We ship agents.We speak ERP.
Both fluencies, one team: agents answering a production RFQ inbox today, and years inside Infor and SAP implementations.
A workload first. A platform second.
Most enterprise automation starts with a demo environment. Corin started with an inbox that could not go down.
- The workload
- Corin's team built and operates the production email-AI workload that became Corin RFQ. It runs today on the sales inbox of a top-5 aviation MRO distributor — anonymized by agreement, RFQ emails read, extracted line by line, priced against live ERP state, and staged as Outlook drafts for a human to send. Manual quote turnaround there ran up to 3 days against buyers who decide in about four hours.
- The platform
- The loop that workload runs on — read the request, ground it in live ERP state, run the dry run, gate it, file the receipt in the ledger — a record of every decision: what changed, who approved it, when. That loop is now the Corin platform: specialist agents across Commerce — Quotes & Orders, Finance — Payables & Receivables, Trust — New-customer checks, and Service — Warranty & Support, for companies running Infor and SAP.
The machine that answers the machines.
Purchasing is automating first. Buy-side purchasing agents already generate and send RFQs at machine tempo — more requests, more follow-ups, tighter expected response windows — straight into supplier inboxes.
The supplier's side needs one too
Corin is the supplier-side answer: a decision engine that reads each machine-sent request, grounds it in live ERP state, and returns a complete, priced answer at machine speed — with a human on every consequential gate. RFQs came first. Orders, invoices, and claims run the same loop.
The buyer's side is a machine
A single buying agent now runs a department's worth of quote cycles before lunch — and it expects a structured answer back.
Hands don't scale
Answering machine-sent RFQs by hand loses twice: too slow to win the line, too expensive to answer them all. Adding people scales with headcount; the inbound volume scales with the buyer's compute.
What production taught us.
Seven rules, learned running a workload where a wrong part number or a missing cert date loses the order.
01Typed inputs, typed outputs
Every agent action declares its input and output schema. If the data doesn't fit the type, the action doesn't run — there is no 'close enough' path into the ERP.
02Evidence per decision
Every decision carries its sources: the message, the ERP records read, the policy version applied. A dispute is a lookup, not an argument.
03The dry run before the write
Nothing posts until the dry run has shown what will happen against live state and the outcome is reviewed. The write is the last step, not the first.
04Read-back
After every write, Corin reads the record back and compares it to what was authorized. A mismatch stops the line and pages a human.
05Replay tests
Every Procedure ships with replay suites run against historical traffic before release. A change that would have mishandled last month's inbox doesn't ship.
06Config-only setup
A deployment is configuration: channels, ERP connections, policies, autonomy levels. No professional-services fork, no custom branch of the product.
07Zero bespoke customer code
If a customer needs behavior the platform can't express, the platform changes — once, for everyone, versioned and tested. Customer-specific code is how systems become unmaintainable; Corin has none.
Three years, one direction.
2024First deployment
The first aviation email-AI deployment goes live: reading the shared RFQ inbox, drafting priced replies for human review.
2025Production scale-out
The workload scales to the full sales inbox — buyers who decide in hours.
2026The Corin platform
The production loop becomes the platform: the work graph, Procedures, the autonomy dial — agents from quotes to warranty claims, for Infor and SAP.