Explore the demo agents
The fastest way to see what Sema4.ai can do is to try the three demo agents on the Welcome page. They're ready to go — no setup, realistic data, and sample inputs included. Open one and start.
Invoice Coding
Upload a vendor invoice (PDF) and the agent extracts it, matches the vendor, and predicts the GL account and cost center for each line — learned from years of how Meridian has coded similar invoices. It flags anything that needs a human: new vendors, anomalous amounts, possible duplicates, and low-confidence guesses. A great look at document extraction + a semantic data model working together, with results in a reviewable data frame.
AP Help Desk
Paste a vendor's "where's my payment?" email and the agent looks it up in Meridian's AP records and drafts a reply — invoice status, what's outstanding, recent history. It shows how an agent answers questions over a database in plain language (no SQL), using a semantic data model.
Receipt Roast
Drop in a receipt or invoice and Receipt Roast gives it the forensic treatment: every line extracted and totaled to the cent, followed by a deadpan verdict on the spending. It's the expense report that judges you back. Behind the dry wit it's a clean demo of document extraction + exact data-frame math — try the built-in examples (Restaurant Guacaya, Restaurant Aster, Car rental Hertz) or feed it your own and brace yourself.
About the demo
The data — Meridian Capital Partners
The accounting agents run on a prewired dataset that mimics the books of Meridian Capital Partners, an imaginary mid-size investment firm in New York. It's a real PostgreSQL database ("AP System Data") with about four years of accounts-payable history — roughly 6,700 coded invoice line items from 55 vendors (cloud providers, market-data vendors, consulting, legal, facilities, fund ops), allocated across 31 departments.
A semantic data model sits on top and maps the raw tables — vendors, invoices, invoice_lines, gl_accounts, cost_centers — into business concepts the agents reason about in plain language. It even captures the messy real-world patterns: Bloomberg always codes to the Trading Desk, while Deloitte splits across Compliance, Finance, and Tax depending on the engagement.
Sample inputs
Each agent ships with inputs you can try right away — sample invoice PDFs, receipt images, and starter prompts. Use those to get a feel for it, then bring your own files and questions.
Make them your own
The demo agents are read-only — you can run them, but not edit them.
To change one, duplicate it (from the agent's ⋯ menu) and edit your copy freely — runbook, tools, data, and settings. You can also build your own agent on the same Meridian dataset: attach the AP semantic data model and start from a blank agent — see Build an agent from scratch.