The Use Cases Everyone Tries and Nobody Ships

Certain agent applications get attempted repeatedly across organizations and stall at the same point. They're worth cataloguing, because the pattern in why they stall is more instructive than the individual failures — and in several cases there's a narrower version that does ship.

The autonomous customer-facing agent with real authority

The attempt: an agent that resolves customer issues end to end, including refunds, account changes, and commitments.

Where it stalls: the blast radius. A wrong action affects a customer, is often irreversible, and creates liability. The gate that gets added to manage that turns it into an assisted tool, at which point the autonomy that justified the project is gone.

The version that ships: the agent gathers, drafts, and proposes; a person approves in one click with the evidence in front of them.

The "ask anything about our company" assistant

The attempt: an agent over all internal documents answering any employee question.

Where it stalls: the corpus. Internal documentation is contradictory, stale, and incomplete, and the agent surfaces that faithfully. Answers are confidently wrong because the source was wrong, and trust collapses after a handful of incidents.

The version that ships: scoped to a curated, maintained corpus for a specific domain, with sources cited and freshness visible. ⚠️ The unglamorous finding is that the project is a documentation project wearing an AI costume, and the teams that succeed are the ones that accept that.

The autonomous data analyst

The attempt: ask a business question in English, get an analysis.

Where it stalls: semantics. The agent writes syntactically valid queries encoding wrong business logic — joins that double-count, filters that miss the exclusion everyone knows about, metrics that don't match the finance definition. The answers are plausible and wrong, and nobody catches it because checking requires the expertise the tool was meant to replace.

The version that ships: curated views with the business rules baked in, or parameterized queries for known questions, with the semantics documented for the agent.

The meeting-to-action-items agent

The attempt: transcribe, extract commitments, create tasks.

Where it stalls: commitment is ambiguous in real conversation. "We should probably look at that" isn't a task, and neither is a suggestion someone made and the group ignored. The output is a list mixing real commitments with noise, and cleaning it takes as long as writing it.

The version that ships: a summary a human turns into tasks, or extraction only from explicitly-marked commitments.

The autonomous incident responder

The attempt: detect, diagnose, remediate.

Where it stalls: an incident is when the system's behavior is least understood, which is the worst condition for autonomous action. Also the diagnosis problem — something always deployed recently and it always looks causal.

The version that ships: gather, correlate, and prepare actions; a human executes.

The "replace the whole workflow" agent

The attempt: automate a multi-step business process end to end.

Where it stalls: compounding. Six steps at high per-step reliability produce disappointing per-task reliability, and errors early are executed faithfully by everything downstream.

The version that ships: automate the two or three steps with complete inputs and mechanical checks, leave the judgment steps to people, and accept that the process still has humans in it.

💡 The pattern

Every stall is one of four things:

• Blast radius exceeds what anyone will authorize.
• The corpus or data isn't good enough, and the agent surfaces that faithfully.
• The semantics live in people's heads and aren't in any artifact.
• Compounding across steps that each seemed fine.

✅ And every shipping version does the same thing: narrow the scope, add a check, keep the human at the decision. Which is less exciting than the original pitch and is the version that produces value.

The takeaway

The repeatedly-attempted applications fail for four reasons — blast radius, corpus quality, unwritten semantics, and compounding — and each has a narrower version that works. Before starting one of these, check which of the four applies to yours and whether you're willing to build the shipping version instead of the pitch. Frequently the real project turns out to be documentation, or data semantics, or a permission model, wearing an agent costume.