward interpretation and strategy�Less time compiling → more time explaining results, framing progress, and shaping the narrative.
• Reduce cost of reporting over time�Lower manual effort and fewer late-stage corrections translate into sustained efficiency gains.

​

5

AI Tools in use

• Sensemaking (Clustering and synthesis)
• SNAP https://snap.cgiar.org/
• Used as first analytical screen for FCAS review
• IWMI preparation for EPMR
• Automated zero draft Technical Report generation
• 2025 P/A Technical Reports, 2025 Portfolio Narrative
• 2024 Portfolio Narrative
• Quality Assurance helper
• 2024 & 2025 Technical Reporting Quality Assurance
• Structured data extraction from unstructured data
• Impact Compendium
• Chatbots
• Performance & Results Chatbot Helper
• Project to Program mapping decision support
• Similarity scoring provides Centers and P/As with decision support

Read the report: https://cgspace.cgiar.org/items/5ca2c041-3ea3-4f5b-84b4-4431230acfb1

6

PRMS 2022-24

Human expert: Conflict transformation and peacebuilding specialist

Product

What it is/isn’t

Version 0 (v0) — AI-Generated Zero Draft

evidence pre-assembly.

The v0 is a machine-generated synthesis created from evidence in PRMS. It is intended to provide a comprehensive, evidence-linked starting point, not an approved narrative.

Authorship

Generated using available data.

No human author.

No institutional ownership.

Purpose

Reduce blank-page effort.

Surface relevant evidence, trends, and signals.

Provide a structured scaffold aligned with the reporting template.

Nature of content

Evidence-compiled, not judgement-based.

May include gaps, inconsistencies, or weak signals.

May contain interpretation that requires verification.

Accountability

• Not citable.
• Not publishable.
• Not an official CGIAR product.

​

​

​

​

​

​

v1+ = human-authored report

sense-making and narrative construction.

The moment a human edits the v0, the document becomes a human-authored report. From that point onward, responsibility shifts fully to the Programme/Accelerator team.

Authorship

Owned by Programme/Accelerator leadership and contributors.

Human judgement becomes primary.

Purpose

Interpret evidence.

Apply context, nuance, and strategic judgement.

Make claims, narratives, and positioning decisions.

Nature of content

Selective, intentional, and curated.

Includes interpretation, framing, and emphasis.

Reflects managerial and scientific judgement.

Accountability

• Citable.
• Publishable.
• Institutionally owned.

​

​

v0 assembles evidence; human drafts create meaning and assume responsibility.

Scaffolding NOT Substitute

8

Data/Information/Knowledge/Wisdom

Wisdom: "integrated knowledge—information made super-useful“ & the ability to make sound judgments and decisions apparently without thought

Knowledge: Knowledge is a fluid mix of framed experience, values, contextual information, expert insight and grounded intuition that provides an environment and framework for evaluating and incorporating new experiences and information.

FRAMEWORK = TOC

Information: organized or structured data, which has been processed in such a way that the information now has relevance for a specific purpose or context, and is therefore meaningful, valuable, useful and relevant.

Data: discrete, objective facts or observations, which are unorganized and unprocessed and therefore have no meaning or value because of lack of context and interpretation

FRAMEWORK

9

Pipeline Steps — Overview

1

Data Consolidation

PRMS and ToC data combined into a single consolidated data source

2

Section Slicing

Python script slices data — creates 1 slice per Annual Report Section

3

Mini-Slicing & LLM Narrative

Per section: Python creates mini-slices (e.g. partners by country); LLM produces targeted narrative

4

Standard Visuals

Python also creates standard charts and visuals for each section

5

Assembly & Editing

All sections assembled into full editable report with prompts to re-generate or discard

6

Word Export

Export as .docx with watermarks — watermarks must be removed before official use

⚠ Important: The LLM is focused on putting the narrative together, NOT on calculating numbers. All numerical computations are handled by the Python pipeline.

PRMS + ToC → Python Slicing → LLM Narrative → Editable Report → Word Export​

​

Annual Report Generation Pipeline

INPUT DATA

CONSOLIDATION

SLICING

SECTIONS

PER-SECTION PROCESSING

PRMS Data

ToC Data

Consolidated

Data Source

Python Script

Slices data into sections

1 per Annual Report Section

Section 1

Section 2

...

Section N

Per-Section Processing

Python creates mini-slices

e.g. partners by country,

KPIs by region, etc.

LLM reviews relevant slices

Produces targeted narrative

Python creates standard visuals

Narr. 1

Narr. 2

Narr. N

⚠ Important: The LLM focuses on putting the narrative together, NOT on calculating numbers.

HOLISTIC REVIEW

ASSEMBLY

EXPORT

LLM Holistic Review

Reviews each section holistically

and produces cohesive narrative

Full Editable Report

All sections assembled

Access to prompts to

re-generate or discard sections

Export: Word (.docx)

Includes watermarks that

need to be removed

before use

PRMS + ToC → Python Slicing → LLM Narrative → Editable Report → Word Export

What gets generated/what sections it covers

An AI-generated draft narrative and a set of draft graphs will be produced in Word for the following sections:

• 2025 Overview (factsheet + executive summary)
• Annual Progress (P/A level progress + detailed Area of Work progress)
• Partnerships (and internal portfolio linkages for Accelerators)

The Word document will be watermarked as Draft – Not for Publication, and is intended for consideration by P/A teams for inclusion in the Annual Technical Report.

The Word doc. will be preceded by an Excel export of the underlying data used to generate the draft narrative and graphs.

​

​

12

Next steps

• Join the testing group: If you would like to test and use the AI tool that generates the annual technical report zero drafts, please contact Nicoleta (n.trifa@cgiar.org ) by Monday, 2 March,12:00 CET. Interested P/A focal points will be added to a small testing group to review the tool outputs and provide rapid feedback.
• Kick-off session: Nicoleta will schedule a demo of the tool, explain the workflow, and define validation process to generate drafts (sensible & effective QA).
• We aim to share the AI-produced zero drafts with P/A teams Monday, 9 March.

​

13

P/A Technical Report: key dates and deliverables

14

Guardrails – workflow and principle

• Drawing from Quality Assured data
• Code checks
• Human validation at different levels
• Iteration cycles
• – within and between e.g. PN
• Remove before flight tags
• Alignment with IAES principles
• Use of Disclaimers

CGIAR uses AI-assisted tools within its reporting workflows to support evidence synthesis and draft preparation. All published content is reviewed, validated, and approved by responsible human authors prior to release.

15

16

Q&A

Is this replacing us?

• No. v0 removes evidence compilation effort.�Interpretation, framing, and judgement remain entirely with the P/A teams. v1 ownership begins the moment you edit the draft.

Will this misrepresent our work?

• v0 is scaffolding, not a product. Nothing is publishable without your validation. All interpretive claims must be reviewed and adopted by you before institutional status.

Is this going to increase our workload?

• The goal is to reduce blank-page effort and evidence hunting.�We expect the first 30–40% of reporting effort to disappear.

Who is accountable if something is wrong?

• AI outputs (v0) have no institutional standing. Institutional accountability begins at v1, once human validation and ownership are assumed.

Are we required to use it?

• No. Use is optional.

What about bias or hallucinations?

• That’s precisely why we formalized the v0 → v1 boundary.�AI-generated content cannot enter institutional reporting without closure of compliance, correctness, methodological and interpretive checks.

What if our data in PRMS is incomplete or wrong?

• The tool will surface inconsistencies. It does not invent performance; it synthesizes what is recorded. This actually improves data quality feedback loops.

Will funders see AI-generated text?

• Funders will see human-authored institutional products.�AI supports drafting but does not replace ownership or accountability.

Do we disclose use of AI?

• Yes we do. Products include a standard note where AI-supported tools were used in the evidence lifecycle.

17

Resources

AI tools for evidence exploration and use

From Food Security to Food for Security (2 pager)

A Review of CGIAR’s Engagement in Fragility and Conflict: Summary Report (10 pager)

Bibliometric Analysis

Portfolio Explorer

Results Dashboard

SNAP

CGIAR Impact Report

2022 CGIAR Technical Reports �2023 CGIAR Technical Reports�2024 CGIAR Technical Reports

CGIAR Technical Reporting Arrangement

CGIAR Performance and Results Management Framework (PRMF)

Innovation Portfolio Management

Reports on Responsible Innovations & Scaling

Plan of Results and Budgets

Scaling Readiness website�Online course on Innovation and Scaling�Readiness Calculator

Performance & Results Hub�Evaluation and Management Response Actions Tracker

​

Email: performanceandresults@cgiar.org

​

​

18