ConceptIllustrative

Applied AI and agentic systems

An agent is useful in finance only when its job, its limits and its uncertainty are explicit. This bounded pipeline shows where an agent helps, where it must stop, and where a person takes over.

Bounded analysis pipeline

Choose a sample case, then step through it. Each stage reports what it did, how confident it is and whether the case may continue.

Sample case

Quarterly report, complete data A synthetic company report with every field present and internally consistent. Price history with missing days A synthetic price series where 8% of days are missing, two in a row around a key date. Conflicting signals Synthetic news sentiment is strongly positive while the price and volume data point the other way.

01 Intake Not run Receives public or illustrative inputs and records their origin and time.
02 Validation Not run Checks schema, completeness and outliers before any analysis runs.
03 Analysis Not run Produces an estimate and a confidence score from validated inputs only.
04 Explanation Not run Writes the reasoning in plain language, including what would change the conclusion.
05 Human approval Not run A named reviewer approves, rejects or sends the case back. The agent cannot skip this step.
06 Monitored output Not run Approved output is published to a log with alerts for drift and later corrections.

Confidence Not yet estimated

Run log

No stages run yet. Select “Run next stage” to begin.

IllustrativeAll cases, scores and thresholds are illustrative and generated for this demonstration.

What the agent may and may not do

May

Read the inputs it was given and record where they came from.
Flag missing, inconsistent or out-of-range data.
Estimate and state its confidence and the reasons behind it.
Stop and escalate when confidence falls below the floor.

May not

Publish, trade or move money.
Fetch new data sources on its own initiative.
Overrule or skip the human reviewer.
Hide a warning to make an output look cleaner.

Designing for failure

Missing data: Gaps lower confidence and travel with the output as a warning. They are never filled silently.
Disagreeing sources: When inputs conflict past a threshold, the agent stops and asks for a person rather than choosing a winner.
Confidence floor: Below an agreed confidence, the system produces an escalation, not an answer.