A Bid Review in Minutes: How OCR Skills and an AI Agent Caught a 3,959-Yuan Quotation Gap
Document-heavy compliance workflows—bid reviews, contract checks, supplier onboarding—still consume hours of manual cross-referencing across PDFs. This pipeline shows that off-the-shelf OCR APIs, wired together by an agent that can self-diagnose configuration failures, can shrink that to minutes while producing a page-cited audit trail that holds up under scrutiny.
A complete bid review pipeline was assembled by combining two Tencent Cloud OCR Skills—one for document field extraction and one for borderless table recognition—inside the WorkBuddy AI agent. The process turned 8 page-by-page PDFs from a fictional printer procurement into four structured deliverables: a compliance matrix, a graded risk list, a quotation comparison CSV, and an executive summary. The agent corrected a supplier's self-assessed "non-compliant" delivery time by recalculating `12 <= 15` against the tender threshold, and it surfaced a 3,959.96-yuan mismatch between the same supplier's quotation summary and its independent pricing sheet.
The workflow was built without manually writing Skill definitions or configuring API keys from scratch. When the OCR Skills failed on first run, the agent diagnosed the missing environment variables and Python SDK dependency, then guided the user through Tencent Cloud's CAM console to create credentials and activate the OCR service. A custom business Skill, `tender-compliance-reviewer`, was generated from a natural-language description of the review rules, locking in the field schema, judgment logic, and evidence-formatting requirements so the process is repeatable.
Processing time per supplier set was tens of seconds to a few minutes, compared with an estimated hour or more of manual cross-referencing. The final output distinguishes hard failures (resolution below spec, missing warranty) from items needing clarification (conflicting amounts, missing attachments), leaving the procurement officer with a structured, page-cited audit trail rather than a black-box recommendation.
The workflow's most practical design choice is separating "hard non-compliance" from "needs clarification" in the output. Conflating a missing attachment with a failed spec would force the procurement officer to untangle the agent's reasoning; keeping them distinct makes the matrix immediately actionable.
Generating the business Skill from a natural-language description, rather than hand-authoring SKILL.md and validation scripts, lowers the barrier to locking in domain-specific review rules. The model still calls deterministic OCR APIs underneath, so the "customization" is about fixing the schema and judgment logic, not hoping a general-purpose LLM gets the comparison right every time.
The agent's self-diagnosis loop—check Skill files, check env vars, check Python imports, report specific failures—turns the usual "it doesn't work" dead end into a sequence of fixable steps. This is a concrete pattern for making AI tooling usable by people who do not know where API keys live or how venvs load environment variables.
The deliberate decision to retain original decimals (152,000.04 yuan, not 152,000) and to avoid overwriting the independent quotation sheet with the response file's summary figure is what makes the output auditable. Rounding or merging sources would have hidden the 3,959.96-yuan gap that became the key finding.
Cache reuse on identical inputs is a double-edged feature: it saves API costs on repeated runs but can mislead performance measurements if a cached hit is mistaken for a fresh OCR call. The note about creating a new session for benchmarking is a small but important operational detail.