6. Operations phase: Model deployment
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Sed do eiusmod tempor incididunt ut labore et dolore magna.
Why should I do this?
Once a model has been developed to meet the expected outcome and performs at a level determined to be ready for use in a real-world environment, deploy it into a production environment.
- 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 (Q15, Q16, Q17)
- Checklist 2: Participatory Co-Creation, Feedback, and Model Refinement (Post-deployment Q1, Post-deployment Q2)
- Checklist 3: Legal & Regulatory Compliance (Q13, Q14)
Questions to consider
Checklist 1 · Q15
What mechanisms are in place to notify users that the system is AI-powered and to explain how outputs are generated and interpreted?
Checklist 1 · Q16
What final security reviews are in place before go-live (runtime safeguards, output filtering, content moderation)?
Checklist 1 · Q17
What mechanism is provided for individuals to contest or appeal AI-driven decisions that affect them?
Checklist 2 · Post-deployment Q1
What mechanisms are in place to collect feedback on model behavior and system outputs to provide suggestions on the model improvement?
Checklist 2 · Post-deployment Q2
How does the feedback process support anonymous reporting and how are confidentiality protections communicated to users?
Checklist 3 · Q13
Is there an ongoing post-deployment process to monitor continued legal and regulatory compliance with the AI system, maintain appropriate records, and ensure readiness for internal or external audits?
Checklist 3 · Q14
Can compliance documentation (especially any records preserved under this checklist) be produced for an internal or external audit, within a defined period (e.g., 72 hours)?
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 AI assistant was deployed more broadly through the AgriConnect platform with ongoing extension officer oversight and monitoring.
Participatory co-creation, feedback, and model refinement
- What are the channels available for stakeholder feedback? Does this include an option for anonymous feedback?
- What documentation is maintained for model iterations?
Lessons learned and challenges that arose
The team found that trust in the platform could decline rapidly even with isolated inaccurate outputs. The team also found that there was a need to distinguish isolated errors from recurring or high-severity failures, and that model updates could introduce new risks even when overall performance appeared stable.
Decisions taken and documentation maintained
Incident severity and rollback/retraining thresholds were formally defined, with escalation procedures for high-severity or repeated failures. A structured feedback and incident-management process was introduced, including mechanisms for anonymous reporting where appropriate. Model versions, changes, evaluation results, incidents and release decisions were maintained in a human-readable model register, release notes, monitoring logs and incident records to support accountability, auditability and controlled updates.
AgroThrive: a policy and advocacy investment
The system was deployed across ministries and policy advisory units with ongoing governance oversight.
Participatory co-creation, feedback, and model refinement
- Are feedback and challenge mechanisms accessible to users?
- Are external stakeholders informed that AI-assisted synthesis was being used?
Lessons learned and challenges that arose
Post-deployment use showed that source selection and weighting was influencing the policy narratives generated by the system, particularly where some perspectives were more extensively documented. Older reports continued to influence outputs even when policy, economic or geopolitical conditions had changed. Users and external stakeholders needed clearer mechanisms to challenge or contextualize AI-assisted synthesis. There was a need for stronger governance oversight to extend beyond technical performance to include the diversity and provenance of the evidence informing outputs.
Decisions taken and documentation maintained
Source currency and weighting rules were introduced, with periodic review of the document corpus and mechanisms to flag outdated evidence. Governance reviews were expanded to consider source diversity, provenance and potential institutional or donor influence, with external advisory expertise included where appropriate. Transparency measures were introduced to identify when policy materials had been AI-assisted and provide mechanisms for feedback or challenge. Accountability, retention and audit requirements for AI-assisted policy outputs were formally documented. Model updates, source changes and weighting decisions, stakeholder feedback, governance reviews and deployment decisions were maintained in audit logs, model/source registers and oversight committee records.
NGBT: a field research investment
The AI educational assistant was deployed more broadly through agricultural programs and extension services.
Participatory co-creation, feedback, and model refinement
- Are communities informed about updates and limitations?
- Are stakeholder feedback loops active post-deployment?
Lessons learned and challenges that arose
Over time, the team recognized that large-scale educational messaging can shape farming norms and social expectations, even when technically framed as optional guidance.
Decisions taken and documentation maintained
Monitoring indicators were expanded beyond technical accuracy to include crop diversity impacts, community perceptions and behavioral trends. Governance reviews and monitoring metrics were documented in operational oversight reports and fairness audit records.