Monday night: I almost lost client data to an AI
At ten o’clock on Monday night, I was sorting client contracts in my Hangzhou rental apartment. The customer-support assistant, spreadsheet assistant, and email assistant were all running: one changed a client’s contact details incorrectly, and another nearly cleared out an old backup. I unplugged everything and restored the files manually. That was when I realized that one AI working alone and several AIs working together are not “more people moving faster.” They are more paths for mistakes to happen at once. I got stuck here too: I originally thought opening a few assistants would save time, only to create rework for myself.
What multi-agent collaboration is, and who is using it
Simply put, it lets multiple AIs divide the work: one reads client emails, another organizes spreadsheets, another drafts replies, and a final assistant compiles everything. It is more like a temporary project team than a smarter chat box. Independent consultant Lin Xiao used this setup for branding projects from a café in Chengdu. She had a “researcher” collect public information, a “writer” organize the first draft, while she handled verification and sending. She later added a rule: any action involving emailing clients, deleting files, or changing prices had to stop and wait for her confirmation. The point of her example was not full automation, but who has the final say. Security incidents involving AI teams have already appeared in the news, and coding assistants remain inconsistent in real-world task tests.
What it costs to replicate today—and the permission boundary to draw first
For testing ready-made products, I would budget RMB 0–200 per month, with roughly 2–4 hours for learning and initial setup. The technical barrier is not high, but I need to understand what each step can read or change; if I cannot follow what happens in the backend, I would not connect it to a client system. The first step is simple: open the AI tool I already use, find the “tool connections, permissions, automation, or workflows” settings, connect only a test file with no sensitive information, and create a workflow that asks for my approval before “sending, deleting, paying, or publishing.” I initially went wrong by connecting a live client spreadsheet, then switched to fictional clients and recoverable copies. Not everyone needs these tools; if my business is not stable enough to justify dividing the work, there is no rush to test it.
Match the setup to my stage—there is no need for an AI team on day one
If I am just starting: I would use one AI assistant for one low-risk task, such as turning public research into summaries, and run it three times before considering collaboration. That costs less and makes it easier to spot where it goes wrong.
If I already have one or two clients: I would create a separate “workspace” for client data, require human confirmation before sending or deleting anything, and test with pseudonyms for a week. The point is not to show off; it is to make sure clients know which steps involve AI and who owns the final check.
If I am scaling: I would turn permissions, backups, and failure alerts into a fixed process, then add AI roles gradually while keeping one person responsible for spot-checking outputs. I would also invite others to start with one sanitized test spreadsheet, walk through the process once, and only then decide whether to bring in live business data.