1. Design phase: Problem definition
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Sed do eiusmod tempor incididunt ut labore et dolore magna.
Why should I do this?
Establishes strategic, operational, and ethical foundations for any AI initiative. Define the problem, intended use cases and users, and ensure the solution is justified, feasible, aligned with organizational priorities and stakeholder governance-ready.
- Workflow activities Miro Board
Target users
- CTOs and technical leaders
- Data scientists and ML engineers
- Product managers
- Research organizations, think tanks
Checklists in play
- Checklist 1: Risk Identification & Mitigation (Q1, Q2)
- Checklist 2: Participatory Co-Creation, Feedback, and Model Refinement (Design Q1, Design Q2, Design Q5)
- Checklist 3: Legal & Regulatory Compliance (Q1, Q2, Q3, Q4, Q5, Q6)
Questions to consider
Checklist 1 · Q1
What roles and decision-making authorities have been established to support effective governance and oversight throughout the AI lifecycle?
Checklist 1 · Q2
How is the intended use case of the AI system documented and communicated to ensure it is understandable and accessible to non-technical stakeholders and affected communities?
Checklist 2 · Design Q1
What evidence will demonstrate that relevant stakeholders were identified and engaged in shaping the AI system's intended purpose, use cases, and objectives?
Checklist 2 · Design Q2
How will stakeholder representation be reviewed for potential bias to reflect the diversity of those affected by the AI system and its outputs?
Checklist 2 · Design Q5
What assessment will be conducted to identify persons/communities at heightened risk
Checklist 3 · Q1
Has the organization conducted, or solicited legal advice on, the jurisdiction-specific applicable laws and regulations (specifically data protection laws, cybersecurity and digital infrastructure laws, any AI regulation around risk mitigation, safety mandates, standards, reporting, and transparency, and miscellaneous laws like intellectual property, consumer protection, and product liability)?
Checklist 3 · Q2
Has the AI system undergone a risk assessment as prescribed under any applicable laws (including legislation, regulations, guidelines, advisories, or industry manuals) on product safety (or specifically around AI safety and risk mitigation), and determined the degree of risk posed by such AI system in development?
Checklist 3 · Q3
Is there any applicable law, regulation, or binding mandate that prohibits the deployment or use of the proposed AI system in the Deployment Jurisdiction(s)?
Checklist 3 · Q4
Has sector-specific compliance (for example, agri-tech or digital agriculture laws, regulations, policies, guidelines, or advisories) been reviewed?
Checklist 3 · Q5
Is the AI system an autonomous decision-making algorithm, with its outputs directly making or indirectly influencing critical decisions on access to products, services, or governmental welfare schemes (like subsidies)?
Checklist 3 · Q6
If Question 5 is answered yes, are there any laws (including legislation, regulations, guidelines, advisories, or industry manuals) imposing legal restrictions or obligations on the design, development, and deployment of such autonomous decision-making AI systems? Specifically, are there any legal requirements for a risk impact assessment of such systems?
Investment types
Every investment project is unique
The application of the six steps will vary accordingly. To provide examples that align with your project, common characteristics of AgDev investments were researched and three ‘investment types’ were developed.
AgriConnect: a digital solutions investment
The AgriConnect team with the AI development agency explored adding a GenAI assistant to the existing agricultural data platform to allow farmers and extension officers to query datasets using natural language.
Workshops were conducted with extension officers, researchers, cooperatives and farmer groups.
Participatory co-creation, feedback, and model refinement
- Are women farmers, low-literacy users and remote communities adequately represented?
- Are stakeholders involved in defining the AI assistant’s intended use and limitations?
- What is the process followed in determining which stakeholder groups were included and excluded?
Lessons learned and challenges that arose
Early workshops revealed farmers were likely to interpret outputs as authoritative advice rather than informational support. The team also identified a risk that the AI could unintentionally conflict with government approved agricultural guidance.
Decisions taken and documentation maintained
The team narrowed the AI assistant’s scope to decision support rather than fully automated recommendations. High-risk categories (e.g. pesticide use, subsidy advice) were flagged for mandatory human review. These decisions impacting the problem statement were documented in a product scope register, the risk classification log was updated and a stakeholder matrix was created, which was then reviewed by project leadership.
AgroThrive: a policy and advocacy investment
AgroThrive with the university team explored adding a GenAI layer that could synthesize agricultural policy reports, donor-funded project evaluations, climate adaptation strategies and investment case materials into concise briefing notes for ministries and investment agencies.
Participatory co-creation, feedback, and model refinement
- Are policymakers, researchers, civil society groups, producer associations and local agricultural actors consulted on intended use?
- Could some stakeholder perspectives become unintentionally deprioritized through synthesis?
- What is the process to be followed in determining what inputs were included and excluded for synthesis and what impact would that have on the model's outputs?
Lessons learned and challenges that arose
Early workshops revealed that users were less interested in “summaries” and more interested in “decision-ready narratives.” This created a risk that AI outputs would shape policy framing itself rather than simply improve information access. The team identified a high risk of automation bias that could dilute sovereign policy functions.
Decisions taken and documentation maintained
The team reframed the system as an “evidence exploration tool” rather than a policy recommendation engine. A governance decision was made that all outputs must preserve traceability to source material and clearly surface areas of disagreement or uncertainty. This was documented in the system purpose statement and governance documentation.
NGBT: a field research investment
NGBT with the government innovation team explored creating a GenAI-enabled educational assistant to support awareness and adoption of the new barley variant through mobile and extension-service channels.
Participatory co-creation, feedback, and model refinement
- Are smallholder farmers, women farmers and extension officers involved in defining the system’s purpose?
- Could the tool nudge behaviors (dark patterns) or unintentionally pressure farmers toward adoption?
- Are concerns about the impact on traditional farming practices being surfaced?
- Which communities will be at heightened risk if this new crop is widely adopted?
Lessons learned and challenges that arose
Early consultations revealed that farmers did not separate education from recommendation. Many interpreted the existence of the tool itself as evidence that adoption was already expected or officially endorsed.
Decisions taken and documentation maintained
Using active design cues, the team reframed the tool to be a crop information and comparison tool to highlight the final decision making lies with the farmer. Governance guidance required the system to present trade-offs, risks and alternative options rather than exclusively promoting the new variant. These principles were documented in communication guidelines and governance charters. A RACI matrix was co-developed by the different parties involved to help bring about alignment and clarity.