Sales and support teams live inside ERPNext, but ERPNext native has no generative layer. Reps retype the same quotation follow-ups, read through a long Lead or Issue timeline before every reply, and answer questions like "which open opportunities in the GCC closed above 40% margin last quarter" by hand-building a report. ERPNext ships list views, report builder, and email templates, but none of that drafts prose, summarizes a 30-comment thread, or turns a business question into a filtered query. That gap is where hours quietly disappear.

ECOSIRE builds erpnext_ai_assistant, a proper Frappe app (its own module, DocTypes, and hooks.py) that wires an LLM directly into your ERPNext workflow. A "Draft with AI" button on Quotation, Lead, Opportunity, and Issue (via client scripts) calls a @frappe.whitelist() server method that assembles context from the linked records and returns an editable draft into the Email/Communication dialog — nothing is sent automatically, the human always approves. A conversational side panel answers natural-language questions by translating them into permission-safe frappe.get_list/query-report calls, so the model reads data through the same role permissions and User Permissions the logged-in user already has — it can never surface rows a user isn't allowed to see. Record and thread summarization runs over the Communication and Comment timeline; lead scoring and next-best-action are computed on a Scheduler event and written to custom fields you can sort and filter on. A configurable AI Prompt DocType lets you tune tone, guardrails, and per-DocType behavior without a redeploy, and an AI Interaction Log DocType captures every prompt, response, token count, and cost for auditability.

Technically it is a clean, upgrade-safe app: LLM calls run through Frappe hooks.py doc events and background jobs so the UI never blocks; API keys live in Site Config (never in code or the database in plaintext); a Knowledge Base retrieval layer indexes your chosen DocTypes and files so answers are grounded in your own content rather than the model's guesses. It targets Frappe/ERPNext v15 and v16, respects the framework's permission model end to end, and adds zero core patches — it is installable and removable with bench install-app / bench uninstall-app.

Because this is build-to-order, we start from your actual workflow. After a short scoping call we confirm which DocTypes, actions, and questions matter to your team, pick and configure the LLM provider you prefer, then build, test on a staging bench, and hand over. Typical delivery is 2-4 weeks from confirmed scope. You receive the installable source for your version, install and configuration on your environment, documentation, a training session, and a post-go-live support window — plus the git repository so your team fully owns what we build.