Thousands of past support emails are full of good answers — yet every new inquiry gets searched by hand or typed again from scratch. SupportFlowAI automatically matches new inquiries against your historical email archive and proposes a ready-made reply draft — via RAG with a local Ollama model, running entirely on your own hardware. No cloud requirement, no dependency on an external AI provider.
No cloud requirement: Embeddings and reply generation run fully locally via Ollama on your own machine or server by default. Cloud LLMs (OpenRouter/OpenAI/Anthropic) can optionally be enabled under "AI Connection" as a deliberately limited fallback — never as a prerequisite.
From importing historical mail to the sent reply, SupportFlowAI covers the complete workflow — without manually training your own language model.
Historical support emails are ingested via file import (EML/MBOX) or live IMAP, deduplicated, and turned into cleaned question/answer pairs — quotes, signatures, and promotional footers are stripped out automatically.
Every new inquiry is compared against the entire corpus of past cases via local Ollama embeddings (cosine similarity) — the most similar historical replies become the source for the draft.
Given sufficient confidence, a local language model drafts a reply based on the matched cases. At low confidence, it deliberately returns a neutral holding reply instead of a freely invented answer.
The three-column view shows the original inquiry, the editable draft, and the underlying source cases side by side — Approve & Send, Reject, or Regenerate in a single click.
Above a high confidence threshold (default ≥ 0.85) a mailbox can reply fully automatically — with a two-step confirmation on activation and restricted to the SystemAdministrator role.
Email addresses, phone numbers, and IBANs are automatically masked in stored training examples. Reviewers and system administrators can see unmasked data when needed; auditors always see masked data.
Four fixed roles — SystemAdministrator, AccountManager, Reviewer, and Auditor — govern who manages accounts, approves drafts, activates full automation, or only observes read-only.
Duplicate candidates above 97% similarity, outdated entries, and recurring inquiries with weak confidence are detected automatically — keeping the knowledge base clean and current.
A central dashboard shows queue depth, open reviews, average generation time, error rate, and the heartbeat of all background services at a glance.
No model training, no prompt engineering on your part — SupportFlowAI learns from your own archive and keeps improving with every reply you send.
Add an email account via file import or IMAP — including selecting both the inbox and sent folder.
Add signature/promotional patterns, run a test import, and check the knowledge base with an ad-hoc test question.
New inquiries land as drafts in the review queue — approve, edit, or request regeneration.
Once you trust the corpus, SupportFlowAI can fully automate sending above a high confidence threshold.
Fully automatic sending is deliberately not a switch anyone can flip: it only engages above a high confidence threshold, can only be activated by the SystemAdministrator role, and requires a two-step confirmation — a checkbox plus re-entering the password.
Instead of a model trained once and then static, the knowledge base keeps growing: every reply that's actually sent automatically becomes a new question/answer pair in the corpus — with no manual retraining step.
SupportFlowAI is built for traceable, controllable replies — with governance, privacy, and operational safety that go beyond plain text generation.
By default, a locally running Ollama model handles both embeddings and text generation. Optionally, a cloud provider can be configured under "AI Connection" as a limited fallback with its own priority — embeddings always stay local regardless.
During initial setup you choose the hardware mode — CPU models (e.g. phi3.5, all-minilm) for lean servers, GPU models (e.g. llama3.1:8b, bge-m3) for higher answer quality when a GPU is available.
The original inquiry, the editable draft, and the source cases sit side by side — so every approval can be traced in seconds instead of switching between windows.
Login, approval/rejection, sending, and every full-automation toggle are logged immutably — customer email addresses are always stored masked.
Queue depth, open reviews, average generation time, the last 24 hours' error rate, and the heartbeat of the three background services (mail sync, draft generation, auto-send) show at any time whether operations are running smoothly.
Nothing missed: A growing backlog in the queue shows up on the dashboard before customers have to wait for a reply.
Three short guides — whether you're setting up SupportFlowAI, using it for the first time, or getting more out of daily use.
Ollama setup for CPU/GPU, .NET 10, first launch and the first admin account.
For new usersFrom an empty system to an automated reply — as a 10-chapter tutorial series.
For daily useFine-tune confidence thresholds, use curation, scope roles correctly.
Write to us — we're happy to show you how SupportFlowAI can be deployed in your environment.
View pricing