Decision cycles

1) What is the decision cycle

• Question → Data → Analysis/Insight → Solution → Action → Effect Measurement → Training → (new) Question.

• OODA (Observe-Orient-Decide-Act) - loop speed is more important than pitch "ideal."
• PDCA (Plan-Do-Check-Act) - quality control and continuous improvement.
• DIKW (Data-Information-Knowledge-Wisdom) - the degree of abstraction from facts to rules.

The goal is to reduce the time from event to action and improve the quality of cost-to-decision solutions.

2) Roles, rights and responsibilities

Decision Owner-Responsible for alternative selection and risk.
Analyst/Data Scientist: formulates a hypothesis, selects a method, considers the effect.
Business owner of metrics: fixed KPI definitions, target thresholds, guardrails.
Operations/Engineering: provides data, tools, SLO, automation.
Compliance/Risk: acceptable risk parameters, privacy and compliance.

Practices: RACI/RAPID, escalation matrix, threshold/rule change rights.

3) Solution typology and contours

Operating (minutes/hours): incidents, alerts, limits, anti-fraud.
Tactical (days/weeks): campaigns, pricing, allocating budgets, UX experiments.
Strategic (quarters/years): product portfolio, markets, architectural principles.

For each type, define: cadence, SLA solutions, escalation channels, reporting format.

4) Reference cycle (process skeleton)

1. The question and hypothesis is to formulate the problem, target metrics (primary/guardrail), MDE.
2. Data and context - sources, freshness, quality, semantic definitions.
3. Analysis/modeling - stat ./ML-methods, scenarios, sensitivity, risks.
4. Decision - selection criteria, risk limits, approval.
5. Action/implementation - feature flags, instructions, responsible persons, deadlines.
6. Effect measurement - experiment design/observability, confidence intervals.
7. Retrospective - lessons, update of standards/thresholds, documentation.

Artifacts: one-pager template, solution card, rollback runbook, assumption log.

5) Decision KPIs

Decision Latency - The time from event detection to the selected action.
Time-to-Insight: from request to correct insight.
Time-to-Action: from insight to execution (including approvals).
Win-Rate solutions: the proportion of solutions that have a statistically significant positive effect.
Effect Size/Uplift - the magnitude of the impact on the primary KPI (and confidence interval).
Cost-to-Decision: money/hours to prepare and execute a decision.
Coverage: the proportion of processes closed by formalized loops (i.e. owner, SLO, runbook).

It is recommended that you enter a Decision Scorecard on the product/process.

6) Architecture of data and tools for the cycle

Collection/delivery: streaming (Kafka/PubSub), CDC, ELT; circuit contracts, freshness tests.
Storage/display cases: Lake + DWH/OLAP; HTAP as needed; aggregates/roll-ups.
Semantic layer: uniform KPI formulas, versions, owners, RLS/CLS.
Insight delivery: adaptive dashboards, prioritized alerts, recommendations/NBA.
Experiments: feature flags, A/B orchestration, experiment log, MDE calculators.
Automation: rules/policies (rule engine), action orchestrators, APIs to systems.
Observability: logs, metrics, traces; auditing decisions and exports.

7) Solution design and risk control

Guardrails: security metrics (e.g. retention, resiliency, complaints).
Threshold politicians: who changes thresholds, how they are validated, how they roll back.
Data confidence: quality tests, lineage, explainability of models (SHAP).
Ethics and privacy: PII masking, RLS/CLS, DSAR, storage localization.

8) Experiments and causality

Randomization/stratification, power analysis, CUPED/permutations, adjustment for multiple validations.
Quasi-experiments (DiD, synthetic control) when RCTs are not possible.
Decision-as-Code - Store hypotheses, metrics, and success criteria in the repository.

9) Speed vs quality: trade-offs

Fast path: pre-agreed runbook actions (auto-app ↔ low risk).
Safe path: full check and A/B (high risk/error cost).
Dual track: quick "trial" solutions for parallel evidence collection.

10) Decision Automation

Rules → ML → RL: from thresholds and heuristics to models and contextual bands.
Human-in-the-Loop: Operators confirm/adjust system offerings.
Explain & Override - explain the reasons for the decision, the ability to temporarily override.
Versioning/rollbacks: rule/model version number, rollback policy.

11) Visual and UX patterns

Priority Tape: Alerts and Descending Delay Cost Solutions.
Solution card: problem → alternative → expected effect → risk → owner → deadline.
Drill-through: From KPIs to primary events/cases for hypothesis testing.
Zero-click insights: brief conclusions and ready-made actions right in the card.

12) Solution catalog and organization memory

Repository: templates, past cases, effects, anti-patterns.
Search and tags: by metrics, domains, risks, owners.
Reuse: "recipes" for recurring situations (incidents, seasonality).

13) Antipatterns

Decisions on correlations without experiment/causal methods.
Chameleon metrics: different KPI formulas in different reports.
Alert storm: no prioritization, deduplication, snooze and runbooks.
Lack of owner's: "collective irresponsibility," prolonged latency.
Broken feedback-loop: the effect is not measured → the organization is not learning.
Complex live requests to OLTP: degradation of productive systems.

14) Implementation Roadmap

1. Discovery: solution map (JTBD), critical KPIs, risks/limitations; assign owners.
2. MVP cycle: 2-3 priority cases; Solution card template basic alerts; A/B infrastructure.
3. Scale: KPI semantic layer, recipe library, alert prioritization, Decision Scorecard.
4. Automation: fast path rules/models, human-in-the-loop, auditing, rollbacks.
5. Optimization: cost-to-decision, bandits/RL, staff training, regular retro.

15) Pre-release checklist

• Fixed solution owners and escalation matrix.
• Primary/guardrail metrics, target thresholds and MDE defined.
• Semantic layer and data quality tests are included in the CI.
• Configured alerts with prioritization, deduplication, and snooze.
• There are feature flags and safe rollback; a log of decisions and actions.
• Describes privacy policies (RLS/CLS, PII masking), auditing enabled.
• Experiments and quasi-experiments documented; there are power calculators.
• Decision Scorecard and retrospective rituals are scheduled in the calendar.

16) Maturity levels

L1 Adic-hoc: point solutions, metrics are heterogeneous, effects are not measured.
L2 Process: there are templates and owners, but weak automation.
L3 Insight product: semantic layer, default A/B, solution catalog.
L4 Automated loops: fast path with rules/ML, human-in-the-loop.
L5 Self-learning system: RL/bands, budget guards, end-to-end audit and explainability.

17) Sample Solution Templates (Quick Billets)

"KPI X anomaly ": if delta> T and guardrail-metrics are normal → start Y mode for Z hours; otherwise escalation.
"Budget redistribution": once a week compare the ROI of the channels; if ROI_A/ROI_B> R → offset Q%.
"Charn-risk": at p (churn)> P and margin> M → offer S; log uplift.
"SLO Incident": at p95> S and cause - bottleneck N → trigger rollback plan/workaround scenario.

Bottom line: Effective decision cycles are not a report or a meeting, but an engineering loop that connects data, people, tools, and rules into a repeatable system. Reduce latency, increase the proportion of confirmed effects, automate secure "fast path," learn every cycle - and your organization's intelligence will grow predictably and controllably.