Bounded Rationality to AI Decision Support
AI expands what organizations can retrieve, compare and model. It does not abolish the limits of attention, information, time, incentives or judgement. It changes where those limits sit.
The managerial promise of AI is often framed as better decisions through more intelligence. The older and harder problem is different: organizations must decide under constraints. Herbert Simon’s bounded rationality remains useful because the constraint was never simply insufficient computation.
Decision-makers work with incomplete information, limited time, finite attention, uncertain consequences and organizational goals that are often multiple or contested. They search selectively. They simplify. They use routines and representations. They frequently accept a satisfactory course of action rather than exhaustively identify an optimum.
The decision-support lineage
The important continuity is that every generation changes the representation available to the decision-maker. Dashboards make states visible. Optimization models compare feasible choices. Predictive models estimate likely outcomes. Generative systems can assemble evidence and articulate alternatives. Agents can retrieve, sequence tools and prepare or execute bounded actions.
None of those capabilities determines which organizational objective should dominate, what risk is acceptable, whose interests count, or who is accountable for the consequence.
AI changes the economics of cognition
Search becomes cheaper
Large bodies of documents, policies, operational records and external information can be interrogated quickly. The practical boundary shifts from “can we find it?” toward “is this the right evidence, current enough, authoritative enough and complete enough?”
Alternatives become cheaper to generate
AI can produce scenarios, counterarguments and candidate actions at low marginal cost. This can counter narrow search—but volume is not diversity if alternatives inherit the same assumptions, source limitations or optimization target.
Unstructured evidence becomes computationally accessible
Meetings, reports, correspondence and technical narrative can enter decision workflows that historically privileged structured databases.
Decision support can become action-capable
The largest architectural break is not conversational output. It is the connection of models to tools, permissions and workflows. Once a system can change a record, trigger a process, commit resources or communicate externally, decision support becomes an authority-design problem.
The organization is still bounded
| Persistent bound | AI can improve | AI can worsen | Design response |
|---|---|---|---|
| Incomplete information | Retrieval, synthesis, anomaly discovery | Plausible completion of missing evidence | Source provenance and explicit uncertainty |
| Limited attention | Compression and prioritization | Overproduction of fluent analysis | Decision-specific evidence views |
| Uncertain consequences | Simulation and prediction | False precision | Ranges, assumptions and scenario comparison |
| Conflicting objectives | Surface trade-offs | Hide value choices inside optimization | Explicit objectives and decision rights |
| Cognitive bias | Challenge assumptions and generate alternatives | Automation bias and selective adherence | Structured challenge and accountable review |
| Organizational politics | Make evidence more visible | Give contested choices a technical veneer | Named ownership and recorded rationale |
OECD work on AI in government highlights automation bias: people may overweight algorithmic recommendations, fail to recognize errors and reduce independent judgement. NIST likewise treats human roles and responsibilities as part of AI risk management and notes that Human–AI configurations range from additional opinion to autonomous decision-making.
Do not automate “the decision.” Architect the decision system.
A useful enterprise architecture separates what is known, what is inferred, what is recommended, who judges, who has authority, and what happened afterward.
Match authority to consequence
A mature system does not ask whether humans or AI should make decisions in the abstract. It allocates work according to reversibility, consequence, evidence quality, uncertainty, regulatory obligation and the cost of delay.
| AI role | Typical capability | Authority boundary |
|---|---|---|
| Retrieve | Assemble relevant evidence | No decision authority |
| Interpret | Summarize, classify, detect patterns | Human validates consequential interpretation |
| Recommend | Rank alternatives and explain trade-offs | Named human decision owner |
| Prepare | Draft action, transaction or communication | Approval before commitment |
| Execute bounded | Act within predefined limits | Controls, monitoring, exception escalation |
| Commit | Make consequential organizational choice | Reserved or tightly governed where impact is material |
NIST’s AI RMF operationalizes risk management through Govern, Map, Measure and Manage and explicitly places AI in sociotechnical context. OECD’s 2026 due-diligence guidance similarly emphasizes embedding AI into management systems, identifying impacts, mitigating them, tracking results and adapting practice. The common implication for decision support is operational: oversight must be designed into the work system rather than appended to a model.
A decision-first AI probe
- Select one recurring, consequential decision—not a generic assistant use case.
- Map the evidence actually used, including gaps, latency and disputed sources.
- Record the current alternatives, judgement points, decision rights and cycle time.
- Introduce AI first at retrieval, synthesis or alternative-generation layers.
- Require source-linked recommendations and explicit assumptions.
- Run in parallel with the existing process before granting action authority.
- Measure decision cycle time, evidence quality, reversals, exceptions and outcomes—not merely user satisfaction.
- Increase authority only where evidence shows the surrounding system can detect error, contain consequence and learn.
Selected sources
- Herbert A. Simon, Administrative Behavior (1947 and subsequent editions): bounded rationality, administrative decision-making and organizational choice.
- Herbert A. Simon, Nobel Memorial Prize lecture and biographical materials, Nobel Prize, 1978.
- NIST, AI Risk Management Framework 1.0 and AI RMF Playbook, 2023; current revision programme noted 2026.
- NIST AI Resource Center, Human–AI Interaction appendix: roles, responsibilities and configurations in AI-supported decision-making.
- OECD, Governing with Artificial Intelligence, 2025: automation bias and operational risks in AI-assisted public decision-making.
- OECD, Due Diligence Guidance for Responsible AI, 19 February 2026.
AI expands the feasible evidence and option space. It does not choose the organizational objective.
The computational boundary moves outward: more evidence can be retrieved, more alternatives generated and more consequences modelled. The decision remains bounded by attention, objectives, uncertainty and legitimate authority.
Evidence enters bounded
Records are incomplete, sources conflict, time is finite and some relevant context remains tacit or unavailable.
Objectives + trade-offs + authority
People still determine which goals dominate, what uncertainty is tolerable, whose interests count and who may commit resources or accept consequence.
DENFORTH synthesis. The principal shift is not from bounded to unbounded rationality; it is a relocation of the bounds as computational search and synthesis become cheaper.
Capability can expand faster than authority.
AI changes the cost of search, interpretation and preparation. Consequence, uncertainty and accountability determine whether that capability should cross the next authority boundary.
Assemble evidence and expose provenance.
Summarize, classify and surface patterns.
Compare alternatives and articulate trade-offs.
Draft an action, transaction or communication.
Act inside explicit limits with monitoring and escalation.
Make a consequential organizational choice.
Figure · Decision authority field. The relevant design question is not whether AI can perform a step, but whether the surrounding work system can detect error, contain consequence and assign accountability at that step.