ICAD Principles / D. Decide / D3
D3

D3: Autonomy is configurable per decision type

Autonomy is configurable per decision type — from fully autonomous execution for routine operations to human-in-the-loop approval where regulatory classification or risk profile requires it.

What does this mean?

D3 addresses the most nuanced question in AI-enabled pharmaceutical operations: how much autonomy should an AI system have? The answer is not a single setting. It is a spectrum, configurable per decision type, per regulatory classification, per risk profile.

Some decisions are routine, low-risk, and high-volume: archiving a completed instrument run, routing a passing system suitability test (SST) result, or updating a fleet health dashboard. These can operate autonomously. Other decisions are high-impact, regulatory-sensitive, or novel: releasing a batch, filing a regulatory deviation, or recommending a process parameter change. These require human review and approval.

D3 requires that the system supports the full autonomy spectrum and that the autonomy level for each decision type is explicitly configured, not assumed.

The autonomy spectrum

D3 defines five levels of operational autonomy, adapted from the Society of Automotive Engineers (SAE) autonomy framework for the pharmaceutical context:

  1. Informational: The AI presents data and analysis. A human makes the decision. The AI has no execution authority. Example: presenting a stability trend summary to a scientist.
  2. Advisory: The AI recommends an action with supporting evidence. A human reviews and approves before execution. Example: recommending batch release with supporting CoA data.
  3. Supervised: The AI executes the action automatically but a human is notified and has a defined window to override. Example: routing SST-passing results to the next workflow step with a 24-hour review window.
  4. Conditional: The AI executes autonomously within defined bounds. Execution outside those bounds escalates to human review. Example: automatically archiving passing results but escalating failing results for investigation.
  5. Autonomous: The AI executes without human intervention. Used only for fully routine, low-risk operations with no regulatory implications. Example: updating instrument fleet health metrics, compiling data summaries.
Operational test

For every decision type in the governed system, can you identify its configured autonomy level? Is that level documented with a rationale linking it to the decision's regulatory classification and risk profile? Can the autonomy level be changed through a governed configuration change (not a code release)? If any decision type operates at an undocumented or unconfigurable autonomy level, D3 is not satisfied.

Regulatory classification drives autonomy

In pharmaceutical operations, the appropriate autonomy level is not a technology decision — it is a regulatory and quality decision. The key factors:

  • GxP classification: Decisions affecting product quality, patient safety, or data integrity under Good Manufacturing Practice (GMP), Good Laboratory Practice (GLP), or Good Clinical Practice (GCP) regulations require documented human oversight proportional to risk.
  • Regulatory impact: Decisions that affect regulatory submissions, compliance status, or product release require higher oversight levels than internal analytical decisions.
  • Novelty: First-time decisions (new product, new method, new site) require more human involvement than repetitions of established decisions.
  • Reversibility: Easily reversible decisions (flagging a sample for investigation) can tolerate higher autonomy than irreversible decisions (releasing a batch to market).
Example — tiered autonomy in QC operations

A QC release workflow configures three autonomy levels: (1) SST evaluation operates at Level 4 (conditional autonomous) — passing SSTs auto-approve and route to the next step; failing SSTs halt the sequence and notify the QC manager. (2) Batch release recommendation operates at Level 2 (advisory) — the AI compiles all results, verifies specifications, and presents the recommendation with supporting evidence; a qualified QC reviewer approves. (3) OOS investigation assignment operates at Level 3 (supervised) — the AI assigns the investigation to the appropriate analyst based on technique expertise and workload; the QC manager has 4 hours to reassign before the assignment takes effect.

Self-driving labs and the governance gap

The industry conversation around self-driving labs and lights-out operations conflates automation capability with operational readiness. A self-driving lab that can execute experiments autonomously is a technical achievement. A self-driving lab that can do so under GxP governance — with traceable decisions, configurable approval gates, and regulatory-grade audit trails — is an operational reality.

D3 does not prohibit high-autonomy operations. It requires that autonomy levels are explicitly configured, documented, and appropriate to the regulatory classification of each decision. The path to higher autonomy runs through governed infrastructure, not around it.

Relationship to other principles

D3 requires D2 (traceability) because configurable autonomy without traceable decision logic is dangerous — an autonomous decision that cannot be reviewed is an unaccountable decision. D3 also depends on D1 (governed inputs) because autonomy levels are only meaningful when the AI operates on trustworthy data. An AI making autonomous decisions on ungoverned data is not automation — it is uncontrolled risk.