That morning, I realized the question was no longer “Should we use AI?” but “Will it keep working for the company?”

At 9 a.m. Thursday, I was revising a client proposal in a Hangzhou café when I saw United restricting unauthorized AI agents from booking tickets, Databricks launching a persistent business assistant, and a Federal Reserve official warning about AI investment risks. I felt a little uneasy: if AI no longer just answers questions but monitors workflows continuously, could small teams’ roles and ways of collaborating change quickly?

Among these updates, small teams should pay closer attention to persistent business assistants

This isn’t a full review of one product; it’s a set of industry shifts. In simple terms, a business assistant keeps handling sales materials, customer questions, or internal information over time, without us opening a fresh chat window and re-explaining the background every time.

That reminded me of Xiaoxiao Chen, a foreign trade consultant in Shanghai. Last year, she had her team use AI to draft replies to repetitive inquiries, then reviewed them herself. At the time, she also worried clients would think the replies “didn’t sound human,” so she kept manual approval and signature steps. Today’s changes suggest that automation is moving from one-off question-and-answer sessions into continuously running parts of the work.

Calculate today’s replication cost by information boundaries first

To experience the thinking behind a persistent assistant, I’d start with a low-risk trial: roughly RMB 0–100 per month, with about 2–3 hours of setup. The technical barrier is knowing which client materials the machine can safely handle. The first step is creating a “customer question drafts” assistant in the AI tool I already use, letting it organize and draft only, without permission to send messages, make payments, or edit contracts.

I’d document permissions, the scope of accessible data, and where human approval is required. Not everyone needs this, and there’s no need to chase every news item immediately.

Here’s how I’d sequence it based on where my work stands

If I’m just getting started, I’d keep the manual workflow and test one repetitive question each week rather than rebuilding the whole operation. If I already have 1–2 clients, I’d use an assistant to organize materials and draft replies, while I still approve every message and track how often it gets something wrong. If I’m scaling, I’d first map which workflows can be automated, then decide whether to connect a long-running assistant, while preserving clear permissions, backups, and accountability.

Industry news can help me spot direction, but the things I need to protect first are client trust and cash flow. It’s fine not to try this now; keeping my own workflow stable matters more.