Hand-crafted OCR pipelines are already being replaced
Retired shift: a late-September document-extraction launch cycle does not establish a durable transition from deterministic OCR pipelines to conversational LLM extraction.
Disappearing action
Human ritual: Treat product launches for conversational extraction as evidence that manual OCR and schema pipelines have broadly been displaced. Old cost: It turns vendor capability claims into a market-wide before-and-after statement while hiding accuracy, audit, throughput, and regulated-workflow requirements.
Source ledger
Publishable sources attached to this record.
| # | Source | Role | Public status |
|---|---|---|---|
| 1 | llamaindex.aisource | primary receipt | source_urls |
| 2 | llamaindex.aisource | supporting receipt | source_urls |
| 3 | llamaindex.aisource | supporting receipt | source_urls |
Why this shift is retired
Recent document-extraction announcements show an active product area. They do not yet show a durable industry transition from hand-crafted OCR pipelines to LLM-led processing. The available evidence is concentrated in one vendor family and one late-September launch cycle. A leaderboard, a schema-oriented announcement, and a workflow example are useful inputs for future evaluation; they are not independent deployment outcomes.
The proposed before-and-after statement also collapses several different jobs. OCR, layout interpretation, schema design, validation, human exception review, and downstream integration have different accuracy and control requirements. An LLM can improve one part of that chain without replacing the whole deterministic pipeline.
What remains useful
Track bounded capabilities: can a system propose a schema, extract a named field set, show source evidence, route exceptions, and meet a specified error budget? Those questions can become a Delta or project when their verification boundary is explicit. They should not be promoted to a Shift merely because several product pages use the word “agentic.”
Where the old interface wins
Stable, high-volume, regulated, or audit-heavy workflows still benefit from hand-authored schemas and deterministic OCR stages. Their limits are known, their failures can be categorized, and their throughput can be forecast. A conversational layer may be a useful adjunct, but it must earn trust against that baseline rather than assume the baseline has disappeared.
Reopen condition
Reopen a Shift only with at least three independent story clusters across two source families, real deployment outcomes, and counter-evidence about regulated or high-volume use. The required comparison is a durable workflow replacement, not another launch announcement.