BISM1201 Chap.10 Business Intelligence, AI and Responsible Decisions
Business Intelligence, AI and Responsible Decisions
Business intelligence transforms governed data into analysis and decision evidence. Source records are selected, validated and integrated; measures are defined; summaries reveal status or pattern; and a responsible decision-maker interprets them in context. A dashboard is only the visible end of this chain.
AI can classify, predict, recommend or generate from patterns, while process mining can discover actual flows from event logs. These tools do not remove accountability. Responsible design defines the supported decision, data boundaries, human review, error consequences, monitoring, access and a route to challenge outcomes. Correlation and prediction should not be overstated as causal proof.
What this chapter covers
- 01
Data-to-information-to-decision pipeline
- 02
Metric definitions, units, periods and owners
- 03
Dashboard comparison and interpretation
- 04
Process mining and predictive capabilities
- 05
Privacy, fairness, review, monitoring and accountability
Governing a service-delay prediction
- +1Define the decision as prioritising human review, not automatic denial or closure.
- +1Validate required fields and document the model’s data boundary.
- +1Compare false-positive and false-negative consequences for different case types.
- +1Require a supervisor to review evidence and record any override.
- +1Monitor performance, drift and relevant group differences after deployment.
- +1Provide a challenge route and keep accountability with the organisation.
Key terms
- Business intelligence
- The governed processes and tools that transform operational data into analysis and evidence for organisational decisions.
- Decision metric
- A defined measure with a formula, unit, period, owner and intended action or review threshold.
- Process mining
- The discovery and analysis of actual process paths from recorded events with case, activity and timing data.
- Predictive model
- A model that estimates an unknown or future outcome from patterns learned in available data.
- Model monitoring
- Ongoing checks of performance, drift, data quality, errors and relevant impacts after a model is deployed.
- Contestability
- The practical ability of an affected person or reviewer to question, correct or appeal a system-supported outcome.
Business Intelligence, AI and Responsible Decisions FAQ
What makes a dashboard decision-ready?
Each metric needs a clear definition, unit, period, source and owner, plus a comparison that supports an action. Refresh status and data-quality limits should be visible so presentation does not imply false certainty.
What does process mining reveal?
It reconstructs actual process paths from event logs and can expose variants, delay and rework. Its conclusions depend on complete case identifiers, event meanings and timestamps, so logging quality must be assessed first.
Why does human review remain necessary?
Models operate within data and design limits, while decisions carry context and consequences beyond a score. Human review provides accountable interpretation, handles exceptions and creates a record for correction and learning.
How should AI error be evaluated?
Separate false positives from false negatives and examine their consequences, frequency and distribution. The more material harm should shape thresholds, review requirements, monitoring and any opportunity to challenge the result.
Exam move
For every dashboard, write the metric contract before judging the chart. For every AI case, answer seven questions: decision, data, model role, human owner, false-positive harm, false-negative harm and monitoring. Practise distinguishing correlation, prediction and causal explanation.
Add one concrete operational control for privacy, fairness, transparency, security and contestability rather than listing principles alone.
Rebuild a dashboard from the decision backwards. Write the action a manager may take, then define the target, comparison, unit, period, refresh timing and source for each required metric. Add a balancing measure that exposes gaming or an unintended consequence.
Choose a visual form only after deciding whether the task is status, trend, ranking, composition or relationship. Write one sentence the chart supports and one it does not.
For process mining, sketch the event-log fields before interpreting a discovered path: case identifier, activity, timestamp and relevant actor or status. Create a missing-event scenario and explain how it could distort the apparent flow.
Compare the discovered process with the designed process, treating difference as evidence to investigate rather than automatic non-compliance.
For AI, separate model output from organisational decision. Build an error table with false positives and false negatives, their consequences, reviewer and remedy. Add controls for data minimisation, access, group performance, explanation, override logging and contestability.
Imagine performance changes after deployment and state the signal, threshold and response. This turns responsible AI from a list of principles into a monitored operating process with accountable human judgement.
Run a pre-mortem on the proposed decision aid. Assume the dashboard or model produced confident but harmful guidance.
Trace possible causes through missing events, shifted definitions, stale refreshes, unsuitable training data, interface failure and reviewer over-reliance. For each cause, name a detection signal, an accountable responder and a recovery action. Compare performance over time and across relevant groups, and preserve overrides as evidence for later review and refinement.
