The email to be routed, the PDFs retyped by hand, the answers that are always the same. An agent takes them on and stops where a signature is needed.
Almost no business process should be automated end to end, and anyone promising otherwise has never put one into production. What gets automated are the steps: reading, classifying, extracting, preparing. The decision stays where it is.
It is also why these projects pay for themselves quickly. They do not take a person out of the process: they take out the part of the process that does not need that person's head.
Email, tickets, enquiries from the website. It recognises what they are about, what language they are in and who they belong to, and files them accordingly.
Invoices, delivery notes, orders, contracts. The fields land in the ERP without anyone retyping them, with the uncertain ones flagged rather than silent.
The draft arrives with the data already looked up across the systems. Whoever answers reads it, corrects it if needed and sends: what you save is the lookup time, not the judgement.
Every action carries a log with input, output and the model used. When something goes wrong you can reconstruct what happened, which in a business process is not optional.
Orders arriving by email, as PDFs, on WhatsApp. An agent reads them, checks prices and availability in the ERP and prepares the order for confirmation.
The info@ inbox receives everything and someone has to sort it by hand. AI reads every email, gives it a category and forwards it to the right department. It replies to nobody: it routes.
Delivery notes, supplier confirmations, contracts, certificates. AI recognises every document, extracts the data for the ERP and flags what a person needs to look at.
Closed tickets, commits and pull requests become the end-of-sprint report and the release notes, in two versions: technical for the team, readable for clients and users.
High-volume, repetitive ones with a verifiable outcome: routing, data extraction, first reply. Rare processes full of exceptions cost more than they return, and we say so up front.
Only if you decide it should, and almost never on the first pass. You start with a draft that needs approval, measure how often it gets corrected, and remove the approval where the numbers say you can.
That has to be designed in advance, not afterwards. For every step you decide what the system does when it is unsure: stop, ask, or hand over to a person. An agent that cannot stop does not go to production.
No. The agent works on the systems you have: it reads from the mailbox you already use and writes into the ERP you already use, whether that is TeamSystem, SAP Business One or another one with an API or an export.
A free 30-minute call, with no commercial follow-up unless you want it.