Switchboard is the intake layer for your AI agents. It takes documents and requests from every channel, checks them, sorts them into your own categories and patches each one through to the right agent, person or process.
Designed for the AI era. Developed by AI.
Live board · illustration with invented documents
Channels
Destinations
One customer case comes in by email, a portal upload, a phone photo and a partner feed, each with its own format and failure modes.
Each agent project writes its own mailbox polling, file parsing and classification, and maintains it alone.
Between channel and agent there is no shared record of what arrived, what happened to it and who has it now.
Every document takes the same path, whatever channel it came from. Each step writes to the document's status record, so operations, IT and the receiving agent all see the same thing.
Inboxes, upload endpoints, chat, partner APIs and SFTP folders become intake channels of one tenant.
A file edited after it was created, or one whose structure could not be verified, goes to a person.
The model you chose sorts the document into one of your own categories, with a confidence score.
Text and tables come out of PDFs, scans and photos, so the agent gets content, not just a file.
Your rules pick the destination: an AI agent, a review queue or a downstream system. The outcome flows back.
Switchboard is installed into your AWS account as a versioned package. Documents are stored and processed there, in the region you choose, and you decide which model reads them.
Your AWS account · your region
Outside, only if you choose
Another model provider Set per kind of document in the model registry. Switching is configuration, not code.Each version ships with a signature, a software bill of materials and a licence list.
You apply each release when it suits you. The current and the previous version are supported.
Everything Switchboard does is visible to the security tooling you already run.
Ask us about the pilot programme in a discovery call.
No mailbox polling, no file sniffing, no guessing what a PDF is. Subscribe an agent to a route and it receives each matching document with its category, source and extracted content.
→ one versioned event per routed document
→ the original file plus the extracted text
→ report the outcome back into the record
→ any agent framework, any language
first_notice_of_loss → claims-agent
policy_change → policy-agent
complaint → person
onboarding_document → kyc-agent
loan_application → credit-agent
statement → archive
bill_of_lading → shipping-agent
customs_document → person
proof_of_delivery → archive
supplier_invoice → invoice-agent
hr_document → person
contract → archive
For companies where documents carry personal and contractual data, and where IT needs to know exactly what runs where.
Channels, rules, models and documents belong to a tenant. One business unit never sees another's traffic.
Each classification, routing decision and handover is recorded with its time and the model or rule that made it.
Decide which models may read which documents, and change that centrally without touching your agents.
Flagged and uncertain documents wait in a review queue. A person confirms or corrects, and the document carries on.
In a 30-minute discovery call we walk through your channels, the agents you are building, and what a pilot in your AWS account would look like. The demo runs on invented sample documents.
Write to us
contact@agenticswitchboard.net