Responsible AI & Algorithmic Policy

Why This Policy Exists

A.G.E.N.T uses artificial intelligence to guide community reporters through a structured workflow, generate summaries of child protection cases, suggest next steps for officers, and assist with service directory lookups. These are not trivial functions — they directly influence decisions that affect the safety of real children in Ghana.

An AI that misunderstands a Ghanaian community's dialect, misclassifies a CRITICAL case as low-severity, or generates a summary that omits a crucial detail is not just a technical failure. It is a safeguarding failure.

The governing principle of this policy is this: the AI is the assistant. The officer is the guardian. That hierarchy is absolute.

The Human-in-the-Loop Requirement

Absolute Prohibition on Autonomous AI Decisions:

The following decisions must never be made by AI alone, under any circumstances:

• Whether a child is at risk.
• Whether a case should be escalated, closed, or reassigned.
• Whether a report constitutes abuse, neglect, or exploitation.
• Whether emergency services should be contacted.

Required Human Review Before Action:

Before an officer takes any action on a case, they must read the full structured referral (not just the AI summary), confirm or override the suggested severity, and apply professional judgement. An officer who acts solely on an AI summary without reviewing the full case has not met their professional obligation.

Emergency Override:

If an officer or community reporter signals an emergency — a child in immediate danger — the system stops asking structured questions and immediately provides contacts for the Ghana Police Service (191) and DOVVSU.

Transparency & Labelling

Every AI-generated element displayed in the A.G.E.N.T dashboard must be clearly labelled. They must appear directly adjacent to the output (no hiding in footers/tooltips). Labels include:

• AI-generated summary — review full case before acting
• AI suggestion — officer decision required
• AI-suggested severity: [LEVEL] — confirm or override

Data Used by AI — Prohibited Inputs

The AI may process structured referral fields, anonymised case notes, and service directory entries.

The AI must NEVER process:

• Raw message content from WhatsApp or SMS.
• Audio recordings or transcriptions in their raw form.
• A child's full name combined with their location.
• Reporter identity in any form.

No Training on Case Data: Case data held in A.G.E.N.T must never be used to train, fine-tune, or improve any AI model — whether the Foundation's own model or a Third-Party AI Provider's model. This prohibition applies permanently.

Third-Party AI Provider Obligations

The Third-Party AI Provider used for case summarisation must confirm in writing in the Data Processing Agreement (DPA) that:

• Case data is processed only to generate requested output and not retained.
• Case data is not used to train or improve any AI model.
• No human reviewers read A.G.E.N.T case data at the provider's facilities.
• The provider notifies the DPO within 24 hours of any security incident.

Governance & Calibration

AI models trained on international datasets may not accurately interpret Ghanaian cultural context or dialects. We conduct monthly reviews of the AI Error Log (overridden flags) and quarterly calibration reviews to ensure accuracy for Ghanaian communities.

AI Governance Lead: Rebecca Tiwaah Boafo
Deputy AI Lead: Opoku Nana Kwame Ababio