Skip to content
Abelardo Carlos

PosterDCMI 2026 · Seoul

ImpactAI: generative AI for evidence-informed development decisions

Linxi Wang, Madeline Bassetti, Abelardo Lorenzo, Philipp Zimmer, Satvik Garg, Piriyakorn Piriyatamwong, Giuliano Martinelli and 14 more, Samuel Fraiberger · Development Impact (DIME), The World Bank

How a curated corpus of impact evaluations, a metadata layer and a five-stage workflow turn a practitioner’s question into a source-grounded answer, and what it took to deploy it inside the World Bank.

Start the story DCMI proceedings DOI The ImpactAI system
QuestionWhat is the evidence on cash transfers and women’s empowerment in South Asia?
intervention · cash transfersoutcome domain · women’s empowermentcoverage · South Asia

Conditional transfers … [1][3] Unconditional transfers … [2]

⚑ Evidence gap flagged where few studies match.

every claim → a source record with a persistent identifier

stages from question to cited answer
5stages from question to cited answer
design principles: grounding, traceability, workflow alignment
3design principles: grounding, traceability, workflow alignment
operational use cases, from project preparation to M&E
4operational use cases, from project preparation to M&E
approved for launch and opened to external users
2025approved for launch and opened to external users
01

Evidence exists; using it is the hard part

Across education, health, social protection, governance, jobs and climate, rigorous impact evaluations say what works, for whom and under what conditions. But at the moment of a decision, teams must find the relevant studies, compare approaches and judge whether a result transfers to their context. That work is slow, so evidence and operations drift apart.

The poster frames this as a metadata and knowledge-organization problem: evidence has to be documented, indexed and retrieved in a way that keeps both its original findings and its context.

Design principles

Evidence groundingTraceabilityWorkflow alignment

A decision-support tool, not a replacement for expert judgment.

Project preparation
Comparative effectiveness
Theory of change
Monitoring & evaluation
02

A metadata record for every study

The corpus is curated causal evidence, mainly randomized controlled trials. Each study becomes a structured record that combines Dublin Core-style bibliographic elements with domain attributes, down to the treatment effects.

One study, one structured record

illustrative record

Bibliographic · Dublin Core terms

dcterms:title
Cash transfers and women’s decision-making: evidence from a randomized trial
dcterms:creator
A. Author; B. Author
dcterms:date
2021
dcterms:publisher
Journal of Development Economics
dcterms:identifier
doi:10.xxxx/…

Domain-specific attributes

sector
Social protection
intervention type
Conditional cash transfer
outcome domains
Women’s empowerment · intra-household decision-making
geographic coverage
Bangladesh · South Asia
study design
Randomized controlled trial
target population
Women in poor rural households
treatment effects
d = 0.14 (SE 0.05) on decision-making index

The metadata layer does three jobs. Pick one to see which fields serve it.

The poster’s point: this shifts the evidential burden from the model’s parametric memory to curated, structured metadata.

03

Five stages, one question

The poster walks through a task team preparing a social protection operation. Follow the question through query understanding, retrieval, standardization, grounded generation and attribution.

Stage 1

Parse the question and infer the metadata facets it implies.

Task team asksWhat is the evidence on cash transfers and women’s empowerment in South Asia?

intervention

cash transfers

what was done

outcome domain

women’s empowerment

what was measured

coverage

South Asia

where

Question and stages from the poster (Sections 3–4). Studies, effect sizes and the answer text are illustrative; the random-effects average (DerSimonian–Laird) is computed live from them.

The result is end-to-end traceability, from each claim back to a paper and its bibliographic record, and a first defence against hallucination: when retrieval finds too little, the system says so instead of extrapolating.

04

How it is evaluated

1

Retrieval quality

  • precision
  • recall
  • coverage of evidence

2

Accuracy

  • citation counts
  • treatment–outcome coverage
  • effect sizes
  • task completion

3

Standardization

  • a specific standardization protocol for extracted effects
  • average effect from an econometric model
05

2025: from prototype to institution

  1. 1

    Prototype

    Evidence curation, system design and product development.

  2. 2

    User testing

    Testing with the people the tool is for, before a wider release.

  3. 3

    Governance & risk review

    Responsible-AI review for a high-stakes institutional setting.

  4. 4

    Approved for launch

    One of the first generative AI tools positioned for launch within the World Bank.

  5. 5

    External registration

    First version opened to external users: researchers, staff and partners.

Lesson 1

Not only technical

Deploying generative AI needs institutional alignment, responsible-AI review, user education and expectation management.

Lesson 2

Demand is real

Demand for AI evidence tools is substantial, but must be balanced against evidence quality, transparency and sustainability.

Lesson 3

Structure over fluency

The value is not fluent answers but evidence structured for real operational decisions.

Lesson 4

Metadata is the core

Performance, authenticity and user trust depend on high-quality, consistent evidence curation.

Next: wider evidence coverage, stronger source-grounding and quality assurance, better evaluation and user guidance, and support for questions on external validity, implementation constraints and evidence gaps.

The system behind the poster

The ImpactAI project page goes phase by phase through the pipeline: data collection, PDF parsing, extraction, taxonomy, entity linking, retrieval, grouping, meta-analysis and the answer, with the parts I built.

ImpactAI, phase by phase →

Cite

@inproceedings{wang2026impactai,
  title     = {ImpactAI: Embedding Generative AI into Evidence-Informed Development Decision-Making},
  author    = {Wang, Linxi and Bassetti, Madeline and Lorenzo, Abelardo and Zimmer, Philipp and
               Garg, Satvik and Piriyatamwong, Piriyakorn and Martinelli, Giuliano and Hussain, Saqib and
               Sajid, Nihaa and Gall, Cl{\'e}mence and Khan, Rida and Aggarwal, Aarushi and Mrema, Jackson and
               Sekwao, Grace and Prakash, Jay and Sanghavi, Dhruti and Bentil, Shadrack and
               Sharifkazemi, Mohammadmehdi and Pascale, Sophie and Polles, Flavia and Fraiberger, Samuel},
  booktitle = {Proceedings of the International Conference on Dublin Core and Metadata Applications (DCMI 2026), Posters},
  address   = {Seoul, South Korea},
  year      = {2026},
  doi       = {10.23106/dcmi.952675403}
}
All publications

Poster text: CC BY 4.0, DCMI 2026 proceedings.