Secretary AI : Autonomous B2B Document Agent

Inspiration

Every business runs on documents it hates writing — quotations, purchase orders, customs letters, payment reminders. Teams lose hours copying old letters and swapping names by hand, and the tools meant to help don't: templates just wait to be filled, and AI writers produce nice paragraphs with no idea who the recipient is or what the company's terms are. We wanted something that behaves less like a text generator and more like a junior employee who actually does the task — then stops and asks a human before anything is sent.

What it does

Autonomous B2B Document Agent takes a plain-English goal — "Create a sponsorship proposal for Microsoft" or "Prepare all documents for exporting 100 smartphones to Dubai" — and drives the whole workflow to a reviewable draft. It plans which tools and documents the goal needs (producing a full multi-document package when required), recalls prior approved documents and preferences from a per-company memory, fills structured templates with verified data to reduce hallucination, self-evaluates for grammar, compliance, and factual confidence, and then stops for a human — showing sources, flagged uncertainties, and suggested attachments. Only after approval does it export, send, and log the outcome.

It's privacy-first: company data stays local, only the minimal fields needed go to the LLM, with an optional fully-local mode for banks, legal firms, and customs brokers.

How it works

The core is an Agent Orchestrator, not a bare LLM call, coordinating four layers: a Planner (turns the goal into ordered steps), Memory (per-company store of documents and corrections), Tools (LLM, compliance data, email, PDF/DOCX — selected dynamically), and a Validator (checks facts, compliance, and formatting; produces confidence scores). It runs as a loop — understand → plan → recall → gather → reason → draft → validate → human review → execute → learn — with fallbacks at each step when a tool fails.

Why it matters

A real agent is defined by decisions, planning, memory, and recovery, not by how well it writes. Human-in-the-loop is the feature, not a limitation: keeping a person responsible for the final communication is what makes the tool trustworthy for real business use. And constraining the LLM with verified data beats giving it a blank page — the answer to the genuine cost of a wrong name or address on a live B2B letter.

What's next

Broader multi-document packages, richer per-company memory, and a measured pilot: generation-time reduction, fact accuracy, and approval-without-edits rate against a manual baseline.