The totals don't add up
Line items sum to 6,150. The invoice says 6,450. Relay does not create the bill; it flags the 300.00 gap with the numbers attached and tells #finance.
no bill · flagged · #finance toldAI operations agent · invoices, emails, tickets
Relay reads the supplier invoices and customer emails that land in your inbox, checks them against your own systems, and does the work: bills in QuickBooks, tickets in your helpdesk, a line in Slack. It refuses to guess, retries when an API is down, and hands anything doubtful to a person.
Built on a frontier language model, strict tool schemas, and the boring engineering that keeps agents running at 2am.
01Demo · run it on a real document
A normal supplier invoice. Expect: vendor matched, no duplicate, bill created, finance notified.
Checking for the backend…
Pick a document and press Run. Or edit the invoice and change a number — Relay will notice.
02Works with the systems you already run
03Track record
Every value on this page traces to the document or a tool result. That is what "0 numbers invented" means.
04How it works
The same loop runs on every document. Nothing is assumed, and every step leaves a record you can read.
The document is parsed, nothing is assumed. Vendor, invoice number, dates, every line item and the stated total.
What you see: the reasoning stream, word by word.
Vendor lookup, duplicate check, line-item reconciliation, approval threshold. Any miss stops the run.
What you see: one tool card per check, with its result.
Bill, ticket, Slack, in parallel when the steps are independent. Retried with backoff when a system is down.
What you see: the systems strip ticking as each action lands.
A plain-English summary and a full event log: every turn, call, retry and result, replayable and auditable.
What you see: the summary entry and the token and timing counts.
05Guardrails
Most of the value of an agent that touches money is in the things it declines to do without a person.
Line items sum to 6,150. The invoice says 6,450. Relay does not create the bill; it flags the 300.00 gap with the numbers attached and tells #finance.
no bill · flagged · #finance toldA vendor sends a second notice for INV-1042. Relay finds bill QB-1186, paid 12 Aug, and stops.
no bill · duplicate of QB-1186Anything above 5,000 goes to a person for approval, even when everything else checks out.
held for approvalGuardrails live in code and in the prompt. The model can't skip them.
06What it saves you
Three numbers you already know. The outputs update as you drag.
Your monthly pile300 documents
Model cost, estimated: $15 a month, at about five cents per document. About 1/75 of what it saves.
Typical build: two to four weeks. Pays for itself in the first quarter for most teams above 150 documents a month.
Estimate. Your numbers will differ.
07For the engineer in the room
One loop, six tools, an executor that expects the world to fail, and an event log that the console, the tests, and the recorder all share.
08Where it fits
The loop is the same whatever the paper is. The tools change to match your systems.
AP inbox QuickBooks/Xero
Vendor match, duplicate check, line reconciliation, approval threshold, bill created.
Support inbox tickets with priority and summary
Business impact sets the priority. The summary is what an engineer reads first.
Orders/POs ERP with duplicate and price checks
Every quantity and price traces back to the document or the price list.
What you get
Scope, guardrails and the go-live date are written down in week one, before any code.
09FAQ
Yes. The demo uses mocks so it runs anywhere; the tool contracts are written for the real APIs and swap in without touching the agent.
It stops and queues the document for a person with the reason and the numbers.
Documents are sent to the model API for the run and to your own systems. Nothing is stored by Relay beyond the event log you keep.
About five cents per document in model cost at current pricing, plus your hosting.
Yes, text and PDF; scans go through OCR first.
Every live run calls the model and costs real money. The demo allows a few runs per hour per visitor and a daily cap. Your own deployment sets both in one environment variable each.
Malik A. builds AI agents and API integrations for small companies. Ten years of shipping software, 210+ projects, a 5.0 rating.