HAL · Capability sustainability
An accountable AI operating model cannot depend indefinitely on human judgement while simultaneously removing the experiences through which that judgement develops. This is a system-design question, not a training programme.
The question is not only whether a human remains in the loop. It is whether the system continues to produce humans capable of meaningfully being there.
This is not a fourth Loop pattern and not a ninth HAL domain. It is a cross-cutting condition of meaningful human involvement. For Human-in-the-Loop, Human-on-the-Loop and Human-Accountable-for-the-Loop alike, the operating model assumes a capable human. The method below asks whether that assumption will still be true after the work is redesigned.
It is also not an argument against automation, and not an argument for preserving repetitive junior work because previous generations had to do it. Traditional professional work often combined production and development in the same activity. AI allows those functions to separate. Organisations therefore need to identify which developmental effects remain important and deliberately reproduce, relocate or redesign them.
Two outputs of professional work
For years, professional processes produced both work and workers capable of doing more difficult work later. We measured the first because the second happened gradually in the background. AI allows us to separate them.
Output A
The visible organisational output.
Output B
The less visible output created through participation.
Capability debt is the future human capability placed at risk when developmental experience is removed from a workflow without an alternative mechanism for producing or maintaining the capability it supported. The analogy to technical debt is useful only if it forces a question that ordinary productivity metrics miss. Automation does not necessarily create a debt; some repetitive work adds little, and well-designed AI use can accelerate feedback. The point is that organisations currently have poor visibility of whether it is happening.
The assumption underneath every Loop pattern
Will the organisation still be producing people capable of performing these roles five or ten years from now?
The method
Map the workflow twice: once for what it produces, and once for what participation develops. Then decide deliberately what happens to the capability, test whether the replacement works, and keep watching it. These are not levels and they do not need an acronym.
Map the work
Map the hidden learning
Preserve / Compress / Simulate / Move
Test the capability
Monitor capability over time
The useful question is where the learning happens in the new system, not how to recreate the old job. A single role may need several of these responses.
Keep direct experience where doing the real task is important to developing the required judgement. Preserve the mechanism, not automatically the complete old workflow.
Example. A junior lawyer makes an independent assessment of selected matters before seeing an AI-generated analysis.
Remove empty volume while retaining selected examples that contain meaningful variation, ambiguity and exceptions.
Example. Rather than reviewing 300 largely similar contracts, someone reviews 30 selected examples representing important patterns and exceptions.
Create encounters with failures or rare conditions that are too infrequent, risky or unpredictable to leave to chance.
Example. Exercises using plausible but incorrect AI output, conflicting instructions, incomplete information, or technically correct but commercially inappropriate advice.
Develop the required judgement elsewhere in the redesigned workflow rather than recreating the old job.
Example. Junior professionals investigate AI–human disagreement, handle exceptions, trace provenance, analyse failures, or observe senior decision-making.
Use this as a workshop artefact. Complete it for each human role the redesigned workflow changes. Download the worksheet if you want a copy you can fill in.
| Diagnostic | Ask |
|---|---|
| Primary output | What does this work visibly produce? |
| Capability output | What does somebody become better at by repeatedly doing it? |
| Mechanism | What creates that improvement: repetition, feedback, consequence, proximity, variation, or something else? |
| Necessity | Does the redesigned system still need that human capability? |
| Replacement | If yes, will the experience be preserved, compressed, simulated or moved? |
| Evidence | How will the organisation know the resulting person is genuinely capable? |
| Sustainability | What will keep the capability credible as automation increases? |
Do not allow the method to end at “provide training”. If the organisation believes it has replaced the developmental effects of traditional work, it should test whether the intended capability actually exists. Independent manual performance is not an absolute requirement in every case. The deeper requirement is sufficient independent capability to exercise the assigned oversight or accountability role.
Test with unfamiliar scenarios, plausible but incorrect outputs, incomplete information, conflicting sources, ambiguous cases, novel exceptions and realistic time pressure. The objective is not only whether they spotted a planted error. Did they know why it mattered? Did they recognise uncertainty? Did they know what to do next? Could they challenge an apparently authoritative machine output?
Record an evidence state without adding a ninth HAL domain or calculating another score. Do not average it with the eight-domain control profile. Where Review, Monitor or Own relies on human judgement, an unsupported assumption weakens Ownership, Evidence, Monitoring and Review even if policies and system controls are present.
| State | Evidence | What it means |
|---|---|---|
| Assumed | Capability is inferred from title, qualification or historic experience. | The control case has not been tested. Treat this as a material assumption, not evidence. |
| Defined | The judgement required for meaningful oversight is explicit. | The organisation knows what the role must be able to do, but not whether anyone can still do it. |
| Designed | Necessary developmental experience has been preserved, compressed, simulated or moved. | The replacement is a hypothesis. It remains unproven until people demonstrate the judgement. |
| Tested | People have demonstrated the capability against realistic uncertainty and AI error. | Current role-holders can exercise the assigned judgement. The pipeline and durability are still open questions. |
| Sustained | Performance, continued exposure and the future pipeline are monitored and acted upon over time. | Capability is treated as an operating-model condition, not a one-off training event. |
Competence asks whether the person can make the judgement required. Capacity asks whether they can give it enough time and attention under actual workload. A reviewer can be highly competent but overloaded, available but incapable, both or neither. Automation may increase capacity by removing volume while weakening competence if people no longer practise the parts of the work on which oversight depends. Neither effect should be assumed; both should be tested.
Can this person exercise the required judgement today?
Does the role provide enough continued exposure, feedback and practice for that judgement to remain credible?
Does the operating model continue to produce people able to perform the role later?
A reviewer who starts from an AI answer is not in the same cognitive position as someone producing an independent answer. Finding the consequential 2% in a draft that is 98% correct may require more judgement than producing the original first draft, not less.
Where does the capable reviewer eventually come from?
These do not automatically prove capability loss. A low disagreement rate may show that the system improved. Each is a reason to investigate, compare with outcomes, and test the human role directly.
Keep returning to the operating-model question: if your AI workflow relies on human judgement, where does that judgement come from and what keeps it credible?
