The Applied Knowledge Gap

Abstract

Evaluations of deployed AI agents score whether the task got done, typically by checking the final state of a system. We present an audit framework that measures something state-based evaluation is structurally blind to: whether an agent volunteers information it has available, recognizes as applicable, and yet was never explicitly asked for. The framework separates four stages of an agent's knowledge use, available, retrieved, applied, surfaced, and isolates failures at the last stage. Across three production documentation agents we find no gap: factual questions, including edge-case eligibility restrictions, are answered correctly and the relevant caveats are surfaced unprompted. The pattern changes in judgment domains. On controlled support agents over realistic business policies, agents recognize an applicable policy and answer correctly when asked directly, yet fail to volunteer it in production-style phrasing, with surfacing falling as low as 50% where clean-eval performance is 100%. The failures concentrate where volunteering the policy conflicts with simply fulfilling the user's stated request. The contribution is methodological: a framework for measuring the distinction between knowing a policy and volunteering it.

1. Four stages

Output-based agent evaluation collapses a pipeline. We separate it:

AVAILABLE  ->  RETRIEVED  ->  APPLIED  ->  SURFACED
(in corpus)    (in context)  (recognized   (volunteered
                              as relevant)   to the user)

A failure can occur at any stage, and they have different fixes. The stage output evaluations are blind to is the last one: an agent can have the information available, retrieve it, recognize it as relevant, and still not volunteer it, because surfacing requires recognizing that an unrequested policy applies and offering it without being asked. We refer to this final-stage failure as the applied knowledge gap.

2. The scope, stated up front

The central result has two halves, and the negative half is load-bearing.

1. Factual-lookup agents do not have this gap. Three production documentation agents passed cleanly, including questions designed to elicit failures.
2. Judgment-domain agents do. Controlled support agents over realistic policies fail to volunteer applicable policies at measurable rates.

So the claim is not "AI agents fail to disclose what they know." It is scoped: the gap appears where disclosure requires inferring that information applies and volunteering it unprompted, not where it requires answering a direct question. The documentation-agents-pass result is evidence that the instrument discriminates rather than flagging everything. We do not claim factual retrieval is solved field-wide; we report only that the agents we tested showed no gap on the questions we posed.

3. Real agents, factual domain: no gap

We queried three production documentation assistants (Supabase, Resend, Pinecone) with factual questions including deliberate edge cases: free-tier region restrictions, plan-eligibility caveats, domain-verification requirements. All three answered correctly and surfaced the relevant caveats, with citations. Asked "can I use point-in-time recovery on the free plan?", one agent correctly volunteered the plan restriction and the upgrade path. There is no available-but-not-surfaced gap here. Retrieval is what these systems are built for, and they do it well.

4. Controlled judgment agents: the gap appears

We built support agents (embedding-retrieval over realistic, overlapping policies) and probed them with scenarios where a policy should be volunteered but the customer never names it. A customer says cost is the problem and asks to cancel; policy says to proactively offer the retention discount. Disclosure was judged from the natural reply; recognition checked by forced-choice; clean-eval by the direct policy question.

The counts are small and we report them as counts, not rates: these are a controlled demonstration that the gap occurs and where it concentrates, not an estimate of how often it occurs in production. The agents answer the policy correctly when asked directly but fail to volunteer it in the live scenario. A representative miss: the customer cites cost and asks to cancel; the agent helpfully offers to pause the subscription, but does not offer the retention discount that policy says to surface on a cost complaint. The agent sounds helpful. It dropped the higher-value obligation.

5. The mechanism: knowing versus volunteering

Across sectors the misses concentrate on one obligation type: retention and win-back offers. This is not random. The observed failures are not simple retrieval failures. In most cases the agent has access to the relevant policy and can identify it when queried directly. The failure occurs at the point of disclosure. The behavior is consistent with a compliance-default strategy: fulfilling the user's stated request rather than volunteering an applicable policy that was not explicitly requested. Retention is the obligation where this strategy is most costly, and the one tied most directly to revenue. The other obligations the agents handled (proration, recall notices, restocking waivers, price match, autopay discounts, downgrade warnings) do not pit fulfilling the request against volunteering an unrequested policy in the same way, and the agents surfaced them.

The distinction is not whether the agent knows the policy. The distinction is whether the agent volunteers the policy when it becomes applicable. Modern agents retrieve and answer well; the harder problem is recognizing what should be said when nobody asked.

We deliberately do not commit to a single mechanism. The compliance-default reading, fulfilling the literal request rather than volunteering more, is consistent with the sycophancy literature, but it is not the only candidate. The same failures could come from brevity or conciseness training, from genuine uncertainty about whether the agent has permission to volunteer an offer, or from ambiguity in the policy itself about when proactive disclosure is required. What this work establishes is the gap, measured; identifying which of these mechanisms drives it in a given system is a separate question, and one the audit is designed to help answer rather than presume.

6. Why output evaluations miss this

A clean evaluation asks "what is your retention policy?" and the agent answers correctly (clean-eval 80-100%). Production phrasing ("money's tight, cancel me") never names the policy, and surfacing drops as low as 50%. The gap between these two is the difference between an evaluation that tests retrieval and a deployment that requires volunteering. An agent can pass every direct policy question and still fail eligible users in production.

7. The four metrics

The available-retrieved-applied-surfaced decomposition yields four measurable quantities:

• Retrieval accuracy: does the agent answer correctly when asked directly?
• Applicability accuracy: does it recognize when an unrequested policy applies?
• Volunteering rate: does it proactively surface the applicable policy?
• Disclosure gap: the difference between what it recognizes and what it volunteers.

The insight these operationalize: AI is now very good at answering questions; the harder, less-measured problem is recognizing what should have been said when nobody asked.