An AI accountant that handles client requests over email
Whitelists, residency, attachments, and Ask before booking—mailbox setup that used to take half a day becomes an API create.
Discuss this post in AI
Send a pre-filled prompt to ChatGPT, Claude, Gemini, or Perplexity — get a summary, ask follow-ups, or compare ideas from this guide.
Clients still send bookkeepers and accountants requests by email: “Please book this,” “Where is my report?,” “Updated IBAN attached.” An AI accountant that only works inside a portal ignores how clients already behave. The workable pattern is a governed mailbox per practice or per client engagement, not a chatbot bolted onto a login wall that clients will route around.
Whitelists and scope
Start with an allowlist of client domains and known senders. The Neuro OS role ignores or quarantines everything else. Classify inbound into document intake, question, deadline change, or escalation. Attachments go through malware scanning and filename sanitization before object storage. Project memory stores the engagement context; the ledger connector proposes entries without posting them blindly into the books of record.
When classification is uncertain, forward to a human rather than guessing a booking. False confidence in finance is more expensive than a short delay.
Residency for financial mail
Financial correspondence often includes personal data and bank details. Choose mailbox residency that matches your obligations, including EU and 152-FZ constraints where they apply. Running Neuro OS with connectors and mail data in a region you control is a product decision, not a slogan. Keep credentials server-side. Limit retention of raw messages to what compliance requires, and document that retention next to the role definition.
Ask before booking
Anything that changes books—new entries, vendor master changes, payment instructions—defaults to Ask. The human accountant sees the draft entry, the source attachment, and the client thread. Questions that need only a status reply can be Allow after quality checks. Block outbound that would confirm wire changes without a second factor of verification through your existing process, even if the email looks authentic.
From half a day to API create
Historically, standing up a client-facing mailbox meant DNS tickets, shared credentials, folder rules, and a fragile IMAP integration that broke when a password rotated. With Agent Inbox, creating an engagement inbox is an API call during onboarding. Bind it to the role in Neuro OS, attach the whitelist, and document the public address for the client. Teams typically see cycle time drop from hours to minutes on routine document intake once the thread and the file land in one place.
Clear client expectations
Tell clients which address to use, what the AI will handle, and when a human will reply. Forward sensitive disputes to a named partner. Publish a short service description so clients do not treat the mailbox like an unbounded human associate. The AI accountant is a role with a mailbox, not a promise that judgment disappeared from the practice.
Email remains the client channel. Give the practice an identity on that channel with residency, Ask, and an onboarding path measured in API calls instead of afternoon-long mailbox archaeology.
Hand-off language clients understand
Clients should know whether a reply came from the AI accountant role or from a human partner. Use a short signature convention and escalation sentence when Ask routes to a person. Clarity reduces duplicate sends and reduces the chance a client pastes bank details into the wrong thread. Governance is partly UX copy, not only connector policy.
Agent Inbox gives each role a mailbox people can reply to, with forwarding into Neuro OS when a human must see the thread. Outbound mail defaults to Ask. Run the role on Neuro OS. To scope the first inbox, get started.