About Services Case Studies Pricing Blog Book Free Audit
AI-NATIVE CRM20262 min read

AI-Native CRM for Sales Inbox Automation

A CRM where every incoming Purchase Order, Quote, Invoice, and RFQ is automatically classified, extracted into a structured deal record with line items, matched or created against contacts and companies, and turned into a ready-to-work opportunity — with zero manual tagging, forwarding, or data entry.

CLIENT
Inflow
INDUSTRY
Sales Tech / CRM
TIMELINE
~14 weeks
ROLE
End-to-end product engineering (email ingestion, AI document classification/extraction pipeline, CRM data model, dashboard UI)
// overview

Inflow connects to a sales team's inbox, continuously scans incoming mail, and uses AI to detect and extract structured data from business documents — without a rep ever forwarding, tagging, or manually re-typing a single field. A PO lands; a fully populated deal record with line items, a matched or newly-created contact, a flagged deadline, and a draft reply are what the rep sees next.

// the problem

The deal already exists in the inbox — the CRM just doesn't know it yet.

Re-typing PO and Quote data from an email or attached PDF into a CRM form is tedious enough that reps skip it under time pressure. Urgent requests like "quote needed by Friday" carry no structural tracking, new senders never become tracked contacts, and every vendor formats documents differently — tables, unstructured text, scanned PDFs — breaking a one-size-fits-all parser constantly.

  • Continuous, automatic inbox monitoring — no forwarding, no manual tagging
  • Structured extraction from both email body and attachments
  • Automatic contact and company resolution from sender and domain
  • Ready-made deal records with line items already attached, plus a human-approved draft reply
// approach

How we built it.

01

Continuous inbox ingestion via provider push events

Inflow subscribes directly to inbox events — Gmail push notifications, Microsoft Graph change notifications — so every incoming message is scanned as it arrives.

02

A two-stage pipeline: classify, then extract

A cheap first pass classifies whether a message represents a PO, Quote, Invoice, or RFQ; only messages that pass go through the more expensive structured-extraction step.

03

Claude tool-use for structured extraction

The extraction step uses tool-calling to force a structured JSON schema directly out of email body text and parsed attachment content.

04

Vendor-specific learning as a stored feedback loop

When a rep corrects an extracted field, that correction is stored against the sending vendor/domain and biases future extractions from that same sender.

// what we shipped

The solution.

01

Zero-touch inbox monitoring

Inflow connects to Gmail and Outlook and continuously scans incoming mail with no forwarding or manual tagging required.

02

Two-stage classify-then-extract pipeline

A cheap classification gate filters out irrelevant mail before the more expensive structured extraction runs, keeping cost and false positives low.

03

Structured extraction across email and attachments

Sender, company, line items, quantities, pricing, and delivery terms are pulled via Claude tool-use, with a vision fallback for scanned documents.

04

Automatic contact and company creation

Unmatched senders become new CRM contacts and companies, auto-populated from their email signature and domain.

05

Ready-made deal records

Every detected PO or Quote becomes a deal/opportunity with line items already attached, so a rep opens a populated record instead of a blank form.

06

Draftable replies with mandatory human approval

A suggested reply is generated for every detected opportunity, but always requires rep review before it's sent to a customer.

// the results

Outcomes that shipped.

New deals appeared in the pipeline automatically with line items already attached, replacing manual re-typing of PO/Quote data. New contacts and companies were created automatically from unmatched senders, and extraction accuracy compounded per vendor as corrections accumulated instead of repeating the same mistakes. Validated a two-stage classify-then-extract pattern for high-volume inbox automation that keeps AI cost and false positives low, with a human-in-the-loop reply step protecting against acting on a wrong extraction.

// stack
PythonFastAPIGmail APIMicrosoft Graph APIClaude (tool-use)PostgreSQLRedisCeleryReact
// learnings

What we took with us.

  • 01Build a shared extraction-accuracy view sooner — confidence and accuracy trends per vendor were tracked informally at first.
  • 02Handle ambiguous documents with a clearer escalation path — early versions defaulted ambiguous messages to lower confidence rather than routing them to a rep for quick review.
  • 03Ship the vision-based fallback for scanned PDFs earlier — a meaningful share of real vendor PDFs turned out to be scanned images rather than native text.
// ready when you are

Have a build that needs this level of thinking?

Free 30-min audit. We'll tell you what to build, what it'd cost, and whether we're the right shop.