The decision framework used across this site has five questions: Signal (what changed?), Context (what does it mean for us now?), Options (what could we do?), Evidence (why this one?) and Decide (who has the authority?). After the decision come two more steps — Act, where the choice executes in operational systems, and Learn, where the outcome is measured and fed back — which close the loop so every decision improves the next.
The loop in one picture
↺ Every outcome feeds the context of the next decision.
The first five steps are the questions a decision must answer. The last two — Act and Learn — are what turn a good answer into a result and a better next answer. Decision intelligence engineers all seven; most organizations stop at a dashboard somewhere around step two.
Signal: What changed?
A signal is a change that crosses a threshold someone cares about: a supplier slips, a vibration passes alarm, a queue stops clearing. Good signals are specific, timely and tied to an owner.
How to measure it
→ Time from event to detection
→ Share of signals with a named owner
→ False-alarm rate
What goes wrong without it
Dashboards nobody reads; alerts nobody owns; thresholds set once and never tuned.
Context: What does it mean — for us, right now?
Context joins the facts this decision needs — orders, schedules, inventory, policy, history, capacity — and scores its own completeness. When something is missing, it asks rather than guesses.
How to measure it
→ Context completeness (health) score
→ Missing inputs requested vs assumed
→ Time to assemble context
What goes wrong without it
Training a model on everything and grounding it in nothing; decisions made on half the facts.
Options: What could we do?
Real options include doing nothing, the obvious fix and at least one creative alternative. Each carries its cost, benefit, risk and reversibility, scored against the objective by optimization or simulation.
How to measure it
→ Options considered per decision
→ Share of decisions where do-nothing was priced
→ Value gap between chosen and best option
What goes wrong without it
The first idea wins because it arrived first; “wait and see” is treated as free.
Evidence: Why this one?
Evidence combines forecasts, similar past cases, rules and model outputs into a calibrated confidence, with the drivers shown. It is what lets a person trust, challenge and audit the call.
How to measure it
→ Confidence calibration (predicted vs actual)
→ Share of recommendations with drivers shown
→ Override rate and reasons
What goes wrong without it
Black-box scores; confidence that is never checked against outcomes.
Decide: Who has the authority?
Policy sets who decides at what autonomy level — human, augmented, on-the-loop, automated or agentic — based on consequence and confidence. Then the action executes in the system of record and the outcome is measured.
How to measure it
→ Decision latency (signal to action)
→ Share executed within policy
→ Outcome vs expected outcome
What goes wrong without it
AI acting without a mandate — or good recommendations waiting forever for someone to press yes.
After the five questions: act and learn
Act
The chosen option executes in the system of record — a purchase order, a work order, a schedule change — with an audit trail of who approved what and why. See decision execution.
Learn
The outcome is measured against the expected outcome. Approvals, overrides and results become labelled data that tunes thresholds, models and autonomy. See decision observability.
A decision canvas you can use tomorrow
| Question | Write down | Example — late supplier |
|---|---|---|
| Signal | The trigger and its threshold | Supplier 03 more than 2 days late on an order over $100K |
| Context | The inputs this decision needs — and which are missing | Open orders, line schedule, inventory, policy, Supplier B capacity (missing) |
| Options | At least three, including do nothing, each priced | Wait (−$310K) · move to Supplier B (+$216K) · split lines (+$120K) |
| Evidence | Models, past cases and the confidence you need to act | 14 similar cases, Supplier B on-time 96%, 87% confidence |
| Decide | Owner, autonomy level and the policy limit | Procurement lead approves above $100K; automated below |
| Act | The system and action that executes it | PO re-issued in ERP, customer promise date updated |
| Learn | The outcome metric and when you’ll check it | On-time delivery and margin, measured at shipment |
The same five questions in four industries
| Manufacturing | Healthcare | Logistics | Robotics | |
|---|---|---|---|---|
| Signal | Compressor vibration past alarm | 14 patients boarding in the ED | Storm cell over I-35, 14 trucks inbound | AMR drive current +18% mid-mission |
| Context | Historian, CMMS, plan, spares | Census, discharges, staffing | Telematics, windows, driver hours | Telemetry, queue, service bay |
| Options | Run · swap bearings in slowdown · overhaul | Divert · expedite discharges · overflow unit | Hold · reroute 9 · reroute all | Continue · slow and service · stop now |
| Evidence | Wear signature 89% match, 21 days left | Admissions forecast, 11 similar Mondays | Delay model, past storm routes | Bearing signature 91% similar |
| Decide | Reliability engineer approves | House supervisor; clinicians keep clinical calls | Dispatcher on-the-loop | Automated inside the autonomy envelope |
- Answer the five questions in order — each one depends on the last.
- Price “do nothing”; it is never free.
- Measure every question, not just the final outcome.