ERP support triage agent eliminates the Level-1 backlog

Customer support backlog had grown to ~340 open tickets. Level-1 triage took 12–20 minutes per ticket on average, and 35% of tickets were misrouted on first pass — every misroute became a customer-facing escalation churn.

• 1Manual ticket classification: 12–20 minutes per ticket across an overwhelmed L1 team
• 235% first-pass misrouting causing avoidable escalation cycles
• 3Inconsistent first responses depending on which consultant picked up the ticket
• 4No structured handoff brief — escalations arrived without context to the senior consultant

A supervisor-pattern agent that ingests email and form submissions, classifies the issue, queries the customer's Odoo instance for context (open invoices, recent modules, last login, current contracts), drafts a Level-1 response with the right module screenshots inline, and routes complex tickets to the right consultant with a pre-filled handoff brief.

The agent runs as a Python service alongside the existing helpdesk. Inbound channels (Zendesk / email / form submissions) all funnel through one normalising webhook that calls the agent. Outbound — drafts and ticket updates — goes back through the helpdesk's API so the existing audit trail captures everything.

• Inbound normaliser: Zendesk + email + form submissions to a single event shape
• Odoo MCP server: read-only customer / order / invoice / contract tools
• Helpdesk API: drafts written as agent-attributed comments; humans approve before send
• Operator dashboard: override + audit log in the team's existing helpdesk UI

Python service in the partner's existing container platform. Stateless behind a webhook entry; postgres for the agent's own decision log; redis cache for tenant-scoped tokens. Eval suite runs in CI and can promote/block model upgrades.

• Containerised Python service (existing infra)
• PostgreSQL: decision log + eval results
• Redis: tenant-scoped Odoo tokens (TTL 1 hr)
• CI: eval suite gate on prompt/model changes

Every ticket touched by the agent has a Langfuse trace with the ticket ID as session key. Operators can click "show trace" from the dashboard to see the agent's reasoning, the tools called, and the cost.

• Per-ticket Langfuse trace, session-keyed to the ticket ID
• Cost per ticket reported in the operator dashboard
• Override rate tracked per category as the leading drift indicator
• Weekly eval-suite refresh from overridden tickets

72% of Level-1 tickets are now resolved without human touch within four minutes of arrival. The escalations that do reach senior consultants arrive with a pre-filled handoff brief — module, customer context, attempted L1 response — so they start their work already 80% of the way through what was previously the first ten minutes.

Business outcomes

The partner reduced L1 staffing requirements by ~60% on the team handling agent-eligible categories and redirected those consultants to implementation work (their highest-margin offering).

• 01Customer-tier classification matters more than ticket-type classification. The same "invoice question" is a different ticket from a high-tier customer than a low-tier one — that nuance was missed in v1 and added in v2.
• 02The handoff brief is the most-loved feature with senior consultants. They will defend the agent because the briefs save them ten minutes per escalation.
• 03Eval cases sampled from real overridden tickets beat synthetic adversarial cases for tuning. Drift shows up in real data first.

The same Odoo MCP server is now feeding renewals workflows, onboarding agents, and proactive customer health scoring. Once the canonical MCP server existed, each new agent dropped from 8 weeks to ~2 weeks of engineering.

• Renewals follow-up agent reusing the same Odoo MCP
• Customer-onboarding agent: contract → setup checklist → kickoff scheduling
• Customer-health scoring agent: weekly synthesis from Odoo signals
• Cross-agent eval registry shared with the partner's in-house team

Considering an ERP-integrated agent?

A scoping call covers the ticket / workflow you want to automate first, the ERP scopes you are comfortable exposing, and the expected throughput / cost envelope.