Use case • Accounting & Professional Services
AI inbox copilot for a small accountancy firm
Monitors a shared Outlook inbox, classifies every incoming client email, decides if a reply or follow-up action is needed, and drafts or sends responses grounded in the firm's knowledge. The team went from manually triaging >50 client emails a day to reviewing high-quality drafts, one-click approvals and focusing on edge cases.
What this agent does for the team
Built for a growing accountancy startup handling dozens of client emails every day across tax, payroll, and general queries.
Before vs after
Before
One support/admin staff member manually triaged >50 emails / day: categorising, forwarding to accountants, responding to routine questions, and chasing missing information.
After
The inbox is auto-classified and prioritised. Routine queries arrive pre-drafted or already sent; complex cases land in the right folder with clear labels, so the team spends most of their time on approvals and edge cases instead of inbox triage.
How the inbox copilot works
A chain of specialised AI decisions orchestrated in n8n, backed by a vector knowledge base and logging in Google Sheets.
1. Listen to the shared inbox
An n8n workflow monitors a shared Microsoft Outlook inbox. Each new email triggers the agent with message body, headers, thread context and basic metadata.
2. Multi-dimensional classification
The first AI step classifies the email into labels (e.g., tax return, payroll, onboarding), priority (low / medium / high), whether a reply is expected, and whether further action is required.
3. Move & organise the inbox
Based on the classification, the email is labelled and moved into the right folder in Outlook, keeping the primary Inbox clean and organised for the team.
4. Search the knowledge base
If a reply may be needed, the agent queries a vector database in Qdrant using OpenAI embeddings to find relevant documents: policies, FAQs, historic answers or templates.
5. Draft or send a reply
Using the classification, KB snippets and email thread, the agent drafts a human-like reply. Depending on the `agent.enable_autoreply` flag, it either sends directly (marking the email as read) or saves a draft for a human to review and send.
6. Notify on high-priority emails
For high-priority or sensitive topics, the workflow sends a notification so someone can review quickly — even if an initial draft was prepared automatically.
7. Log & evaluate decisions
Key decisions (classification, priority, reply/no-reply, actions) are logged to private Google Sheets for monitoring. A separate evaluation sheet holds test cases and metrics to validate changes before rollout.
Tech stack for this agent
Every classification, priority decision and reply action is logged to private Google Sheets: one for runtime logging, and one for evaluation scenarios and metrics. When prompts or logic change, we re-run evaluations before pushing to production, so you can see accuracy and behaviour improve over time instead of “trusting the model”.
Safety, control & human oversight
Designed so the firm can start with review-only drafts, then selectively enable full autoresponding where it’s safe.
Human-in-the-loop by default
The firm started with `agent.enable_autoreply = false`, so all messages were drafted but not sent. Staff reviewed, edited and sent replies while building trust in the system.
Scoped knowledge & access
The agent is restricted to a vetted knowledge base and specific mailboxes. It cannot invent new policies or send messages outside the authorised accounts.
Tested on edge cases
A separate evaluation workflow runs the agent against tricky scenarios (ambiguous emails, partial data, sensitive topics) and logs metrics before any major change is deployed.
Want a similar automation for your team?
We adapt this pattern to your stack, compliance requirements and tone of voice — whether you’re in accounting, legal, consulting or another professional service.