How It Works
Logic Insights combines two engines that play very different roles. A deterministic graph engine does the measuring: it compiles each process into a graph and computes metrics, defects, paths, simulations, and standards scores in pure code, so the numbers are exact and reproducible. An AI agent fleet does the interpreting: it turns those numbers into narrative, recommendations, benchmarks, and answers to your questions. Keeping the two separate is deliberate. Measurement should never hallucinate, and interpretation should always be grounded in real measurement.
The analysis pipeline
Analysis is set up in two sequential steps. First you choose a project, then you choose which of its processes to analyze:
- Select a project. You pick the project you want to analyze.
- Select the processes. From that project, you choose the specific processes to include, or take all of them.
- Read. The agent lists and reads the selected processes through the Flowon MCP server, on behalf of the signed-in user. Nothing is written back.
- Compile. Each process becomes a directed graph of stages and transitions.
- Measure. The deterministic engine computes every metric, detects defects and anti-patterns, analyzes paths, runs the simulation, and scores the standards audit.
- Score. The measurements roll up into six sub-scores, a composite health score, and a risk tier.
- Interpret. The AI fleet reads the computed analysis and produces the narrative, redesign recommendations, benchmarking, and regulatory mapping.
- Aggregate. Once every process is analyzed, a portfolio roll-up is produced across the whole project.
- Stream. Results are streamed to the browser over Server-Sent Events as each process completes, so a large project renders progressively instead of blocking.
Architecture
An orchestrator coordinates two layers: a deterministic graph engine that measures, and an AI agent fleet that interprets. The AI fleet always works from the already-computed analysis, so its output is grounded in real numbers.
The deterministic graph engine
Each process is run through a set of specialized, code-only engines. None of these call an LLM, so their output is exact and identical on every run.
| Engine | Responsibility |
|---|---|
| MetricsEngine | Computes all seven metric families and the six sub-scores |
| PathAnalyzer | Finds the happy path, critical path, shortest and longest paths, counts all paths, and identifies exception paths |
| AntiPatternDetector | Flags god stages, spaghetti flows, bottleneck actors, SLA blindspots, and manual islands |
| DesignDefectDetector | Flags structural defects such as orphan and dead stages, cycles without exit, impossible SLAs, and segregation-of-duties violations |
| StandardsAuditEngine | Scores the process Pass, Partial, or Fail against five frameworks |
| MonteCarloEngine | Runs a seeded simulation of completion time and per-stage bottleneck probability |
| ReengineeringAnalyzer | Identifies the structural opportunities that feed the AI redesign layer |
The AI agent fleet
On top of the measurements, a fleet of specialist agents produces everything that requires judgment and language. An orchestrator coordinates the run and fans work out per process.
| Agent | Produces |
|---|---|
| OrchestratorAgent | Coordinates the run and distributes per-process work |
| ProcessAnalysisAgent | Turns the metric set into a readable, business-language narrative |
| RedesignAgent | Prioritized, effort-scored redesign initiatives |
| IndustryBenchmarkAgent | Scores the process against benchmark dimensions |
| RegulatoryMappingAgent | Maps process controls to regulatory frameworks |
| PortfolioIntelligenceAgent | Produces the portfolio-level narrative across all processes |
Because every AI agent works from the already-computed analysis, its output is fast and anchored to real numbers rather than guessed from a diagram. Any single enrichment (redesign, benchmark, regulatory mapping) can be regenerated on demand for one process without re-running the whole analysis.
Reproducibility
Two properties make Logic Insights trustworthy in a review:
- Deterministic measurement. Metrics, defects, paths, and standards scores are computed in code. The same process always produces the same numbers.
- Seeded simulation. The Monte Carlo simulation uses a fixed seed, so the forecast for a given process is stable across runs and comparable over time.