Skip to content

Accuracy Tolerance Principle

The Question

When evaluating whether a workflow can use probabilistic AI (LLMs, agentic patterns) vs. requiring deterministic processing, ask: what happens when the output is wrong 2-5% of the time?

If humans review the output before acting on it → probabilistic is acceptable. If the output feeds directly into downstream systems without human review → deterministic processing is required.

Decision Framework

Workflow Accuracy Requirement Tolerates AI? Reasoning
Document text extraction ~95% acceptable Yes OCR errors exist regardless; lawyers review extracted text before relying on it
Transcript summarization ~95% acceptable Yes Summaries are presented alongside source material; humans verify before citing
Search indexing Fuzzy by nature Yes Relevance ranking is inherently probabilistic; precision/recall tradeoffs are expected
Document classification ~95% acceptable Yes Misclassified documents surface during review; bulk reclassification is available
PR code review Advisory Yes Findings are suggestions, not automated actions; developer decides what to act on
Bates stamping 100% required No Sequential numbering with zero gaps or duplicates; legally binding page identification
Billing / metering 100% required No Financial accuracy required; errors create legal and contractual liability
Case database operations 100% required No Data integrity is foundational; corrupted case data is unrecoverable
Event routing 100% required No Misrouted events cause data loss or duplicate processing; must be deterministic
Migration / ETL 100% required No Data transformation must be exact and reproducible; rollback depends on it

The Pattern

Workflows split into two categories along a clear boundary:

Probabilistic-tolerant (AI-suitable): - Output is presented to humans for review before action - Errors are detectable and correctable after the fact - The alternative (no processing) is worse than imperfect processing - Volume makes manual processing impractical

Deterministic-required (traditional processing only): - Output feeds directly into downstream systems or legal records - Errors compound silently through the pipeline - Correctness is contractual or regulatory requirement - The system cannot self-detect when output is wrong

Applying This to New Features

When proposing a new feature that might use AI:

  1. Identify the output consumer — Is it a human reviewing results, or a downstream system consuming data programmatically?
  2. Assess error impact — If the output is wrong 2-5% of the time, does someone notice before harm occurs? Or does it silently propagate?
  3. Check for correction mechanisms — Can errors be found and fixed after the fact? Is there a feedback loop?
  4. Consider the alternative — Is the current manual process actually 100% accurate? If humans already make errors at similar rates, AI with human review may be a net improvement.

If the answer to #1 is "human" and #3 is "yes" → AI is appropriate. If the answer to #1 is "system" and #2 is "silently propagates" → deterministic only.

Relationship to Other Principles

  • Guardrails pattern (patterns/guardrails.md): Gate 2 (output verification) and Gate 3 (human approval) scale based on accuracy tolerance. Deterministic workflows need lightweight output checks; probabilistic workflows need stronger verification and human review gates.
  • Exception hierarchy (patterns/exception-hierarchy.md): Deterministic workflows use Recoverable/Permanent/Silent routing. Probabilistic workflows may need a confidence threshold that routes low-confidence outputs to human review rather than the DLQ.
  • Checkpoint pipeline (patterns/checkpoint-pipeline.md): Plan-and-Execute orchestration is deterministic by design. Probabilistic AI features should use the Hierarchical pattern (parallel agents + aggregation) instead.
Ask the Architecture ×

Ask questions about Nextpoint architecture, patterns, rules, or any module. Powered by Claude Opus 4.6.