What to automate, and where a person still has to look

Fully manual: an analyst performs every step, which is where complex cases and high-risk jurisdictions belong. Assisted: the system gathers, the analyst decides. Semi-automated: the system decides routine cases and routes exceptions. Fully automated: no human at all, for low-risk high-confidence matches.

Most mature programmes run semi-automated, and the metric that captures it is the straight-through processing rate — the share of applications that complete without a person touching them.

Data collection and aggregation, because it is retrieval across sources that do not coordinate. Entity resolution, because it is probabilistic matching at a scale no reviewer can hold. Watchlist and sanctions screening, because the lists change daily and the comparison is mechanical.

Document verification, because format and consistency checks are rules. And risk scoring, because a consistent rule applied identically to every file is exactly what a reviewer cannot promise across a shift.

Complex ownership structures, where the chain runs through jurisdictions that publish nothing and the judgement is about what the absence means. Adverse media assessment, which is a question of relevance and severity rather than of matching. Watchlist hit disposition — a name match is not a finding, and deciding whether it is the same person is the work.

Exceptions and edge cases, which are by definition the ones the rules did not anticipate. And relationship decisions, where the question is not whether the business is real but whether you want it.

It is tempting to raise the straight-through rate by moving the thresholds — approving more of the middle automatically. That does not add information; it just relabels uncertainty as confidence, and the errors arrive later as losses rather than as review queue.

The honest lever is more independent evidence, which moves applications out of the middle by making them genuinely clearer. That is a data problem, and it is the one worth solving.