Early development · ATLI 1.9

Depth of hands-on use — scored 1 of 5 (Frequency, range, and duration of personal AI engagement.) Candidate describes AI through briefings, training programmes, or delegated demonstrations. Cannot name specific tools used in the past 30 days or specific personal outputs. References AI in third person ("our team uses..." or "we have access to..."). No artefacts uploaded, or uploaded artefacts are obviously not personally authored. Self-reported hours/week ≤ 2.

Anchor L1

Candidate describes AI through briefings, training programmes, or delegated demonstrations. Cannot name specific tools used in the past 30 days or specific personal outputs. References AI in third person ("our team uses..." or "we have access to..."). No artefacts uploaded, or uploaded artefacts are obviously not personally authored. Self-reported hours/week ≤ 2.

Anchor L3

Uses 1–2 tools weekly for clearly described task types. Can describe specific recent outputs (a memo drafted, a code snippet generated, an analysis run). At least one workflow described in concrete terms. Uploaded artefacts include at least one personally authored prompt set or output. May have encountered failure cases and described them generally.

Anchor L5

Uses multiple tools across multiple task domains. Can produce a specific recent output stream — what was created in the past 7 days. Has refined at least one workflow through iteration; can describe what was kept, what was discarded, and why. Multiple uploaded artefacts across taxonomy types. Failure cases are described specifically. When asked about a specific deliverable they have produced in the past 7 days, identifies the AI components used without prompting and can describe the boundary they manage between AI-supported and AI-independent work.

Calibrated trust and probabilistic reasoning — scored 2 of 5 (Critical evaluation of AI outputs; reasoning about confidence intervals; systematic verification.) Treats AI outputs as authoritative or as untrustworthy in either direction — does not show calibrated discrimination. Two L1-consistent signals required to fire the anchor: (a) when asked about a recent AI use, describes the output without engaging with verification or accuracy considerations, AND (b) when probed directly on verification practices, defers to others ("the team checks that") or describes generic verification posture without a specific example. Conflates fluent prose with factual accuracy where evidence is available. May have acted on hallucinated information without recognising it as such. At L2, some L3 evidence is present but not the full L3 pattern.

Anchor L1

Treats AI outputs as authoritative or as untrustworthy in either direction — does not show calibrated discrimination. Two L1-consistent signals required to fire the anchor: (a) when asked about a recent AI use, describes the output without engaging with verification or accuracy considerations, AND (b) when probed directly on verification practices, defers to others ("the team checks that") or describes generic verification posture without a specific example. Conflates fluent prose with factual accuracy where evidence is available. May have acted on hallucinated information without recognising it as such.

Anchor L3

Has a stated verification habit ("I always check citations" or "I read the output critically") and can point to one example where verification caught a problem. Differentiates verification effort by task — at least two task categories named with different verification practices, with the differentiation tied to the cost of being wrong rather than the visibility of the task. Can describe a hallucination they have encountered and how they recognised it.

Anchor L5

Treats verification as a function of the task, not a blanket rule — articulates which tasks need verification and why. Demonstrates discriminating trust through specific behavioural examples (verifies X but not Y for stated task-specific reasons); reasoning about reliability may be qualitative or quantitative — the criterion is consistent task-by-task differentiation, not the form of articulation. Has built personal heuristics for spotting hallucinations. References specific failure modes from own experience. Can articulate a documented threshold above which they will not act on AI output without verification, and can describe a recent case where the threshold was applied.

Fit judgement — scored 1 of 5 (Pattern recognition for where AI succeeds and where it does not.) Treats AI as universally applicable or universally inappropriate. Cannot distinguish workflows that yield to AI from those that do not. Adopts vendor or peer-organisation framing without evidence of personal evaluation. The "jagged frontier" concept is absent — speaks of AI as a single capability rather than a heterogeneous one.

Anchor L1

Treats AI as universally applicable or universally inappropriate. Cannot distinguish workflows that yield to AI from those that do not. Adopts vendor or peer-organisation framing without evidence of personal evaluation. The "jagged frontier" concept is absent — speaks of AI as a single capability rather than a heterogeneous one.

Anchor L3

Identifies workflows where AI works well and where it does not — but the boundary is generic (e.g. "AI is good for drafting, not for decisions"). Can name two or three specific workflows in their own organisation that have yielded value and two or three that have not. References the experience that informed the discrimination.

Anchor L5

Predicts workflow fit at high resolution and can defend the prediction with reference to specific properties of the workflow (data structure, decision cost, verification cost, repetition). Identifies that AI capability is heterogeneous within a domain rather than uniform across a domain — provides at least two examples within a single workflow type where AI performs strongly and at least two where it performs weakly, and can articulate the difference in task-property terms. May or may not use specific framework labels (e.g. "jagged frontier") — the construct is the discrimination, not the label. Has changed their fit judgement on a specific workflow as the technology changed. Identifies non-obvious use cases that competitors have not yet identified.

Agentic supervision and orchestration — scored 1 of 5 (Directs, bounds, and supervises AI agents (not just prompts models); sets autonomy/permission boundaries; recognises when an agent exceeds its competence and intervenes.) Treats AI exclusively as a single-turn assistant (chat, drafting, lookup). Has not configured an agent or multi-step automation. Cannot describe what an autonomy boundary is or why one would be set. When asked what they would let an AI system do unsupervised, answers in generalities or defers entirely to a vendor/IT function.

Anchor L1

Treats AI exclusively as a single-turn assistant (chat, drafting, lookup). Has not configured an agent or multi-step automation. Cannot describe what an autonomy boundary is or why one would be set. When asked what they would let an AI system do unsupervised, answers in generalities or defers entirely to a vendor/IT function.

Anchor L3

Has personally configured at least one multi-step agent or automation (e.g. a tool-using assistant, a scripted workflow, a custom GPT with actions). Can describe the guardrails set and at least one failure that was caught. Supervision is real but ad hoc rather than systematic. Can articulate, at least informally, where they were willing to grant autonomy and where they were not.

Anchor L5

Routinely orchestrates and supervises agents as part of how work gets done. Articulates a personal model for autonomy tiers and escalation (what an agent may do unsupervised, what requires approval, what is never delegated). Can describe both a guardrail that worked and one that failed, with the change that followed. Names a specific authority granted to an agent unsupervised AND a specific boundary refused, with the reasoning that distinguishes the two (cost of error, reversibility, verification cost). Distinguishes tasks suited to autonomous execution from those requiring a human in the loop, in task-property terms.

Workflow redesign and change architecture — scored 2 of 5 (Redesigns processes for AI-augmented operation; applies defensible change methodology.) Has not redesigned workflows; describes change as "rolled out tools" or "trained the team". When asked about methodology, defaults to generic project-management language or names a methodology in passing without showing application. No personal theory of change. Describes redesigns where the design was someone else's. At L2, some L3 evidence is present but not the full L3 pattern.

Anchor L1

Has not redesigned workflows; describes change as "rolled out tools" or "trained the team". When asked about methodology, defaults to generic project-management language or names a methodology in passing without showing application. No personal theory of change. Describes redesigns where the design was someone else's.

Anchor L3

Has redesigned at least one workflow end-to-end (not just inserted AI into an unchanged process). Demonstrates a methodology in practice — either by naming a published framework (Kotter, ADKAR, Lean Change, or equivalent) AND showing how it shaped a specific decision, OR by articulating their personal change methodology with sufficient structure that the methodology is reconstructable from their description. The construct is methodological discipline, not methodology-naming. Outcome described qualitatively (faster, better, fewer errors).

Anchor L5

Has redesigned multiple workflows and articulates a personal theory of change derived from those redesigns. Methodology is used as a working tool, not a label — references which steps were emphasised or skipped and why. Outcomes are evidenced — either quantified (cycle time, error rate, capacity) or qualitatively documented through specific stakeholder feedback, downstream-decision behavioural change, or other observable signals appropriate to the workflow type. Has described at least one redesign that did not go as planned and the specific revision that followed. Demonstrates that the redesign extended past the AI insertion to the surrounding workflow (handoffs, exception cases, performance metrics).

Empathic mobilisation — scored 1 of 5 (Reads organisational political/emotional state; mobilises through change; builds commitment over compliance.) Treats stakeholder resistance as a compliance problem. Describes overcoming resistance with authority, escalation, or attrition (waiting out the resistor). Cannot describe the resistor's reasoning in their own terms. Does not distinguish commitment from compliance. Communication artefacts (where described) are top-down broadcasts.

Anchor L1

Treats stakeholder resistance as a compliance problem. Describes overcoming resistance with authority, escalation, or attrition (waiting out the resistor). Cannot describe the resistor's reasoning in their own terms. Does not distinguish commitment from compliance. Communication artefacts (where described) are top-down broadcasts.

Anchor L3

Can describe a specific resisting stakeholder and what the resistance was about. Has worked directly with the stakeholder (individual coaching, multiple conversations) rather than only through committee or memo. Distinguishes resistance from disagreement and engages disagreement through structured discussion (named decision processes, surfaced reasoning, willingness to revise position when persuaded) rather than escalation or avoidance. Communication artefacts attempt to engage the audience's frame rather than override it.

Anchor L5

Reads the political and emotional state of the organisation accurately — describes stakeholder positions in stakeholders' own terms before describing the disagreement. Has coached a specific stakeholder through an identity transition (e.g. from "expert technologist" to "learner alongside AI"). Distinguishes commitment from compliance behaviourally — names what each looks like in observable terms. Has created cultural conditions where the change continued without the candidate's ongoing intervention. Communication artefacts close perception gaps (e.g. address the BCG/Columbia 51-point pattern).

Governance, safety and regulatory architecture — scored 2 of 5 (Designs decision rights, risk frameworks, audit structures, regulatory posture, AND the operational safety control plane (permission boundaries, kill switches, silent-degradation monitoring, named-owner accountability).) Governance described purely as compliance-receiving rather than governance-shaping — when probed on the candidate's decision-making contribution to the governance structure, defaults to "we follow the regulator's guidance" or "our legal team handles that" without showing the candidate's own governance judgement. Regulatory posture is naive (assumes regulators will handle it) or defensive (treats regulation as obstacle). No personal involvement in decision-right design. Note: in heavily-regulated industries, appropriate deferral to specialist functions is itself a T3 construct — knowing what to own and what to defer; L1 fires only when the candidate cannot describe their own judgement on that boundary. At L2, some L3 evidence is present but not the full L3 pattern.

Anchor L1

Governance described purely as compliance-receiving rather than governance-shaping — when probed on the candidate's decision-making contribution to the governance structure, defaults to "we follow the regulator's guidance" or "our legal team handles that" without showing the candidate's own governance judgement. Regulatory posture is naive (assumes regulators will handle it) or defensive (treats regulation as obstacle). No personal involvement in decision-right design. Note: in heavily-regulated industries, appropriate deferral to specialist functions is itself a T3 construct — knowing what to own and what to defer; L1 fires only when the candidate cannot describe their own judgement on that boundary.

Anchor L3

Has designed at least one governance structure for an AI deployment — decision rights, escalation paths, risk thresholds. Can articulate how the structure adapts to changing model capability. Regulatory posture is informed: knows what applies in their jurisdiction and what is likely coming. Has navigated at least one regulatory question for an AI deployment.

Anchor L5

Designs governance as a working architecture, not a control overlay — decision rights are matched to the risk profile of each decision class. Has built audit structures and can articulate the design trade-off they made between regulator-required visibility and operating-team workload, with reference to specific decisions about scope, frequency, or sampling that managed that trade-off. Regulatory posture is strategic: anticipates future regulation and positions the organisation to influence and absorb it. References specific cases where governance design changed the trajectory of an AI initiative for the better. Safety control plane (v0.2): has designed permission/authority boundaries for an autonomous or semi-autonomous system, an escalation or kill-switch path, and monitoring for silent degradation, and can describe a production case where a guardrail caught (or would catch) an AI system failing quietly. Ties AI decisions to named human owners.

Systems integration — scored 1 of 5 (Anticipates cross-functional ripple effects; recognises emergent patterns; Cynefin-grounded sense-making.) Treats AI deployment as a local change with local effects. Has not anticipated cross-functional ripple effects; surprises documented as "no one saw that coming". When asked about second-order effects, defaults to risk-register language. No reference to complex-adaptive-systems thinking, even informally.

Anchor L1

Treats AI deployment as a local change with local effects. Has not anticipated cross-functional ripple effects; surprises documented as "no one saw that coming". When asked about second-order effects, defaults to risk-register language. No reference to complex-adaptive-systems thinking, even informally.

Anchor L3

Can describe a specific case where a cross-functional ripple effect emerged from an AI deployment — and can describe what was learned about anticipating it. Sequences AI deployments rather than parallelising all of them. Recognises that some decisions are not amenable to upfront planning and uses pilot/iterate approaches for those.

Anchor L5

Articulates a working understanding that some decision contexts have non-linear feedback loops and irreducible uncertainty about second-order effects, and applies different sense-making practices to those contexts — may reference Cynefin, complexity theory, or systems thinking explicitly, or may articulate the same insight in their own operational terms. The construct is the working understanding, not the framework labels. Distinguishes ordered from un-ordered decision contexts and chooses different practices for each. Has designed AI deployments with feedback loops that capture emergent effects. Can describe a case where they failed to anticipate a ripple effect and the specific revision to their approach that followed. Sequences and modifies deployments because they anticipate downstream effects in others' workflows.

Operational execution and embedding — scored 2 of 5 (Pilot-to-production discipline; AI-calibrated training infrastructure; metrics for non-deterministic workflows; recursive/self-improving infrastructure.) Has not led an AI initiative from pilot to operational embedding. Pilots are described as ending at the demonstration of feasibility rather than at production embedding. Training infrastructure is generic. Performance measurement is deterministic-system thinking (uptime, defect rate). No iteration on production data. At L2, some L3 evidence is present but not the full L3 pattern.

Anchor L1

Has not led an AI initiative from pilot to operational embedding. Pilots are described as ending at the demonstration of feasibility rather than at production embedding. Training infrastructure is generic. Performance measurement is deterministic-system thinking (uptime, defect rate). No iteration on production data.

Anchor L3

Has led at least one pilot to production embedding. Built training infrastructure that addresses AI-specific features (capability evolution, probabilistic outputs, user-as-tester). Performance measurement is non-deterministic-aware (sampling, drift monitoring, qualitative evidence triangulation). Iterates on production data; can describe at least one production-data-driven refinement.

Anchor L5

Distinguishes pilot from production as a discipline, not a phase — articulates a specific threshold or condition for the transition. Has built training infrastructure that evolves with the model (refresh cadence matched to model release cadence). Performance metrics are designed for the non-deterministic case (distributional measures, drift alerts, calibration tracking). Embedding evidence: can name the moment when AI usage became default behaviour for the team rather than special-occasion. Has iterated through multiple production cycles. Recursive-improvement element (v0.2, L4→L5 discriminator): describes a specific evaluation/feedback loop or monitoring harness they built that compounds — performance improves across cycles without a human re-tuning it each time — and can name the metric it moved. A genuinely self-improving loop, not a repeated manual optimisation, is what separates L5 from L4 on T5.

Strategic patience and value reasoning under uncertainty — scored 3 of 5 (Defends investment through the J-curve; constructs investment cases when value is genuinely uncertain.) Has defended at least one transformation investment through pressure to walk it back, and can describe what the pressure was and how the defence was reasoned. Demonstrates that value uncertainty was actively managed in the defence — describes specifically what was known and unknown at the time of decision, what was decided in advance vs. left to be decided based on emerging evidence, and what the off-ramps were. Distinguishes capability investment from value-capture investment and can articulate why the timing differs.

Anchor L1

Defends investments by emphasising sunk cost or authority, not reasoning. Cannot describe a J-curve experience or names the J-curve concept without showing application. Investment cases described in single-point ROI terms; no reference to options thinking, portfolio framing, or expected-value reasoning under uncertainty. Folds under pressure to walk back commitments.

Anchor L3

Has defended at least one transformation investment through pressure to walk it back, and can describe what the pressure was and how the defence was reasoned. Demonstrates that value uncertainty was actively managed in the defence — describes specifically what was known and unknown at the time of decision, what was decided in advance vs. left to be decided based on emerging evidence, and what the off-ramps were. Distinguishes capability investment from value-capture investment and can articulate why the timing differs.

Anchor L5

Demonstrates reasoning about investment value as a probability distribution rather than a point estimate — describes managing the value case by what could go right and what could go wrong with rough likelihoods attached, decisions designed to preserve downside while keeping upside open, and treating the investment as one bet within a wider portfolio. May reference real options, portfolio-of-bets, or expected value under uncertainty as working tools — or may articulate the same reasoning in the candidate's own operational language. The construct is the reasoning pattern, not the framework labels. Has defended an investment through a long horizon where value was not yet visible. Has also walked away from an investment when evidence justified it, and can articulate what shifted. Manages the J-curve as a deliberate strategy.

Tolerance for ambiguity — scored 2 of 5 (Cognitive comfort with uncertainty and incomplete information.) Resolves ambiguity prematurely — defaults to false certainty (acts as if uncertainty did not exist) or paralysis (defers decision until information is complete). Decision quality drops sharply in incomplete-information contexts. References "we need more data" as a substitute for a decision. The E2 criterion is cognitive — fires on the candidate's reasoning pattern under incomplete information, independent of any S1 (emotional regulation) signal in the same response. At L2, some L3 evidence is present but not the full L3 pattern.

Anchor L1

Resolves ambiguity prematurely — defaults to false certainty (acts as if uncertainty did not exist) or paralysis (defers decision until information is complete). Decision quality drops sharply in incomplete-information contexts. References "we need more data" as a substitute for a decision. The E2 criterion is cognitive — fires on the candidate's reasoning pattern under incomplete information, independent of any S1 (emotional regulation) signal in the same response.

Anchor L3

Operates competently when information is incomplete. Distinguishes between ambiguity that can be reduced through investigation and ambiguity that is structural. Makes provisional decisions and revises them as information arrives. Acknowledges uncertainty explicitly rather than hiding it.

Anchor L5

Provides specific examples of operating in deeply ambiguous contexts (e.g. M&A diligence under information asymmetry, regulatory uncertainty, technology choice in immature category) and can describe the working practices that enabled clear thinking in those contexts — typically: separating their preferences from their conclusions, naming which working assumptions were most fragile, making provisional decisions with revision triggers attached. Can make and communicate consequential decisions with explicit uncertainty bounds, in a way that maintains organisational confidence rather than undermining it.

Learning velocity — scored 2 of 5 (Speed of belief updating as new evidence and capabilities emerge.) Cannot identify a recently-updated strongly-held belief. References "always learning" as a posture rather than naming specific updates. Updates beliefs only when evidence is unambiguous and overwhelming. New AI capability is responded to with resistance or skepticism by default. At L2, some L3 evidence is present but not the full L3 pattern.

Anchor L1

Cannot identify a recently-updated strongly-held belief. References "always learning" as a posture rather than naming specific updates. Updates beliefs only when evidence is unambiguous and overwhelming. New AI capability is responded to with resistance or skepticism by default.

Anchor L3

Can identify at least one recent strongly-held belief that has been updated, and name the trigger. Has a reading or engagement practice that actively exposes them to new evidence. References at least one moment in the past year where their understanding shifted materially.

Anchor L5

Actively seeks disconfirming evidence on positions they hold. Can identify multiple recently-updated beliefs (≥ 2 in past year on consequential topics), with at least one update that was politically expensive to make publicly — and can describe how the public update was managed. The criterion is evidence of ongoing belief-updating practice, not just one notable example. Distinguishes which of their beliefs are load-bearing for the organisation versus which are personal — and updates the load-bearing ones publicly so the organisation can update with them.

Emotional regulation — scored 3 of 5 (Steadiness under pressure; absence of volatility patterns that destroy psychological safety.) Maintains composure under typical executive pressure. Recovers from setbacks within a reasonable timeframe. Does not visibly transfer pressure to subordinates beyond what the situation requires. Has navigated at least one high-pressure context where steady leadership was visible to the team.

Anchor L1

Shows visible volatility under pressure — sharp tone shifts, withdrawal, public frustration. Stakeholders describe walking on eggshells. Has destabilised a meeting or team through emotional response. Recovery from pressure is slow. May be effective in calm conditions but unreliable under stress.

Anchor L3

Maintains composure under typical executive pressure. Recovers from setbacks within a reasonable timeframe. Does not visibly transfer pressure to subordinates beyond what the situation requires. Has navigated at least one high-pressure context where steady leadership was visible to the team.

Anchor L5

Calm under acute pressure in ways that have observably steadied others. Has been visibly steady in a context where the team expected volatility — describes the specific situation, the pressure source, the candidate's own internal state during it, and the behavioural signal the team would have observed. Where 360 input is available (Part 2+) the team's corroborating account strengthens the score; in Part 1 the candidate's specific behavioural narrative is the primary criterion. Models emotional regulation deliberately. Recovers quickly from setbacks and the recovery is visible (not hidden).

Status security — scored 3 of 5 (Capacity to be visibly inexpert; learns from juniors in public; shares own mistakes without face-saving.) Can describe a specific recent occasion when they publicly admitted not knowing something or being wrong. Learns from juniors on specific topics where the juniors have more recent expertise. Does not require public credit for ideas that came from elsewhere.

Anchor L1

Avoids being visibly inexpert. Will not admit not-knowing in public or in front of juniors. Defends status through assertion or by changing the subject when challenged on competence. Has not been observably wrong in public in a meaningful way. May rely on positional authority to override input from juniors who know more on a specific topic.

Anchor L3

Can describe a specific recent occasion when they publicly admitted not knowing something or being wrong. Learns from juniors on specific topics where the juniors have more recent expertise. Does not require public credit for ideas that came from elsewhere.

Anchor L5

Status-secure as a default. Routinely learns from juniors in public in ways the team can describe. Frames "I don't know" or "I was wrong" as standard professional behaviour rather than as a vulnerable moment. Has built team norms around not-knowing — describes specific practices (e.g. starting meetings with "what's unclear?", or routine "I was wrong about..." entries in shared documents). Status security is not performance; it is integrated.

Decisive action under uncertainty — scored 3 of 5 (Acts before complete information is available; distinguishes prudent governance from disguised indecision.) Makes consequential decisions on incomplete information when the situation requires it. Can describe a decision taken under uncertainty and what the reasoning was. Distinguishes between "let me investigate further" as a legitimate decision and as a deferral.

Anchor L1

Delays decisions indefinitely under uncertainty. Mistakes process for progress — calls for additional analysis or another committee review when a decision is needed. Cannot distinguish prudent governance from disguised indecision. Indecision affects subordinates' ability to act.

Anchor L3

Makes consequential decisions on incomplete information when the situation requires it. Can describe a decision taken under uncertainty and what the reasoning was. Distinguishes between "let me investigate further" as a legitimate decision and as a deferral.

Anchor L5

Acts decisively under uncertainty with clear acknowledgement of what was not known at the time of the decision. Distinguishes prudent governance from disguised indecision and can name examples of each — has accepted a higher-than-comfortable error rate in exchange for organisational tempo when the situation warranted. Has reversed a decision when evidence justified — describes the reversal in terms that take responsibility for the original call. Some face-management framing is acceptable; the criterion is the underlying ownership of the reversal. Decisive action is calibrated to consequence — slow when cost-of-error is high and reversibility low; fast when cost-of-error is recoverable.