The Proxy Learner Problem

A completion record can be entirely accurate and still stop proving the one thing it was built to prove: that a human learned.

Your LMS shows the course as complete. The seat time is logged, the quiz is passed, the certificate is issued. But you cannot tell whether the employee gained the capability or an agentic assistant helped produce the completion signal on their behalf. That gap is the failure mode. The completion event can stay true as a system event while losing its value as proof that a human is now able to do the work. For a CLO or CHRO, that is the moment a familiar metric can quietly stop answering the question it was built to answer.

This is not a cheating panic, and it is not a case for banning assistants. It is a measurement and control question: who performed the learning action, what human capability was demonstrated, and which controls keep activity telemetry from substituting for proof of learning.

The question is no longer whether the module reached done. It is whether the organization can still show the credited person can do the work.

1. A clean completion record is becoming weaker evidence of human learning. Where an agentic assistant can navigate, summarize, answer, or complete parts of a learning task, a completion event can stay accurate as a system event while becoming weaker proof that a human learned. This is a can-and-may risk, not a claim that every record is invalid.
2. The unit of accountability moves from the activity to the actor. The governing question shifts from “did the activity happen” to “who performed it” — separating the human who is credited from whatever performed the work.
3. Completion metrics become a weak control where assistance is plausible. They remain legitimate signals of engagement and flow, but they need corroboration before they can carry a high-stakes claim. Completion does not stop mattering; it stops being self-sufficient proof.
4. The risk concentrates where learning records are reused downstream. Compliance attestations, manager decisions, role-readiness judgments, skill inference, and workforce-transformation metrics can turn a soft activity signal into a load-bearing decision input.
5. The remedy is proportionate validation, not surveillance or a ban. Actor provenance, work-embedded checks, transfer tasks, and assistance-limited demonstrations — scaled to stakes — can preserve completion records while adding proof of capability beside them.
6. The defensible claim is narrow. The evidence supports a can-and-may mechanism, not a measured prevalence. This brief does not assert the pattern is already widespread across corporate training, and most AI-assisted completions may be entirely legitimate.

The record looks clean; the evidence is not.

The LMS was built to record activity inside a controlled workflow: starts, seat time, attempts, completion. Those signals remain useful, and they were trustworthy when only the credited human could generate them. Agentic browsers and learning assistants can change that premise. When an assistant can navigate, summarize, answer, or complete parts of a learning task, a completion event can stay accurate as a system event while becoming weaker as evidence that a human learned.

The shift is from “did the activity happen” to “who performed it.” A clean record is no longer self-evidently proof of human capability; it can become a claim that needs corroboration. This is an evidence problem, not an accusation. The defensible reading is that completion records can become ambiguous where agentic assistance is plausible — not that every record is invalid, and not that this is already widespread across corporate training. The useful posture for leaders is narrower: treat the record as a starting signal rather than the conclusion, at the moment you need it to prove readiness or compliance.

The proxy learner changes the unit of accountability.

The control question moves from the activity to the actor. When an assistant can act as a learner proxy inside an LMS-style workflow — navigating, summarizing, drafting, or answering — activity telemetry alone may no longer confirm that the human engaged, understood, or demonstrated the intended capability. Accountability separates into two strands the LMS historically treated as one: the human who is credited, and whatever performed the work.

For a CLO or CHRO, that reframes the unit of measurement. The thing worth governing is the provenance of the learning action — who initiated it, who answered or demonstrated the assessment, and what human capability is on the record — rather than the completion count itself. This is actor identity in the narrow sense of who did the learning. It is not a non-human identity or access-governance program; those may be supporting context at most. The center of gravity is the integrity of the learning evidence — whether the record can still answer “which human can now do this work.”

Completion metrics become a weak control.

Completion metrics still matter. They remain legitimate operational signals of engagement and flow, and the argument here is not that completion never proves anything. The narrower, defensible claim is that completion metrics can become weak evidence wherever agent assistance is plausible — and that they therefore need corroboration before they can carry a high-stakes claim.

The risk grows when learning records are reused beyond the course itself: compliance attestations, manager decisions, role-readiness judgments, skill inference, and workforce-transformation metrics. Where those downstream uses cannot distinguish human performance from mediated activity, an organization may be measuring throughput where it intends to measure capability. The control weakness is not the metric; it is leaning on the metric as if it were proof. The remedy is not to discard completion data — it is to pair it with evidence that the credited human can perform the work.

A control model: provenance, transfer, and validation.

A practical model can preserve completion records while adding proportionate validation around them. Frame it as risk management scaled to stakes — not surveillance, not accusation, and not a blanket ban on assistants.

1. Actor provenance. For high-stakes learning, capture who initiated the action, who answered or demonstrated the assessment, and what assistance was permitted. The aim is a clear chain of who did the work, not monitoring of the learner.
2. Work-embedded checks. Build evidence into real work — artifacts, work samples, applied tasks — so the proof of capability is something the human produced, not just a module marked complete.
3. Transfer tasks. Use tasks that require applying the capability in a new context, where simply having reached completion does not carry the learner through.
4. Assistance-limited or proctored demonstrations, where stakes justify them. For compliance, safety, certification, or role-critical upskilling, add a stronger proof path — proctored or assistance-limited demonstrations, oral checks, or manager validation — reserved for cases that warrant the cost.

For low-stakes learning, disclosure of allowed assistance plus standard telemetry may be enough; the heavier proof paths should be the exception, applied where a wrong capability judgment is expensive. The design question for each course is simple: what artifact would prove the human can perform the work?

Autonoma will track how enterprises respond as agentic assistants reach inside learning workflows.

1. Completion is reclassified as an activity signal. Watch for learning teams to treat AI-assisted completion as a flag to corroborate rather than a conclusion, especially for records that carry downstream weight.
2. Actor provenance enters learning systems. Watch for LMS and learning-platform owners to capture who initiated a learning action and who demonstrated the assessment, not only that a module was marked complete.
3. Validation tiers scale to stakes. Watch for stakes-based models where completion can stand alone for low-stakes learning while compliance, certification, and role-critical upskilling require a stronger proof path.
4. Work-embedded and transfer evidence gains weight. Watch for assessments built around applied work and new-context tasks, where reaching completion does not by itself carry the learner through.
5. Downstream reuse gets scrutinized. Watch for organizations to identify which learning records feed compliance, mobility, and skill inference, and to require corroboration before those records are reused.

We weighed three counterarguments.

The first: this is not new.

Assessment has always been an imperfect proxy, and completion was never a perfect measure of capability. That is true. What changes is that an assistant can now help generate the completion signal itself, so the gap between activity and capability can widen faster — and in records that used to be safe to trust.

The second: most AI-assisted completion is legitimate.

This is not an accusation that learners are cheating. Most AI-assisted completions may be entirely appropriate. The argument is narrower: completion on its own should no longer close the question for anything carrying real downstream weight.

The third: this is a mechanism, not a measured prevalence.

The evidence supports a can-and-may risk — completion records can become weaker, more ambiguous evidence of human learning where agentic assistance is plausible — not a finding that the pattern is already widespread across corporate training. The right response is to corroborate where stakes are high, not to discard completion data or assume the worst. The organizations that adopt these assistants well will be the ones that keep completion and add a proof of human capability beside it.