Most institutions do not suffer from a shortage of dashboards. They suffer from a shortage of reliable answers. The principal can see asset allocation, cash, projects, incidents and deadlines across polished screens. Then a serious question arrives: Why did this number change? Which source is authoritative? Who can act? What is uncertain? What happens if we wait? The dashboard cannot answer because it was designed to display state, not support judgment. An intelligence layer is different. It connects evidence to decisions.
A dashboard compresses; intelligence preserves meaning
Compression is useful. It is also lossy. The green indicator hides the threshold. The consolidated total hides the entities. The trend line hides a change in methodology. The risk score hides disagreement among advisers. A decision-grade system should let the user move from signal to source without abandoning the context of the question. It should preserve:
- Provenance: where the fact came from, its date, owner and version.
- Transformation: how the source became the displayed figure or conclusion.
- Uncertainty: what is estimated, disputed, stale or missing.
- Authority: who may decide and who may execute.
- Consequence: what changes if the institution acts, waits or is wrong.
Without these, the system creates confidence faster than understanding.
Do not begin with the screen
Begin with recurring high-consequence questions. For example:
- What requires the principal’s decision in the next fourteen days?
- Which obligations or exposures changed materially since the last review?
- Which facts are being reported differently by two advisers?
- Where is authority concentrated or about to expire?
- Which transaction assumption is unsupported by current evidence?
For each question, identify the authoritative sources, freshness requirement, decision owner, action path and evidence to retain. The interface comes after the information contract.
Create decision objects
A useful intelligence layer should not merely produce alerts. It should create structured decision objects. Each object contains:
- The question or decision required.
- The triggering change and its source.
- Relevant facts separated from assumptions and interpretation.
- Available options, constraints and material trade-offs.
- The person or body holding authority.
- The deadline and consequence of inaction.
- The final decision, execution record and review date.
This structure allows humans and AI systems to assist without quietly rewriting the institutional record.
Search is not intelligence
Enterprise search retrieves what matches a query. It may surface ten documents and still leave the user unable to know which one governs. The intelligence layer resolves identity, version, ownership and conflict. It can say that two records disagree, that a mandate expired, that a figure was derived from an estimate or that the apparent answer sits outside the user’s authority. The system should be comfortable returning “not established” when the evidence is insufficient. A fluent answer without provenance is often less useful than an explicit gap.
Design the AI role carefully
AI can be excellent at assembling timelines, finding inconsistencies, extracting obligations and drafting options. It should show its work. It should not invent certainty, conceal missing records or make a binding decision simply because the interface makes that action easy. Use the model to reduce the cost of understanding, not to erase the distinction between source, inference and authority. For sensitive domains, retrieve the minimum context, cite the exact records and retain a human checkpoint before external or irreversible action.
Measure decisions, not engagement
A dashboard vendor may celebrate daily active users and time on screen. An intelligence layer should aim to reduce both if the institution reaches sound decisions faster. Measure:
- Time from material change to decision owner awareness.
- Percentage of consequential assertions with current provenance.
- Unresolved conflicts between authoritative sources.
- Decisions delayed by missing authority or evidence.
- Actions executed outside the approved decision.
- Questions the system correctly refused because the evidence or permission was insufficient.
The interface should become quieter as the system improves
Good intelligence is selective. It does not make every data point compete for attention. It routes routine matters through standing authority and brings forward the few changes that require judgment. The objective is not a more impressive command centre. It is a private decision environment in which the principal can understand what changed, what is known, who can act and what the institution will remember after the action. Build that layer. The dashboard can remain a view inside it.