Data and AI: Readiness, Quality and Governance

Most AI failures are data failures

When an AI project stalls in a public institution, the cause is usually upstream of the model. The records exist but are duplicated across systems, incomplete in the fields that matter, undocumented so nobody can say how they were collected, or held under terms that do not permit the intended use. No model repairs any of these conditions; it inherits them and expresses them at scale.

This is why data readiness deserves assessment before a use case is approved rather than after a pilot disappoints. The use case discovery and readiness playbook scores exactly this dimension alongside value and risk.

What AI-ready data means

Four properties, in order of how often they are missing. Findable: someone can say where the authoritative copy lives, rather than which of nine spreadsheets is closest. Explained: fields have definitions and provenance, so a value can be interpreted the same way twice. Trusted: quality is measured rather than assumed, with known rates of duplication, missingness, and staleness. Lawful for the purpose: the basis on which the data was collected permits the use being proposed, which is a separate question from whether the data is technically available.

A useful test: if a competent analyst who joined last month cannot answer where a figure came from and what it excludes, the data is not ready regardless of its volume.

Record linkage and deduplication

Large registration systems accumulate multiple records for the same person: names transliterated differently, dates entered inconsistently, a household split across registrations, a person re-registering after losing documentation. Exact matching finds almost none of these, because the fields that would match are precisely the ones that vary.

Probabilistic record linkage handles the problem by scoring how much agreement on each field raises or lowers the likelihood that two records describe the same person, then setting thresholds for automatic merge, automatic reject, and human review. The middle band matters most: in any setting where a wrong merge affects a person’s entitlements, the ambiguous cases belong with a caseworker rather than with a threshold.

The payoff is operational rather than statistical. Data integrity analysis on UNHCR registration data recovered 63,000 missing contacts, restoring the ability to reach people who had effectively become unreachable in the system. Contact data is the precondition for every other channel: an appointment service has nobody to call without it. That work is described in the AI for refugee operations hub.

Knowing what you hold

Governance is unenforceable over an estate nobody has inventoried. A digital discovery across UNESCWA inventoried 181 databases and 145 portals and assigned each a lifecycle verdict, which converted a fragmented estate into a documented basis for consolidation and secured director-level endorsement to proceed to remediation. That inventory is what allows a data governance rule to be applied rather than merely published.

Data responsibility where the subjects are vulnerable

In humanitarian contexts the people described in the data may be at risk from the authorities they fled. Minimisation, retention limits, and access control are therefore protection measures rather than compliance paperwork, and the question of where inference happens is a safeguarding question as much as an architectural one. The operating tension is real: data minimisation reduces the value available to any model, and the resolution is to decide what the system must accomplish before deciding what it may collect, not the other way round.

Related: AI for humanitarian organisations, AI security and assurance, and public sector AI governance. Where those systems then act rather than only report, the operational examples are on agentic AI for humanitarian operations. A current example of that constraint in the field is Peru’s AI career coach, which works in metropolitan Lima and stops at the connectivity line.

Related reading

These go deeper on parts of this subject:

Data minimisation and AI value — why collecting less usually performs better.

Synthetic data explained — a privacy-preserving route to usable training data.

Poverty mapping — the technical method and where it is applied.

Common questions