What we shipped, and where it stops.

Three agents running on real data. For each one: the scope, the permissions, who approves an action, and how a bad run is undone.

NX-01

Agent Support

Retail, 340 staff

Tier one tickets answered in minutes, not hours.

Infrastructure running the Agent Support deployment
Problem
A support team of eleven handling 1,400 tickets a month, roughly half of which were five recurring questions about orders, returns and account access.
Scope
Draft a first reply for those five categories in English and Dutch. Escalate anything touching a refund over 50 EUR, a complaint, or an account the agent cannot verify.
Permissions
Read on the ticket store and order history. Write limited to draft replies and internal notes. No access to payment data.
Approval
Replies in the five in-scope categories send automatically. Everything else is drafted and waits for a human.
Rollback
Any drafted reply can be discarded before send. Sent replies are logged with the retrieved sources and can be followed up by an agent within the thread.
NX-02

Agent Revenue

B2B SaaS, 60 staff

The CRM updates itself after every call.

Abstract visualisation of the Agent Revenue model
Problem
Eight account executives spending an estimated four hours a week each on call notes, CRM hygiene and follow-up drafting, with inconsistent field completion.
Scope
Summarise recorded calls, propose the next step, draft the follow-up email, and update a fixed set of CRM fields. Never change stage, amount or close date.
Permissions
Read on transcripts and account records. Write on notes, activity and six named fields. Stage and pricing fields explicitly out of scope.
Approval
Notes and field updates apply automatically. Follow-up emails are drafted and require the owner to send.
Rollback
Every write carries the agent identity and run id. Field history is preserved, so a change can be reverted from the CRM audit view in one step.
NX-03

Agent Ops

Industrial distribution, 900 staff

Supplier invoices read and checked without retyping.

Document processing pipeline behind Agent Ops
Problem
Four thousand supplier invoices a month arriving as PDF and scanned paper, keyed by hand into the ERP with a measured error rate on line items and tax codes.
Scope
Extract header and line data, match against the purchase order, flag discrepancies over a tolerance, and stage the result for approval in the ERP.
Permissions
Read on the invoice mailbox and purchase orders. Write limited to a staging table. No posting to the ledger.
Approval
Nothing posts automatically. Matched invoices inside tolerance are queued for one-click approval. Discrepancies route to a named controller.
Rollback
The staging table is the only thing the agent touches, so a bad run is cleared without any ledger impact. Source documents are never modified.

Your process, next.

Tell us which process costs your team the most time. The diagnostic exists to check whether an agent is the right answer before you commit.

Book a diagnostic