Intelligence · Module
Decision Engine
From signal to root cause to approved action to a measured, calibrated outcome
Decision Engine is planned to turn signals into decision records with evidence-labelled root causes, options including "do nothing", and a pre-registered expectation. Execution runs only through the owning app's Actions, and outcomes are measured and calibrated.
Closing the loop dashboards leave open
Dashboards tell you something went wrong. Decision Engine is designed to take the next steps: why it happened, what the options are, who approves, what was done and whether it worked. An alpha is planned for W3 within the Enterprise Brain, with GA in W4. It needs BI Studio and Context Graph.
Decision records
Each record holds the question, level (strategic, tactical or operational), reversibility, domain, context, options, the chosen option, rationale, success criterion, confidence, review date and goal link. Before execution, a pre-registered expectation is recorded: the metric, direction, P10/P50/P90 range and window. It cannot be changed afterwards.
Diagnosis with honest evidence
Signals arrive from dashboard alerts and anomalies, BI anomaly explanations, connector health, ROI below threshold and later forecast breaches, and land in a triage queue. Root-cause v1 combines metric-tree decomposition, segment drill-down with minimum sample sizes and multiple-comparison control, Context Graph traversal and change-point alignment. Each hypothesis carries an evidence label: observed correlation, quasi-experimental evidence or experimentally confirmed. Causal language needs at least quasi-experimental evidence.
Options and a critic
Options come from organisational memory, similar past decisions and domain templates, and "do nothing" is always one of them. Each option shows an impact range from history, cost, time to effect, reversibility and risks. A critic reviews against a domain checklist: missing options, optimistic estimates, goal conflicts and second-order effects. Critical findings block progress unless overridden with a reason. Recommendations below the data-quality threshold are refused with a fix list.
Execution and learning
Recommendation cards offer Approve, Test first or Dismiss with reason, in the app, the Management Cockpit and the Approvals Inbox, and later on WhatsApp with step-up for high-risk tiers (policies to be verified). Execution only happens through the owning app's Actions, with approvals by risk tier and the Delegation-of-Authority matrix. Agents may propose but never approve their own recommendations. At the review date, actual is compared with expected. Calibration (Brier score, interval coverage, WAPE) is tracked by actor, agent, model and domain once enough decisions are resolved. Workforce, credit, clinical-adjacent and legal-outcome decisions are advisory only; a person decides.
Features
What Decision Engine includes
Decision records
Question, level, reversibility, options, chosen option, rationale, success criterion, confidence, review date and goal link.
Pre-registered expectation
Metric, direction, P10/P50/P90 and window, recorded before execution and immutable after.
Signal triage
Alerts, anomalies, connector health and ROI thresholds in one queue with owner assignment.
Root-cause engine
Metric trees, segment drill-down, graph traversal and change-point alignment with evidence labels.
Option generation with "do nothing"
Impact ranges, cost, time to effect, reversibility and risks per option.
Critic review
Domain checklist; unresolved critical findings block approval unless overridden with a reason.
Recommendation cards
Approve, Test first or Dismiss with reason, in app, cockpit and Approvals Inbox.
Execution through Actions only
Approvals by risk tier, reversibility and Delegation of Authority; autonomy budgets respected.
Data-quality gate
Recommendations refused on metrics below threshold, with a fix list.
Outcome tracking and Decision Journal
Review reminders, actual versus expected, and search across past decisions.
Experiments and causal checks (W4)
A/B, holdout or geo tests and difference-in-differences or synthetic control with refutation tests.
Calibration
Brier score, reliability, interval coverage and WAPE by actor, agent, model and domain.
ROI linkage
Agent-executed decisions post investment and return to the Investment & ROI Ledger.
Decision capture from approvals
Approved agent decisions above R2 create decision records automatically.
Workflow
Signal to calibrated outcome
- 1
A signal enters the triage queue and gets an owner
- 2
Root cause is drafted with evidence labels
- 3
Options are generated, including "do nothing"
- 4
The critic reviews the options
- 5
The owner approves, tests first or dismisses, and registers an expectation
- 6
Owning-app Actions execute with approvals
- 7
At review, actual is compared with expected and calibration updates
AI
AI skills and Digital Employees
Root-cause drafting
Hypotheses with evidence levels; always reviewed before options.
Option generation
Options with ranges from history, always including "do nothing".
Critic review
Always-on checklist review that can block approval.
Decision capture
Creates decision records from approvals above R2.
Quasi-experimental checks
Upgrades evidence levels when tests support it, with audit.
Digital Employees that use this module
They act through the same governed Actions as people; risky steps wait for approval.
Frequently asked questions
Can the Decision Engine change data on its own?
No. It never writes to app tables. Every effect is an Action of the owning app with its own approvals.
Does it claim causes?
Only with at least quasi-experimental evidence. Otherwise findings are labelled observed correlation.
What does it require?
BI Studio and Context Graph. An alpha is planned for W3 and GA for W4.
Does it decide on people or credit?
No. Workforce, credit, clinical-adjacent and legal-outcome decisions are advisory only; a person decides.
See Decision Engine in your workflow
Tell us about your team and the outcomes you want. We will walk you through the apps, the Digital Employees and the plan for your industry.