Skip to content
View case studies
AI / Operations2025United States

AI Workflow Assistant

Document understanding wired into an existing business process.

Role
Full Stack Product Engineer
Industry
AI / Operations
Client
SaaS Company
Market
United States
Duration
4 months
Status
Live
ExtractionCited
invoice_noINV-20418
total_due$12,400.00
po_refneeds review
Overview

What it is

An assistant that reads incoming operational documents, extracts the fields a team actually keys in by hand, and routes anything it is unsure about to a human, with the source passage attached.

Problem

What was going wrong

An operations team was reading PDFs and emails and retyping the same eight fields into an internal tool. It was slow, it was inconsistent between people, and a generic chatbot pilot had failed because nobody could tell where an answer came from.

  • Extract structured fields from unstructured documents
  • Make every extraction traceable to its source text
  • Route low-confidence results to a human instead of guessing
  • Keep inference cost proportional to document volume
Solution

How it was built

Documents are chunked and embedded into a vector index; extraction runs as a structured tool call against Claude with a strict schema, so the model returns typed fields rather than prose. Each field carries the passage it came from and a confidence value, and anything below threshold lands in a review queue where a human correction is captured as training signal. Responses are cached by document hash so re-processing costs nothing.

Challenges

The parts that were hard

01

An assistant nobody trusted

WhyThe earlier pilot returned confident prose with no provenance, so reviewers had to re-read the document anyway and the tool saved nothing.

FixEvery field returns with the exact passage it was drawn from, turning review into a glance at a highlight rather than a full re-read.

02

Cost that scaled with retries, not with work

WhyRe-processing a document after any pipeline change re-paid for inference that had not changed.

FixResults are cached against a hash of the document and the prompt version, so only genuinely new work reaches the model.

Outcome

What changed

  • Manual re-keying replaced by review of pre-filled, cited fields
  • Uncertain extractions surfaced for humans instead of silently guessed
  • Repeat processing made effectively free through hash-based caching
Takeaway

What I'd carry forward

The hard part of shipping AI in a business process is not the model. It is provenance, thresholds and a human path for the cases the model should not decide.

Start here

Have a product idea, or one that stalled?

Tell me what you’re building and who it’s for. If I’m the wrong fit I’ll say so on the first call and point you somewhere better.

Based in Jaipur, India (GMT+5:30). Replies within one business day.