Define the task
Classify a specific use of AI: a defined task, carried out by a particular system, in a particular setting. The product name alone tells you too little. The same assistant might draft a briefing, rank requests for assistance or send messages. Record these as separate uses because their consequences differ.
Identify capabilities
The categories below describe AI capabilities, or the kinds of task a system can perform. They use shortened labels from [A, sections 1.1 and 2.1]. A use can combine analysis with generating content. Agentic behaviour means combining capabilities and pursuing actions across a task. It is recorded separately from the type of result produced.
| Code | Capability | Result to identify |
|---|---|---|
| C-PRED | Prediction | An estimate or forecast. |
| C-CLASS | Classification | A label assigned to a record. |
| C-GROUP | Clustering | Groups based on similarities. |
| C-ANOM | Anomaly and change detection | A flag for an unusual pattern or change. |
| C-EXTRACT | Perception and extraction | Information extracted from images, audio or documents. |
| C-RANK | Ranking and decision support | Items placed in order, or a suggested choice. |
| C-GEN | Content generation or transformation | New or revised text, speech, images or translations. |
| C-SYNTH | Synthetic data generation | Artificial records for a defined purpose. |
Assign a code to each significant step. A system that reads messages, judges their urgency and drafts replies needs several capability codes. If you lack evidence about a result, mark it as unknown. You can also record automation that follows fixed rules, using a separate non-AI tag.
Then describe the system’s permissions: what it may read, draft, change or communicate. If it works across several steps, record its tools, permitted actions, stopping point and when it must refer work to a person. Naming a reviewer does not demonstrate that the person can control the work effectively.
Identify the humanitarian function
The proposed codes below group humanitarian work by its purpose, across sectors. They need testing. They are neither an established humanitarian standard nor a complete list of possible uses.
| Code and function | Include | What to distinguish or check |
|---|---|---|
| H-ASSESS / Understanding needs | Analysing assessments and identifying information gaps. | Distinguish a population estimate from a finding about a household. |
| H-ANTICIPATE / Preparing for change | Interpreting hazards and planning for possible demand. | A forecast does not authorise a response. |
| H-PLAN / Planning a response | Comparing programme options and preparing resource plans. | Record whose priorities and constraints define the comparison. |
| H-ACCESS / Access to assistance | Registration support, referrals and assistance-related case review. | Explain how an AI result could lead to refusal, delay or reduced support. |
| H-DELIVER / Delivering services | Supporting frontline services and following up assistance. | Distinguish administrative help from professional judgement or a protection decision. |
| H-SUPPLY / Moving goods and services | Supply planning, routing and operational scheduling. | Record consequences for people when a proposed route or schedule fails. |
| H-COMMUNICATE / Communicating with people | Translation, information services and accessible communication. | Identify who verifies advice before people rely on it. |
| H-LISTEN / Feedback and remedy | Sorting feedback and supporting complaints handling. | Common themes must not hide an urgent individual concern. |
| H-LEARN / Monitoring and learning | Reviewing implementation, evidence and programme results. | Separate a generated interpretation from verified evidence. |
| H-SUPPORT / Organisational support | Internal drafting, coordination and record preparation. | Internal use can still expose sensitive information or affect people. |
Choose the main function that matches the purpose of the use, then add other functions where needed. Record the sector, crisis setting, language, geography and phase of the operation separately. A coordination task that spans several sectors should retain those distinctions.
If none of the categories fits, write a provisional description and flag it for review. Add a category only when the existing definitions cannot cover the work. If you later change a classification, retain the earlier code and explain the change.
Keep a reviewable record
Create one record for each defined use. Keep supporting evidence in the organisation’s existing governance documents and link to it from this record. The proposed AidGPT record organises the classification and references. It does not repeat an assessment or replace the SAFE AI Transparency Card.
| Field | What to record |
|---|---|
| Identity and purpose | Record identifier, title, responsible role, version, date, intended outcome and alternative without AI. |
| Function and setting | Primary H-code, secondary functions, sector, location, languages and operational conditions. |
| Task and capabilities | Information supplied, significant processing steps, result, C-codes and anything not yet known. |
| System and authority | System version, how it pursues actions across a task, connected tools and permissions to read, draft or act. |
| People and consequences | Who uses the system, who is affected indirectly and what happens if a result is wrong. |
| Data and evidence | Where sources are kept, restrictions, validation evidence and gaps. Link to sensitive records without copying them. |
| Participation | Who influenced the design, what changed, who was missing and where disagreement is recorded. |
| Assessment references | Separate links, versions and status for each applicable CDAC and DCI assessment. Mark incomplete assessments as unassessed. |
| Human decision | Authorised role, decision, reasons, conditions, scope and unresolved matters. Do not assume approval. |
| Operation and review | Who monitors the system, how incidents are reported, the fallback, who may suspend use and what requires reassessment. |
Start with a real task and describe its steps. Ask another reviewer to classify it independently, then discuss any differences. Clarify definitions before using the taxonomy to compare cases. Date and explain changes so that earlier records can still be understood.
Mark a claim [V] when it has been checked against a source, [R] when it is an attributed report, [I] when it is an inference and [X] when sources disagree and the conflict is unresolved. Date figures and retain their sources. An incomplete assessment means evidence is missing. It does not mean risk is low.
Sources
A. A Taxonomy for AI in Social Protection: Defining Capabilities and Application Areas