The Problem
Large language models can interpret objectives, reason about possible actions, and operate software tools. Yet most AI interactions remain temporary: a request arrives, a response is produced, and the computational context disappears.
Real operational work is different. It runs for hours, days, or weeks. It waits on external events, requires human approval, encounters partial failures, and must resume safely after an interruption. Building AI systems for that environment takes more than prompt engineering. It takes a durable model of execution.
From AI Responses to Cognitive Execution
SmartFlow is Aon Intelligence's research program examining how AI reasoning can participate in long-running, governed workflows.
The central idea is to separate cognitive decisions from operational guarantees. AI may interpret an objective, classify a situation, select a capability, or recommend a next action. The surrounding execution system remains responsible for preserving state, enforcing constraints, validating outcomes, and recovering from failure.
That separation lets probabilistic intelligence contribute where interpretation and adaptation are genuinely valuable — without asking the model to become the system of record.
Research Questions
- How should an AI system represent progress toward an objective?
- Which decisions belong to a model, and which must stay deterministic?
- How does a workflow resume after a process, service, or model fails?
- How should uncertainty affect execution and escalation?
- When should a person review, approve, modify, or stop an AI-selected action?
- How can an execution history support evaluation, audit, and later learning?
- Can the system change its plan without losing commitments and evidence already established?
A Durable Cognitive Loop
Conceptually, SmartFlow treats AI-enabled work as a repeating loop:
- Observe the current state and relevant evidence.
- Interpret the situation in relation to the objective.
- Select a constrained next action.
- Execute through an approved capability.
- Validate the result.
- Record the resulting state and evidence.
- Continue, wait, escalate, compensate, or conclude.
The workflow must stay inspectable throughout. Decisions, tool calls, approvals, failures, and state changes should form a durable execution record — not disappear into a model transcript.
Research Direction
Our work focuses on durable state, bounded action selection, recoverable execution, evaluation, human oversight, and the interaction between deterministic workflows and probabilistic models.
We're particularly interested in measuring more than whether a final answer looks correct. A cognitive execution system should also be judged on completion reliability, unsupported actions, recovery effectiveness, human intervention rates, repeatability, execution cost, and the quality of its decision history.
Why It Matters
AI systems will increasingly be asked to perform work rather than merely discuss it. Without durable execution, they stay difficult to trust in operational environments.
SmartFlow is meant to bridge that gap — letting AI participate in complex work while preserving the control, visibility, and resilience expected of serious software systems.