The AI Learning Hub Journal

Where It Genuinely Helps Today

Documents in, structure out — and the test that decidesfour conditions have to hold together, and the fourth does most of the work HIGH-FIDELITY SOURCEthe text is in front of the model,not recalled from training VOLUME IS THE CONSTRAINThuman reading time is the actualbottleneck in the process REVIEW ALREADY EXISTSthe control is native to theworkflow, not bolted on THE REVIEWER CAN TELLan error is detectable against thesource sitting next to itall four hold— or one fails IF THE FOURTH FAILS the task has quietly moved intoa much higher risk category —however routine it feels EARNS ITS PLACE TODAYonboarding — entity extraction, ownership resolution, adverse-media summaryalert triage — case assembly, prior-alert summary, first-pass narrativeclaims and contracts — intake, covenant and clause location, consistency checksresearch — compressing filings, transcripts and reports read at speedclient-service drafting — a first pass a person reviews before it is sentwhat stays human: the escalation or filing decision and its recorded rationalemissed alerts are silent — measure what the assistant deprioritised, not only what it surfaced DOES NOT EARN ITS PLACEinvestment recommendations and client adviceengages the local advice regime — MiFID II suitability in the EU,Regulation Best Interest and the fiduciary standard in the USautonomous executionremoves the review step that made the others acceptable, and losesmoney faster than anyone can intervenea language model in place of a scored credit decisionforfeits the stability and explanation consumer-credit rulesin the US and EU assume THE COMMON FAILURE EACH TIME: NO GROUND TRUTH IN FRONT OF THE MODEL, HIGH CONSEQUENCE, A DUTY THE OUTPUT CANNOT MEET THE REVIEWER MUST BE ABLE TO SPOT A WRONG ANSWER AGAINST THE SOURCE anything sent to a client is also a retained, supervised business communication
Ask which obligation attaches to the output before asking whether the model is good enough at the task.

The Shape That Works: Documents In, Structure Out

The tasks where these systems currently earn their place share a recognisable shape. The source material is text and it is in front of the model rather than being recalled from training. The volume is high enough that human throughput is the binding constraint. A review step already exists in the workflow, so a human check is native rather than bolted on. And the reviewer can tell whether the output is right, because the ground truth is the document sitting next to it. That last condition does most of the work. Where a reviewer could not detect a wrong answer, the task has quietly moved into a much higher risk category regardless of how routine it feels.

  • Source text supplied by you, not recalled from training — extraction and summarisation, not knowledge
  • High volume where human reading time is the actual bottleneck in the process
  • A review step that already exists in the workflow, so the control is native rather than added
  • The reviewer can spot an error against the source — if not, the task is not in this category

Onboarding Review and Financial-Crime Alert Triage

Know-your-customer onboarding is document work at scale: extracting entity details from incorporation papers and identity documents, resolving ownership structures across filings, reconciling names across systems and summarising adverse-media results. Anti-money-laundering alert triage is similar — transaction monitoring generates far more alerts than analysts can investigate deeply, and most close without a filing. Assembling the case file, summarising prior alerts and drafting a first-pass narrative is real leverage. What stays human is the decision itself and its rationale. The judgement to escalate or to file — a suspicious activity report in the US and UK, a suspicious transaction report under EU regimes — and the reasoning recorded behind it belong to the firm and to a named person.

  • Onboarding: entity extraction, ownership resolution, name reconciliation, adverse-media summarisation
  • Alert triage: case assembly, prior-alert summary and a first-pass narrative for the analyst
  • The escalation or filing decision and its recorded rationale stay with a named human, always
  • Missed alerts are silent — measure what the assistant deprioritised, not only what it surfaced

Disputes, Contracts and Research Summarisation

The same shape recurs across the institution. Dispute and chargeback handling is document intake, extraction and consistency checking before a person adjudicates. Credit agreements and derivative contracts contain covenants, definitions and terms that teams currently locate by hand across hundreds of pages. Research and market summarisation compresses filings, transcripts and reports that an analyst would otherwise skim. Client-service drafting turns a query and the relevant policy text into a first-pass reply. In each case the model is compressing or locating text that you supplied, a person reviews before anything leaves the building, and errors are visible to that person. Note also that anything sent to a client is usually a business communication subject to retention and supervision.

  • Dispute and contract work: intake, extraction, covenant and clause location, consistency checking
  • Research: compressing filings, transcripts and reports an analyst would otherwise read at speed
  • Client-service drafting is a first pass — and the sent message is a retained, supervised communication
  • In all four, the human review that already existed is what makes the risk acceptable

Why Not the Other Things

It matters as much to say where this does not earn its place. Investment recommendations and client advice put a generative system inside a regulated advice perimeter: MiFID II suitability in the EU, and in the US Regulation Best Interest for retail broker-dealer recommendations plus the adviser fiduciary standard. Autonomous trading removes the review step that made the other uses acceptable and can lose money faster than anyone can intervene; faulty trading software has built enormous unintended positions within minutes. Replacing a scored credit decision with a language model surrenders the explainability and stability consumer-credit rules in the US and EU assume. The common failure each time: no ground truth in front of the model, high consequence, and a duty the output cannot meet.

  • Recommendations engage the local advice regime — MiFID II suitability in the EU, Regulation Best Interest and the adviser fiduciary standard in the US
  • Autonomous execution removes the human review step that justified the other use cases
  • A generative model in place of a scorecard forfeits the stability and explanation consumer-credit rules expect
  • Ask which obligation attaches to the output before asking whether the model is good enough at the task
◆ See it for yourself
Open the Base-Rate calculator in the library →

Put in your own volumes and see the false-alarm arithmetic.

Prefer slides, quizzes, and saved progress? Read this lesson in the library — free, no sign-up.