Inbox Agent
AI that reads your inbox so your team does not have to
Inbox Agent is an Appoly-built AI pipeline that turns inbound email into structured records. It reads procurement orders, booking confirmations, invoices, and portal-gated CSVs in whatever format they arrive — no templates, no regex, no data-entry team.
Most businesses think of email as communication. A lot of it is actually data ingress. Purchase orders from suppliers. Booking confirmations from OTAs. Invoices that land in a different layout from every contractor. Portal links that lead to a CSV someone still has to go and fetch by hand.
People have been trying to solve this for twenty years with rule-based parsers, regex, and per-supplier templates. They all break the same way — a supplier changes their layout and someone has to write another parser. LLMs changed that. A decent model does not need a template. It reads things the way a person reads things, and given the right scaffolding, it can act on what it reads.
We built Inbox Agent as a proof of concept, who handle procurement emails from a roster of suppliers that keeps growing. Every supplier has their own format — plain text, CSV, XML, scanned PDFs, portal links. Inbox Agent ingests whatever arrives, works out what is in it, and writes structured records straight to the database.
Four layers, each doing one thing well. The model handles comprehension. The tools handle side effects. The queue keeps it honest.
Email routes through Mailgun into a webhook on the Laravel app. The raw message is stored immediately — body, headers, attachments. Everything goes on a queue, so the system absorbs spikes without dropping messages and any email can be replayed later.
Before the AI sees anything, attachments are parsed to text. PDFs become readable content. CSVs and XML get extracted. Links in the body are pulled out. The agent gets as complete a picture as possible, in a form it can actually reason about.
An LLM with a small set of tools: visit a web page, download a file, store an order, store an order item, look up a supplier. Extraction is the model’s job. Tools only handle the things the model cannot do itself — fetching, persisting, checking. Decisions are emergent, actions deterministic.
Every agent decision, every tool call, every token used, every order produced is logged to dedicated channels. When it fails, the failure is obvious. When it works, the trail is there for anyone who needs to see it.
Use Cases:
If someone on your team currently opens an email to copy its contents into another system, there is a pipeline worth building. A proof of concept running against your real emails is a two-to-three week engagement. From there, it becomes a production system.