Parallel processing in SIU: how running 15 investigation phases simultaneously changes the claims workflow

Manual SIU investigation is slow for one reason more than any other: the work is sequential. An investigator pulls the claim file, then reviews prior claims, then queries NICB, then reviews medical records, then runs OSINT, then reviews statements. Each step waits for the previous to finish. Two weeks of calendar time is, in actual work terms, maybe 40-80 hours of focused investigator work stretched across serial handoffs and external dependencies.

Autonomous AI investigation agents do not replace the individual phases. They run them in parallel. A week of sequential work becomes a few hours of parallel work. The compression is architectural, not algorithmic. This piece walks through the 15 investigation phases, which can run in parallel, what dependencies exist, and how the workflow changes at the operations level.

This is a companion to the broader rules-based vs autonomous AI piece. If you want the investigation workflow in its original manual form, start with how insurance companies investigate fraud.

Why sequential investigation is slow

Sequential investigation has three latency sources, all of which compound:

• Active-work latency - the actual investigator time on each task, typically 30 minutes to 4 hours per task.
• Wait-state latency - time spent waiting on external parties (database queries with turnaround SLAs, vendor deliverables, statement recordings), typically 24-72 hours per hop.
• Context-switch latency - the investigator's time cost to pick a case back up after a wait state, typically 15-30 minutes per resumption.

A case that requires six external hops (NICB query, LexisNexis query, medical records request, OSINT scrape, financial records request, witness scheduling) accumulates ~4-5 days of wait-state latency before any investigator work even happens. The investigator's actual focused time on the case might be 30-40 hours; the calendar time stretches to 14+ days.

The 15 investigation phases

A complete autonomous investigation runs 15+ phases per case. The exact phase list varies by claim type (auto, workers' comp, property, liability, medical), but the structure is consistent:

Phases 1-3 are ingest and context. Phases 4-11 are evidence gathering, each operating on independent data sources and independently parallelisable. Phases 12-15 are synthesis, analysis, and output - these have dependencies on the evidence-gathering phases completing, so they wait on phases 1-11 before starting.

Dependency graph and parallelism

A properly architected autonomous investigation agent resolves phase dependencies automatically. The dependency graph has three tiers:

Evidence-tier phases run in parallel bounded only by external API concurrency limits (rate limits from NICB, DMV, medical record providers, etc.). On a typical case, this means ~8 phases completing in 1-2 hours total rather than 1-2 hours each serially (8-16 hours of sequential work).

Synthesis-tier phases have a hard dependency on evidence-tier completion - you cannot synthesise findings until you have the findings. But synthesis is computationally cheap (minutes, not hours) once the evidence is in. The wall-clock time from last evidence phase completing to final report ready is typically 5-15 minutes.

What changes in the workflow

At the SIU operations level, parallel processing changes the workflow in four ways:

1. Investigations complete before the investigator opens the case

The investigator's first view of a case is the completed investigation report with evidence, findings, and recommendation - not an empty case file to start work on. The first action is review, not evidence gathering.

2. Investigator time-per-case drops to review duration

A manual investigator spent 40-80 hours of active work per case across 14+ days. A review-oriented investigator spends 30-60 minutes per case - reading the report, evaluating the findings, making the determination, and signing off. Caseload scales from ~10 to 800+ per month as a direct consequence.

3. Queue dynamics change

Under manual workflows, cases queue because investigators are working on other cases. Under parallel autonomous investigation, there is no 'other cases' blocker - every flagged claim runs through the agent in parallel. The queue shifts from 'waiting for an investigator' to 'waiting for a reviewer', which is a much shorter wait.

4. Coverage rises mechanically

Because the throughput constraint lifts, every flagged claim can receive full investigation. The 75% of flagged claims that currently close without investigation become in-scope. For the economic consequences, see insurance claims leakage: how uninvestigated claims drain profitability.

Error handling and human sign-off

Parallel processing introduces error modes that sequential processing does not have. Three classes matter:

• Partial-failure errors - one of the 15 phases fails (rate-limited API, data unavailable, format mismatch). The system should surface the missing phase in the report, not silently omit it.
• Contradictory-finding errors - two phases produce contradictory findings (e.g., OSINT contradicts a statement). These should be flagged, not resolved by the AI - the investigator decides.
• Confidence-scoring calibration - each finding should have a confidence score; low-confidence findings should be explicitly marked for human review.

The human sign-off step is where these error modes are resolved. An investigator reviewing an AI-generated report is reviewing for: missing phases, contradictory findings, low-confidence items, and context the AI could not have. The sign-off is not a rubber stamp - it is the quality control step that keeps the automated pipeline compliant with the investigator-decides regulatory standard.

Parallel vs sequential benchmarks

For the full benchmark set, see benchmarking SIU performance in 2026. For the economics, see from three weeks to four hours.

Frequently asked questions