Data and AI

About 40 AI assistants, one per person

Every member of staff has their own AI assistant in the company chat. It does real work in the ERP with that person's own permissions, and every conversation is traced.

~40assistants, one per person

~6,400model calls a day

~42,000conversation turns traced

how it works

What's left on my plate today?

personamemoryskills

sandbox $ erp wt list --mine --status open

your token 3 open tasks

You have 3 open tasks. The most urgent is this week's invoice reconciliation.written by AIawaiting approval

trace

turn

llm

erp

llm

the problem

A shared chatbot doesn't know who you are, isn't allowed to do anything in your systems, and nobody can check what it did. Staff needed an assistant that knows their job, can do real work, and leaves a clear trail.

So everyone gets their own: separate config, memory, skills and sandbox, signed in to the ERP with that person's own permissions.

results

Assistants do real work: lookups, task creation, invoice reconciliation, a daily ads review.
No assistant can do more than its owner is allowed to.
Scheduled jobs moved to a cheaper model after they turned out to be about 92% of spend.
Tool loops, errors and overly long turns are tracked so each assistant can be tuned.

stack

Hermes AgentMattermostDockerMCPERP CLILangfusesystemd

Next projectMongoDB 4.4 to 8.0 without stopping sales

want something like this?

Tell us about your problem. We reply within one working day.

Email [email protected]