Autonomous outbound sales engine for an industrial manufacturer

Context

In the conservative B2B market for construction composites, sales rest on specific people. A manager spends years corresponding with developers, contractors, and procurement staff — and all of that knowledge lives in a mailbox. A formal CRM often does not exist: the mailbox is the CRM, unreadable to the business. When the manager drops out, the customer base, deal history, and tone stay locked in correspondence. And the work happens on a live corporate mailbox, where a sending mistake is instantly visible to real customers — automation is acceptable only inside a strict approval workflow.

The task

The brief was to seat an agent directly in a sales manager's working mailbox: connect over IMAP, reconstruct the manager's work, mine the customer base and price lists from the correspondence, then reactivate dormant relationships — in the tone and under the signature of a specific person. The product line is narrow and technical — composite and fiberglass rebar, mesh, fiber — so the agent had to name the line correctly and hit the right price tier for each order volume.

Approach

The email transport is pure Python standard library — imaplib and smtplib, zero external dependencies — a deliberate choice for a tool on live corporate infrastructure. Credentials live in .env; mailbox access uses an app password.

Stage one was reconstruction: the system surveyed a mailbox of 14,754 incoming emails and rebuilt the manager's working picture — correspondents, cadence, signature — collecting style samples so the generators write like this specific person. Stage two was data: structured CSVs of inbound customers and 898 real correspondence contacts — the company's de facto CRM — plus the client's PDF price lists parsed into a JSON catalog for quote calculation.

Architecture

The live environment bit immediately: the Sent folder was empty — all outgoing mail sat in the INBOX — so 834+ outgoing emails were extracted in a separate FROM-self pass to reconstruct history, tone, and signature. The system is deliberately a set of single-responsibility scripts rather than a monolith — connection check, mailbox survey, contact mining, style export, email and quote generators, a guarded sender — each verified independently.

The price catalog is a multi-tier grid — three tiers for rebar, two for mesh — and quote.py selects the tier from the order volume on its own. The second source was 1C:UNF: with OData closed by restricted permissions, the base was obtained via UI export to xlsx and parsed with scripts — 3,712 contacts, segmented into 1,242 with email, 2,892 with phone, 1,203 with both channels.

Approval workflow

send.py defaults to dry-run: nothing goes out without an explicit --send flag, and even then only emails from an explicit APPROVED list, with every send logged — one-click mass mailing does not exist by design. Approval runs on an interactive HTML board: a checkbox per email, previews, automatic supplier/competitor and duplicate-domain flags, one-click copy of the approved list. The reports record a key caveat: the base is mostly procurement staff and engineers — an entry point, not final decision-makers — and the reactivation strategy is built accordingly.

Outcome

By handover the engine had reconstructed the manager's history from 14,754 incoming emails, consolidated a working CRM of 898 correspondence contacts, and extended it with 3,712 contacts from 1C:UNF — the company now holds its customer base as a manageable, segmented asset. The quote generator prices orders against the real catalog, and drafts are queued: 10 first-wave reactivation emails and a mass wave of 85 emails across 69 companies. Response and conversion will be measured from the first wave's actuals; the engineering result — a repeatable model for recovering commercial expertise from a mailbox — is already in hand.

What we built

• Connecting to the live mailbox

  An IMAP/SMTP connector in pure Python stdlib on top of the working mailbox; credentials live in .env and are never committed.

• History and style reconstruction

  The manager's outgoing emails extracted straight from the INBOX keyed on FROM self (834+); tone and signature reconstructed from the bodies for the generators.

• Mining the customer base

  CSVs assembled from the mailbox: inbound customers, 898 real correspondence contacts (the CRM), and a list of warm repeat buyers for the pilot.

• Price catalog from PDF

  The supplied price lists parsed into JSON with a three-tier grid for rebar and a two-tier grid for mesh; the quote generator runs on this catalog.

• Email and quote generators

  Scripts assemble reactivation emails and the mass wave (85 emails / 69 companies) with name substitution from a dictionary; quote.py calculates quotes for the order volume.

• Second source: 1C:UNF

  With OData closed, the database was obtained via a UI export to xlsx and parsed with scripts: 3,712 contacts, phone and address normalization, channel segments.

• Approval board

  Interactive HTML with checkboxes, email previews, "supplier/competitor" and "duplicate domain" flags, and one-click copying of the approved list.

• Guarded sending

  send.py defaults to dry-run, actually sends only emails from the APPROVED list with the --send flag, and keeps a log; a human in the loop is mandatory.

Engineering challenges