Decision intelligence

Decision Intelligence (DI) is a discipline that turns data into manageable decisions and a measurable effect. DI integrates causality, forecasting, decision economics, policy design, and MLOps/operations into one life cycle.

1) DI Framework: OODA/SSDL

Observe (Signal): standardized events, quality/freshness, dedup and context.
Orient (Sense): interpretation: cohorts, segments, causal graphs, risk profiles.
Decide: policy (rules/models/bandits), consideration of limitations and cost of errors.
Act: orchestration of actions, channels, idempotence, priorities.
Learn: causal effect estimate, update thresholds/policies/models.

2) Economics of solutions

Value function: revenue/retained damages/retention/quality of service.
Cost of error: FP/FN in money and risk (RG/compliance/reputation).

Expected Value (EV):

[
EV = p_{\text{uspekh} }\cdot Value - p_{\text{vred} }\cdot Harm - Cost
]

Action is allowed if 'EV≥0' and guardrails are normal.
Risk appetite: limits on FPR, frequency of interventions, harm/complaint budget, error budget.

3) Causality and prognosis

When prediction is sufficient: low-risk allocations, ranking by probabilities.
When causality is needed: ROMI, price/limit policy, safety/compliance. Use A/B, DiD, RDD, IV, synthetic control; when targeting - uplift and CATE.
Counterfactual loop: forecast → effect → effect → conversion of uplift/thresholds.

4) Types of policies

Rules (policy-as-code): deterministic, explainable; base and fail-safe.
Score-based: probabilities/rate, hysteresis, cost-sensitive thresholds.
Contextual (bandits): ε -greedy/Thompson for choosing offers/channels.
Sequential (RL): multi-step constrained (safe RL) strategies.
Components: cascade - security/compliance → economy → UX.

5) DI architecture

Data: canonical events (UTC, versions), fichestor (online/offline parity), directory.
Models: registry/versions, calibration, drift monitoring (PSI/KL), PR- AUC/Recall@FPR≤x%.
Semantics and metrics: a single dictionary of KPI/guardrails, SLO freshness.
Policy Engine: decision tables, ABAC/contexts, hysteresis, rate-limits, priorities.
Action orchestrator: guaranteed delivery, retrai, idempotency 'action _ id', DLQ.

Observability: trace'correlation _ id', funnel "signal→decision→action→outcome."

Security: RLS/CLS, PII masking, access and decision log.

6) DI Metrics

Quality of solutions

Decision Precision/Recall: by true success of actions.
Regret/Opportunity Loss: lagging behind optimal policy.
Coverage: the proportion of objects that received an action.
Latency p95: Signal→Decision/Decision→Action.
Fairness/Harms: segment error difference, complaints, appeals.

Business impact

ROMI/ROI actions, uplift @ k, Qini/AUUC.
Net Benefit: effect − cost − harm.
Time-to-Impact: time from signal to measurable result.

7) Decision design

1. Frame the question as an effect: "What is the hold gain from X at Y over T?"

2. Draw a DAG, define confounders/colliders.
3. Choose design: A/B, quasi experiment or net forecast + ex-post score.
4. Define action and alternatives, restrictions and guardrails.
5. Set the value function and risk budget.
6. Describe the policy in the decision table: conditions → action → channels → cooldown.
7. Plan the evaluation: effect metrics, duration, CATE segments.
8. Define the incident runbook and fallback rules.

8) Hysteresis, frequency and collisions

Hysteresis: input/output thresholds are different; prevents interventions from "flashing."

Cooldown: pauses between contacts/constraints on the same object.

Policy conflicts: priority matrix; "security takes precedence."

Quotas/Rate-limit: per channel, segment, user; fair distribution.

9) Levels of autonomy

1. Ad-hoc: a person decides there is not enough data.
2. Assisted: the system offers a solution + explanation.
3. Automated: auto-solutions within guardrails.
4. Adaptive: auto-tuning thresholds/selection of offers (bandits).
5. Safe-Autonomy: Autonomy under formal restrictions and auditing.

10) Solutions under uncertainty

Scenario planning: basic/stress/extreme; effect ranges.
Robustness: a strategy that is robust to parametric errors.
POMDP intuition: act with incomplete information; value the cost of information (what experiment to do).
Bayesian updating: Combine historical knowledge and current data.

11) Model ↔ policy dialogue

The model produces a rate/distribution of results.
The policy takes into account the cost of errors, limitations and fairness.
The partition line is in an explicit decision threshold policy with a version log.
Threshold revision - by EV, not only by ROC/PR.

12) Documents and artifacts

Policy Passport (template)

Code/version, purpose and KPI of the effect

Conditions/features/model, hysteresis/cooldown

Actions and channels, priorities and mutual exceptions

Guardrails (FPR≤x%, latency p95≤y, RG/compliance)

Score: test design, metrics, duration

Audit/Explanation to User, Owners

Decision Table (example)

Condition	Context	Action	Cooldown	Guardrails
`churn_uplift ≥ 0. 08` & `value_q ≥ 0. 8`	retention	offer L	7d	ROMI≥0, cap=1
`rg_risk ≥ τ` & `night`	RG	pause + tip	1d	FPR≤1%
`fraud_score∈[τ1,τ2]`	payment	manual check	—	SLA 2h

"End-to-end" solution logging scheme

`signal_id` → `decision_id` → `action_id` → `outcome_id` (+ `correlation_id`).

13) Governance and compliance

Unified dictionary of metrics and formula versioning.
Policy Committee: Risk Officer, Product, Data, Compliance.
Audit of decisions: explanations, reasons for refusals, appeal channels.
Ethics and fairness: monitoring errors by group; exclusion of protected features from the rules, where required by law.

14) Frequent errors

Optimization of proxy metrics instead of business effect (Goodhart).

Confounding predictions and causality; ROMI "by correlation."

Absence of hysteresis and cooldowns → spam/" blinking. "

Unaccounted for cost of errors and user harm.
Quiet edits of thresholds/formulas without versions and changelog.

Actions without effect evaluation and "cycle closure."

15) Pre-Release DI Policy/System Checklist

Objective is stated as causal effect, value function and risk budget are given
DAG drawn; assessment design (A/B/DiD/SC) and metrics selected
The policy is described in the decision table; have hysteresis/cooldown/priorities
Models are calibrated; thresholds derived from error cost (EV)
The orchestrator of actions is idempotent; signal→decision→action→outcome Log Enabled
Guardrails and alerts are configured; runbooks and fallback rules are ready
Dashboards: funnel of decisions, effect (uplift/ROI), harm/complaints, fairness
Versions/owners/access rights/compliance documented

Total

Decision intelligence is a system, not a set of models: uniform data and metrics → a causal and economic view of effect → explicit policies and safe orchestration → rigorous evaluation and continuous learning. Such a system reduces risk, increases ROI, and makes decisions reproducible, explainable, and manageable.

Decision intelligence