back-office agents
AI agents that clear your back-office queue in 30 days
Someone on your team reads a ticket, opens three systems, copies fields between them and writes a note, all day. Engineers with 25+ years in production build the agent that does it, with approval on every write, live in 30 days.
Send the queue today with volume and minutes per case. Within 48 hours you have the savings math and a price.
People triage tickets, copy between systems and follow a runbook that lives in a PDF.
- ticket queuehumans triage
- copy-paste between systemsmanual
- runbook.pdftribal knowledge
- RPA scriptbreaks on redesign
30 days to the first queue, and the next one goes faster
An agent with typed adapters, an explicit approval gate and one audit line per decision.
- agentGo · tool calls
- system adapterstyped
- human approvalexplicit gate
- audit trailper decision
From manual queue to agent in 30 days
- 01days 1–3
Pick the queue by arithmetic
Volume times minutes per case, minus the cases that still need a person. We do the math with you on day one and start with the queue that pays back fastest.
- 02days 3–10
Write the adapters
Typed calls into your systems with the ugly failures handled: rate limits, half-written records, the endpoint that returns 200 with an error in the body.
- 03days 10–20
Run it in shadow
The agent proposes, a person decides, and every disagreement goes into the eval set. Nobody's job changes yet, and you get an accuracy number measured on your own cases.
- 04days 20–30
Open the gate action by action
Actions with a clean record and a small blast radius go automatic first. The rest stay gated as long as you want. By day 30 the agent is working the queue.
What the platform does, what our engineers do
Reading the runbook and the ticket history
Platform
Including the cases handled against the runbook. That is where the real rules live.
What the agent may do without asking
Engineer
We recommend per action type. You sign it before anything runs.
Building the adapters
Both
Generated from the API contract, then fixed by an engineer against what the API really does.
Judging shadow-mode disagreements
Both
Automatic where the record shows the outcome. A person decides where your experts would argue.
The audit trail
Platform
Inputs, tools called, decision, approver. One searchable line per case, right or wrong.
Escalation paths
Engineer
What the agent does when it is unsure, and who picks up the case at 6pm on a Friday.
Shadow mode produces the accuracy number, measured on your own queue. You open the gate knowing what the agent does with every case.
The people behind the agents
SAP, Salesforce, Jira, Dynamics, the AS/400, a SOAP endpoint from 2006, the CSV dropped on an FTP server every night: our engineers have been connecting systems like these in production for more than 25 years. An agent is only as good as the integration under it. Integration is what these engineers have done their whole careers, which makes them the best people to build yours.
Shadow mode comes before any action goes automatic. The agent works your real queue, a person decides, and you open the gate holding an accuracy number from your own cases.
What operations leaders ask us
We already bought RPA. Why an agent?
RPA drives the screen, so it breaks on every redesign and understands nothing about the case. An agent calls APIs and handles the ticket that does not match the script, which is where your people spend their day. The approval gate and the eval set keep it predictable. Send us the queue your RPA gave up on.
What stops it doing something expensive?
It cannot call a tool it was not given, and every tool that writes is gated by default. Blast radius is set per action type before anything runs, in a document you sign. The proposed limits come with your quote.
How do we know it is right?
Days of shadow mode on your live queue, an eval set built from every disagreement, and a log line per decision you can open when someone complains about a specific case. You see the accuracy number before a single action goes automatic. Send the queue and we start measuring.
Does our data have to leave the building?
No. Our fine-tuned models run inside your network, air-gapped if needed, on your hardware or ours. The agent reads what it needs at the moment of the case, and what it read is in the log. Put your constraints in the form and the quote accounts for them.
What happens when we change a system it talks to?
The adapter is typed, so a contract change breaks the build the day it happens and nothing fails quietly at month-end. The next queue goes faster because the adapters already exist. Start with one queue today.
Hand us one queue
Describe the work as a person does it today, including the step where they double-check. Add cases per month and minutes per case. In 48 hours you have the savings math and a fixed price. In 30 days the agent is working the queue.
Case volume and handling time decide most of the price. The form asks for both.