AI assistant for a one-person business
An assistant setup for a one-person business: business notes, task preparation and an email integration trial. This case describes the setup and its limits, not automatic processing of current customer enquiries.
Setup source review: 11 September 2026. This description does not certify current operation.
For example: from email to reply draft
Invented example. This illustrates a way of working, not a completed client task or an actual agent output.
The email
“Could you prepare a report for next week? What would it cost?”
What needs clarifying?
- Which data should the report cover?
- Who is the report for?
- Scope, deadline and price have not been agreed.
A possible reply draft
“Thanks! Which data would you like the report to cover, and who is it for? That will help me assess the scope and prepare a quote.”
I review the draft. Price and the deadline need my decision, and sending the email needs separate approval.
This is not an automatic responder. The example does not prove that every email is found or that every tool stops without approval.
What was built?
A containerised agent service connected an allowed messaging channel, local interface, scheduler and Markdown knowledge vault. The Gmail connection used Google sign-in permissions (OAuth) to check unread emails. Sending required a separate approval step. Secret masking in logs and a policy-rule prototype were also examined. These parts did not constitute proven universal data-leak prevention.
Business context makes the draft useful.
- 01Customer enquiry
Read the question and identify the need.
- 02Business context
Find the offer and relevant agreements.
- 03Check missing facts
What is known? What still needs clarification?
- 04Prepare the draft
Connect the question with actual business facts.
- 05Meelis reviews
Check facts, pricing and promises to the customer.
Human decision - 06A human sends
The agent prepares. You send the final message.
Human decision
Workflow and review boundary
Task
The business owner assigns a bounded task.
Business context
Instructions and the knowledge vault relate the task to the business.
Read and prepare
Tools can check unread messages and prepare a summary or draft.
Human approval
Sending and external commitments require a separate decision. A prohibition in a prompt is not equivalent to a technical barrier covering every tool.
Verify
Read back the result and distinguish a real action from the agent’s claim that it completed the work.
What operation revealed
Model claims must match the actual configuration.
Documentation described different model routing from the active configuration. Privacy and cost assessments must use the actual destination of data rather than an old description.
A prototype is not an enforced safeguard.
A local policy pack existed, but the source review found it was not connected to tool dispatch. Its check could not be rerun in the inspected environment.
Masked identifiers do not make a conversation anonymous.
Log and identifier masking did not remove every personal detail from message bodies. Security descriptions must explain both coverage and exclusions.
My contribution
I defined operating practices, approval boundaries and business context. AI helped with research, drafts and configuration. This adapts an existing agent runtime; it is not a customer-support platform built from scratch.
Data boundaries
A local vault does not mean its used content stays on the server. Context can go to an external model service; email and messaging services have separate boundaries. This case does not establish that all data remains in Europe.
What the review checked
16 tests passed for a narrow identifier-masking function
Local source-review check, 11 September 2026. The tests covered platform identifier masking, not whole-assistant reliability or full message anonymisation. This was not a new test run performed here.
Does something similar keep coming up in your work?
Describe one task and what you want to review yourself. I will reply to say whether I can help and suggest a sensible first step. Sending a message does not place an order.
Tell me about your task →