Playbook · Part D Anti-Overreliance Triad
The AI-era discipline. Three protections, applied together.
Anti-confirmation defences (Part C) are always-on but undifferentiated. The Triad adds frontier-position uncertainty as a prior-stage defence — an ex-ante estimate that an AI-augmented task is inside, at-edge, or outside the AI tool's capability frontier — and routes low-confidence outputs to additional scrutiny before they enter the engagement chain.
The empirical anchor.
The Triad is anchored on Dell'Acqua and colleagues' jagged-frontier finding (Dell'Acqua et al. 2026) — a pre-registered randomised field experiment with 758 BCG consultants demonstrating that AI assistance produces materially different outcomes inside vs outside an invisible "jagged" capability frontier.
Inside-frontier tasks: +33.9% quality, +12.2% completion, 25.1% faster.
Outside-frontier tasks: 19 percentage points correctness lost — while AI makes the wrong answers more persuasive (+17.9–25.1% subjective coherence quality on incorrect outputs).
The mechanism: highly-trained domain experts overrelied on AI for outside-frontier tasks without ex-ante frontier-position awareness.
Naming whether you're inside the model's reliable zone.
For each decision class in an engagement, the practitioner names: inside-frontier (high confidence the model handles this), at-edge (uncertain), or outside-frontier (high confidence the model doesn't reliably handle this). The judgement is treated as a first-class governance artefact, not an unstated assumption.
Specification
- For each AI-augmented task in the engagement chain, name inside / at-edge / outside-frontier.
- Refresh at the cadence of the underlying capability shift — fast-moving model capabilities: at engagement start; stable tasks: quarterly.
- Outside-frontier and at-edge outputs are routed to D.2 disconfirming-signal authoring before entering the engagement chain.
Write the signal down before the answer arrives.
For each AI-assisted output, produce in advance the specific signal that would indicate the output is wrong. Look for it. Act on it when it appears.
Specification
- The discipline lives in writing the signal down before the answer arrives. Authoring disconfirming signals after the AI output exists is too late — the answer biases the authoring.
- Each disconfirming signal has an observable indicator the practitioner can check.
- Where the disconfirming signal surfaces during the engagement, the corresponding AI-assisted output is treated as suspect and rerun through human-only analysis.
The discipline is in writing the signal down before the answer arrives. Once the answer exists, the framing of what would disconfirm it is already compromised.
Where AI was wrong, and the organisation didn't catch it.
Periodic structured review of AI failure modes across the engagement portfolio. Honest learning loops, instrumented at the Reflexive stock level. The review's value is in admitting what was missed; the format that doesn't permit admission produces no learning.
Specification
- Engagement-portfolio-level, not engagement-specific: patterns of AI failure across engagements compound into Pattern Library entries (Part E).
- The review feeds the methodology's own consequentiality discipline (Part C.6).
- Where AI-assisted findings have proved wrong, the Pattern Library carries the amendment.
How the Triad sits inside the engagement protocol.
The Triad runs in parallel with the five-phase engagement protocol, not as a separate workstream. Each phase has Triad-specific protections:
-
B.1 Signal Selection
Frontier-position estimate for any AI-assisted signal-detection tooling. Outside-frontier signal classes routed to human-only detection.
-
B.2 Delay Diagnosis
Disconfirming signals authored in advance for any AI-assisted measurement (e.g. transcript analysis, sentiment classification). At-edge measurements re-run independently.
-
B.3 Bottleneck Finding
Output 3b alternative interpretation is itself protected by Triad discipline. Where AI augmented the binding-constraint finding, Output 3b is human-authored and AI-checked, not AI-authored.
-
B.4 Intervention Design
The intervention prescription is the most consequential single artefact. Frontier-position estimate applied to any AI-augmented prescription tooling. Outside-frontier prescriptions are human-only.
-
B.5 Consequentiality Review
Anti-confirmation review (D.3) runs alongside consequentiality follow-up. Where AI-assisted findings proved wrong, the Pattern Library amendment is filed.
Lineage & references for Part D
- Dell'Acqua et al. 2026 Dell'Acqua, F., McFowland III, E., Mollick, E. R., Lifshitz-Assaf, H., Kellogg, K., Rajendran, S., Krayer, L., Candelon, F. & Lakhani, K. R. (2026). Navigating the Jagged Technological Frontier. Organization Science 37(2): 403–423. DOI 10.1287/orsc.2025.21838. The empirical anchor for the entire Triad — frontier-position uncertainty added in response to this finding.
- Olivier 2026 (SRD Working Paper v0.3.1) Olivier, A. (2026). Signal-Response Distance: A Diagnostic Construct for AI-Era Organisational Responsiveness. Working Paper v0.3.1. §6.8 frontier-position uncertainty amendment. Soaring Wings Consulting.