The Intent Gap: A Taxonomy of Real-User Failure Modes in Frontier AI Agents

AbstractPublic benchmarks for frontier AI agents (GAIA, OSWorld, AgentBench, GDPval,BrowseComp, SWE-bench) are uniformly researcher-imagined: tasks aredesigned by AI researchers, in clean declarative form, with ground-truth answersfixed in advance. Real users — as captured in 1M+ public conversationcorpora — do not interact with frontier models this way. They assume context,change their minds mid-task, contradict earlier requests, abandon, andphrase requests in registers no benchmark has ever sampled. We hypothesizethat benchmarks systematically miss the deployment-relevant failure mode wecall the intent gap: cases where the model literally answered the prompt butmissed what the user actually wanted.We mine WildChat-1M and LMSYS-Chat-1M using a three-stage frustrationsignalfilter (regex → embedding → LLM-as-judge) to surface a target of ~150–300 high-quality intent-gap conversations. We replay each prompt across fourcurrent frontier models (Claude, GPT, Gemini, and one open baseline) andgrade each replay on two axes: literal compliance and intent satisfaction. Wethen triangulate the resulting taxonomy against ~25 published failure-mode taxonomiesfrom research institutes worldwide and against 5 documented productionfailure incidents (lawsuits, regulatory filings, public post-mortems).Headline preliminary findings (from a 50,000-conversation v2 pilot,before cross-model replay or full hand-review):1. 52.8% of English real-user conversations end after a single exchange.A substantial fraction of these are likely silent abandonmentafter intent-gap failure — invisible to deployment-quality pipelines builton user-feedback signals.2. Only 3.85% of multi-turn conversations contain an explicit repairsignal, even under a permissive 45-phrase filter that scans every user turn.The intent gap exists, but users rarely verbalize it.3. The phrase-list expansion from v1 (17 academic phrases) tov2 (45 natural-language phrases) increased hit rate by ~31×.This implies the dialogue-breakdown detection literature undercounts realworldrepair by more than an order of magnitude because its phrase listsare researcher-imagined.4. Among successfully-judged v2 candidates, 79% (15/19) wereconfirmed as real intent-gap failures by an LLM judge. When the regex fires and the judge can run, the flagged candidate is overwhelminglya real failure.5. The judge model refused to grade 62% (31/50) of real-user candidates,most plausibly on content-moderation grounds. This is itself adeployment-relevant finding: safety-filtered judge models systematicallymiss failures in the most permissive parts of the user-prompt distribution.A hand-reviewed and LLM-judge-filtered corpus of ~150–300 high-quality intentgapconversations is released alongside the filter pipeline, the grading rubric, andthe triangulation matrix mapping our taxonomy onto ~25 prior published failuretaxonomies plus 5 documented production-failure incidents.