Decision context is the specific data, business rules, history, constraints and human knowledge that one decision needs — not all the data an organization owns. Good decision intelligence assembles that context per decision, scores whether it is complete and fresh, and asks for what is missing instead of guessing.
Why context beats more data
Throwing every table and document at a model increases cost and noise without guaranteeing the one fact that matters is present. Decisions fail when a critical input is missing, stale or misread — the supplier’s real capacity, today’s crew roster, the patient’s latest lab, the robot’s last calibration. Context engineering starts from the decision and works backwards to the inputs it requires.
The four layers of context
Context health: a score every decision should carry
Before recommending anything, a decision intelligence system should report context health — how complete and fresh the required inputs are.
| Input | Source | Freshness required | Status |
|---|---|---|---|
| Open orders | ERP | < 1 hour | ✓ 12 minutes old |
| Line schedule | MES | < 1 shift | ✓ current |
| Inventory | IoT / WMS | < 15 minutes | ✓ live |
| Procurement policy | Policy store | Current version | ✓ v4.2 |
| Supplier B capacity | Supplier portal | < 24 hours | ✗ missing — requested |
With one critical input missing, context health might read 61%. The right behavior is to request the missing input and hold or lower confidence — not to produce a confident answer anyway.
What to do when context is missing
- AskRequest the missing input from the system or person who owns it, with a deadline.
- DegradeRecommend a conservative, reversible option and say why.
- EscalateHand the decision to a human with the gap highlighted.
- AbstainIf the decision is out of the domain the system knows, say so and do nothing.
Example: the same decision, two contexts
| Without decision context | With decision context | |
|---|---|---|
| Signal | Supplier 03 is four days late | Supplier 03 is four days late |
| Known | Open orders | Orders, schedule, inventory, policy, lead-time history |
| Missing | Unknown — not tracked | Supplier B capacity — requested with a 2-hour deadline |
| Recommendation | Move to Supplier B (confident) | Hold at 61% context health; move to Supplier B once capacity confirms |
| Result | Supplier B can’t deliver; orders miss | Capacity confirmed; orders ship Thursday |
Context for LLMs and agents
For large language models and agents, decision context is delivered through retrieval, tool calls and structured state rather than training. The same rules apply: retrieve for the decision, cite what was used, detect contradictions between sources, and never let an agent act on context it could not verify. See decision safety.
- Start from the decision and work back to the inputs it needs.
- Context includes rules, history and human knowledge — not just data.
- Score context health and make missing inputs visible.
- Missing context should change behavior: ask, degrade, escalate or abstain.