Case study

ERP Agent with more than 20 skills integrated across 5 systems

Less than 2 months from kickoff to production. Effectively handling meeting summaries, ticket triage, finance queries, internal lookups across systems and more. Integrated with the ERP system, Microsoft Teams, Jira, email, and Telegram.

Live · ERP Agent

20+

skills in production

Reclaimed weekly

30-40h

across the senior team

Kickoff to production

<2 months

live and stable since

Last updated:

01. The problem

The pain wasn't technology - it was attention

  1. The ops lead spent a meaningful chunk of each day answering "where did we land on X?" in DMs.
  2. Every Friday afternoon, someone manually composed the weekly status across three systems for the leadership thread.
  3. Junior staff would ask the same five questions at different times, each pulling someone senior out of focus.
  4. Customer-facing tickets coming in through email and Teams were triaged manually - slow, inconsistent, occasionally lost.

They were using ChatGPT custom GPTs for small repetitive tasks. It was not integrated with any system, required manual upload of data. Lacked broader context and long-term memory. This was leading to inconsistent results and low usage across the team.

The Brief

A system that the senior team would actually trust, that knew the same facts they knew, that could act on its own.

02. The stack we built

Twenty narrow skills around one orchestrator

  • Orchestrator + skills architecture

    Single coordinator routes incoming requests to one of 20+ specialised skills.

  • Tool layer

    Typed wrappers around the client's ERP, ticketing, communication channels, and document store. Permissions enforced at the tool level - the agent literally cannot fetch what the calling user doesn't have access to.

  • Models

    Hybrid: Anthropic Claude for the reasoning-heavy skills (multi-step queries, writing, judgment calls); Gemma 4 on a smaller GPU for the high-volume operational lookups. Routing happens at the orchestrator.

  • Channels

    Microsoft Teams (primary) + Telegram (for limited number of admin users).

  • Persistence

    Postgres for conversation history, audit log, run trace. Custom analytics layer powering the monitoring system.

  • Monitoring

    Custom analytics layer with per-skill success rates, p50/p95/p99 latency, token and cost telemetry, full run-trace replay, anomaly detection on rejection-rate spikes, per-user usage patterns, tool-call analytics, and operational alerts on error-rate or latency-budget breaches. Live dashboard at an internal URL.

03. Implementation

Less than two months from kickoff to full rollout

  1. Shipped
    01Week 1-2

    Discovery - sitting in on actual work for two days, building a backlog of 30+ candidate skills, scoring by frequency × time saved × risk. Picked the first 5 to ship.

  2. Shipped
    02Week 3-4

    First five skills shipped to a private Teams channel with two pilot users. Daily standup with the ops lead; weekly demo with the leadership team.

  3. Shipped
    03Week 5-6

    Expanded to all senior admins. Tightened the orchestrator. Added the audit log + run trace for every skill execution.

  4. Shipped
    04Week 7-8

    Full rollout to the wider team. More than 20 skills in production. Wrote the runbook + handover doc.

04. The numbers

Production traffic, single-digit error rates

Skills running in production
20+
Daily active users on the agent
~20
Conversations / month
2000+
Skill execution success rate
97.6%
p95 latency · retrieval
2.4s
p95 latency · multi-step
8.1s
Time reclaimed per week
~30-40h
Tickets shipping without human touch
~22%
Uptime · 90 days rolling
99.4%
AI that already runs

Bring us a similar setup.

Most ERP integrations have similar shapes - five to fifteen named workflows, the same questions asked in DMs every day, attention concentrated in a handful of people. Diagnostic call is free; written summary either way.

Live · powered by Gemma-4 · running on our hardware in Sofia