Deploying AI in a humanitarian organisation: policy, governance, training and final project

1Why a use policy

In many organisations, staff already use AI, often with personal accounts and without rules: this is known as shadow AI. A total ban is rarely effective: use continues, in secret. Ignoring it is dangerous: sensitive data goes to providers without a contract.

A short and clear AI use policy makes it possible to secure and encourage good uses. It fits with existing policies: data protection, IT security, communications, safeguarding, anti-fraud, human resources.

A good policy fits in two to four pages, is translated into the working languages and comes with training. A thirty-page policy that nobody reads protects nobody.

2The content of a use policy

• Purpose and scope: who it applies to (staff, consultants, volunteers, and expectations of partners).
• Authorised tools: list of approved tools and plans (for example a professional plan of Claude or Gemini), and the rule for others.
• Data classification: what is allowed with which tool (public, internal, confidential, sensitive), with concrete examples from the sector.
• Prohibited uses: decisions about people, protection and security data, misleading content, generated images presented as real.
• Responsibility and checking: the human author remains responsible; mandatory fact-checking.
• Transparency: when to disclose AI use (donors, publications, evaluations).
• Incidents: what to do in the event of a data leak or serious error, whom to inform.
• Training and support.
• Review: at least once a year, given the speed of change.

3Governance and tool selection

Light but real governance:

• a designated owner in management;
• a small group bringing together data protection, IT, programmes, MEAL, communications, and representatives of the field and national staff;
• a register of uses: which tool, for what, which data, which owner, which risks;
• a data protection impact assessment for any use involving personal data.

For tool selection, a procurement grid: data processing terms (location, retention, training, sub-processors), security (single sign-on, logs, administration), quality tested on real tasks and in the working languages, accessibility in the field (weak connection, mobile phone), total cost, discounts for non-profit organisations, availability in countries of operation, possibility of switching providers.

Procurement follows the organisation's usual procedures and, where applicable, those of donors.

4Training and support

The tool is not enough: value comes from skills. A realistic training plan:

• A mandatory core module (2 to 3 hours) for all users: how it works, limits, data, checking, policy.
• Workshops by job: programmes, MEAL, finance, logistics, HR, communications, with real cases.
• Champions in each office or department, who help their colleagues and report needs.
• A shared request library and examples.
• Inclusion: training in the working languages, adapted to weak connections, open to national staff and, where possible, to local partners.

Also plan to address concerns: fear of losing one's job, fear of doing it wrong. Management's message must be clear: AI should free up time for work with human value, and the organisation trains its staff rather than replacing them without discussion.

5Measuring results and costs

A deployment is steered with simple indicators:

• Adoption: share of people trained and active users.
• Time saved on tracked tasks (estimated by survey and spot measurements), net of checking time.
• Quality: errors detected, feedback from reviewers and donors.
• Incidents: data leaks, published errors, reported prohibited uses.
• Cost: subscriptions, training, management time.

Standard calculation. 40 professional licences at USD 25 per month cost 40 × 25 × 12 = USD 12,000 per year. If each user saves on average 2 net hours per week over 44 weeks, that represents 40 × 2 × 44 = 3,520 hours. At an average hourly cost of USD 15, the value of the time freed is about USD 52,800. This calculation is an estimate: it assumes that the time saved is actually reinvested, and it must be compared with real measurements after six months. Prices vary by provider, plan and discounts: check them.

6The final project: your adoption plan

To close the programme, prepare (for yourself or your organisation) a 4- to 6-page AI adoption plan:

• 1. Situation: current uses, needs, risks, maturity level.
• 2. Priority use cases: 3 to 5, chosen according to value and risk (start with high value and low risk).
• 3. Tools and plans selected, with the justification on data, quality and cost.
• 4. Use policy: the main rules.
• 5. Governance: owners, register, impact assessments.
• 6. Training and support.
• 7. Budget and indicators.
• 8. Timeline: 3-month pilot phase, evaluation, scale-up.
• 9. Risks and mitigation measures.

The diploma exam assesses these skills through questions and a case study. The plan itself is not graded by the Academy: it is a tool for your work.