Trengo's inbox now takes requests from Claude and ChatGPT. A Monday morning in five requests, and where to slow down

What Trengo switched on

Trengo is a shared inbox for customer messages: email, WhatsApp, website chat and social messages, handled by a team in one place. On October 29, Trengo made its MCP server live for every customer. MCP is the standard that lets AI tools like Claude and ChatGPT connect to other software and act in it.

In practice, a support lead can now open Claude or ChatGPT, connect Trengo, and ask for things in plain language. Trengo lists 88 supported actions. The pattern is the same each time: you describe a goal, the AI proposes actions, you review them, and you approve.

That last step is the whole game. Asking is free. Approving is the job. Some of these requests only read your inbox. Others change it for everyone, or send something to a customer. To make that concrete, here's one illustrative Monday morning at a small online shop, built from the examples Trengo itself gives, with a note on where to slow down.

8:30. "What happened over the weekend?"

The support lead starts with a briefing: what happened over the weekend, and which conversations need attention.

This is the easy one. It reads conversations and summarises them. Nothing changes in the inbox. If the summary misses something, the cost is that a person finds it a little later, which is what would have happened anyway.

Two habits still help. Ask for links to the conversations it highlights, so you can open the originals rather than trusting a paraphrase of an angry customer. And remember that the summary goes into whichever AI tool you're using, so check that your company is comfortable with customer messages passing through it. That's a policy question to settle once, not every Monday.

9:00. The AI agent's gaps, then the unassigned pile

Next, a request Trengo uses as an example: find what the shop's AI agent couldn't answer this week, and help fix the top three.

The first half is reading, and it's useful: a list of gaps in your help content is exactly what you want on a Monday. The second half, "help me fix," is where you read more slowly. Fixing might mean changing the knowledge your AI agent uses to reply to customers. A wrong fix here doesn't stay in the inbox. It goes out to every customer who asks the same question this week.

So read each proposed change as if you were approving a new paragraph for your returns policy, because that's roughly what it is. If the answer involves money, refunds or delivery promises, check it against the actual policy before approving.

At 9:30 comes the cleanup: find unassigned delivery questions and assign them to the support team.

This is a bulk write, and bulk writes are where small errors grow. Before approving, look at the list of conversations the AI proposes to move. Does it include the wholesale customer asking about a delivery of fifty units, which should go to sales? A complaint that's really a refund request? A good proposal shows you the list. If it doesn't, ask for it before you approve.

If your team has rules about who handles what, write them into the request: "unassigned delivery questions from individual customers, not trade accounts." The AI can only follow the rules you give it. And approve in batches you can actually read. Twenty conversations you've looked at are better than two hundred you haven't, and the rest can wait for the next request.

10:00. "Build a Journey that routes delivery and product questions"

Later, the lead asks for something more lasting: build a Journey, Trengo's routing automation, that sends delivery questions to support and product questions to sales.

This is the request to slow down on most. A Journey keeps running after you close the chat. If it misroutes a kind of message, it'll do so all week, quietly, for everyone. We've written that a shared inbox without a route is just a pile; a wrong route can be worse, because everyone assumes the pile has been sorted.

Approve the Journey in a draft or limited state if you can, check where the first real messages land, and only then rely on it. Tell the team it changed. And make sure someone other than the person who built it knows it exists.

11:00. Customer-facing messages, and the Shopify combination

By late morning: create a WhatsApp template for a shipping update and send it to a customer. Trengo also describes combining its MCP with Shopify's, so the AI can find delivery complaints in Trengo, check the matching Shopify orders, and draft replies with tracking.

Drafting is a good use. Checking an order and pulling the tracking number is tedious and easy to get wrong by hand, and on a busy Monday it's the kind of job that waits until the afternoon. An AI that pulls the order, finds the tracking and writes a first draft saves real time, as long as a person reads the result. Sending is different. Read every message that goes to a customer before it goes, at least until you've seen enough to trust a particular kind of request. A reply with the wrong tracking number, sent with confidence, creates two more conversations.

Setting it up so Mondays stay boring

A few decisions make all of this safer, and they're quick to make now.

Decide who connects Trengo to an AI tool, and with whose account. The AI acts through that connection, so it should belong to someone whose permissions match what you want done. Decide which requests anyone can make and which need a lead to approve: reading for everyone, bulk changes, routing and customer messages for named people. And keep a simple log of what was approved, by whom, and why. We made the same point about putting a ticket id on every write an agent makes, and in our review of Resend's inboxes.

If you'd like help setting up AI tools on your support inbox with the right approvals, or building an agent that works safely across your inbox and your store, that's the work our AI applications and agents team does. Talk to us before the first Monday.

Related work: Sutton Place.