Practical AI inna financial services — Data governance
Shorter reader edition a di paper weh di Caribbean AI Association Harbour Think Tank prepare under Adrian Dunkley leadership. It use di 11 November 2026 revision an keep all ten proposal an dem action limits.
Dis short edishan go wid di full orijinal ina Inglish.
Shorter reader edition a di paper weh di Caribbean AI Association Harbour Think Tank prepare under Adrian Dunkley leadership. It use di 11 November 2026 revision an keep all ten proposal an dem action limits.
Who make di decision
Data an reporting owner define permitted use, lineage an quality. Authorised stewards decide access, deletion, identity merger an disclosure.
Ten use case an dem controls
01. Source catalogue. Draft entry wid purpose, owner, fields, sensitivity an update frequency. Owner verify before publication. Level 1.
02. Conflicting definition. Compare what each report include, exclude an count over time. Owner approve a dated definition. Level 0.
03. Report lineage. Tie source, transformation an report field to evidence. Engineer confirm links an flag missing steps. Level 1.
04. Quality rule. Turn approved definition into value, date an reconciliation checks. Quality owner approve exception an tests. Level 0.
05. Retention. Match catalogue category to approved schedule an legal hold. Records an legal decide what fi keep or dispose of. Level 1.
06. Sensitive fields. Classify permitted metadata an samples in a restricted queue. Privacy check missed field an false label separately. Level 1.
07. Access request. Check purpose, role, data an duration, ask fi missing evidence. Authorised person decide permission. Level 1.
08. Data sharing. Gather purpose, fields, recipients, locations, retention an safeguards. Legal an privacy review before transfer. Level 0.
09. Possible duplicate. Show evidence fi an against each pair already flagged. Steward decide an keep original records. Level 0.
10. Collection fi an assistant. Inventory approved document, obsolete versions an restriction. Owner test permitted question an expected refusal before use. Level 1.
How fi build an check one test
At level 0 di assistant answer one request an a person carry out every step. At level 1 it gather approved input an prepare a draft. It send no message an change no system. Record source, prompt version, permissions, reviewer an stop condition.
Store source ID an version wid each proposed link. Test labelled record pairs, a missed sensitive field an a collection wid one obsolete an one restricted document. Count false merge, missed duplicate an unnecessary label separately. Owner must reproduce report an expected refusal.
Measurement an limits
Compare equivalent work wid di current process. Record sample size, preparation, review an correction time, material failure, permission an confidentiality problem. Count licence, integration, training, supervision an cost per run where relevant. Scores help select pilots. Di paper report no measured return at any institution. Di complete source give starter prompt, example tool, skills an method fi each case.
Suos dem
- [1] Caribbean AI Association. Think Tanks. Harbour programme remit. Accessed 26 October 2026.
- [2] Office of the Information Commissioner, Jamaica. Data Protection Standards. Accessed 26 October 2026.
- [3] Basel Committee on Banking Supervision. 2013. Principles for effective risk data aggregation and risk reporting. January; scope, paragraphs 13–19 (pages 3–4), and Principles 1–3 (pages 6–9).
- [4] Basel Committee on Banking Supervision. 2026. Implementation of the Principles for effective risk data aggregation and risk reporting. Newsletter 36, 6 January. Informational publication.
- [5] Agent Skills. Agent Skills Overview. Accessed 10 November 2026.