Skip to main content
RoutMe
All posts
Playbooks5 min read

Replace the Zap: a where-is-my-order automation, no glue code

Build a where-is-my-order WhatsApp flow without Zapier or glue code. Trigger, order lookup tool, reply, and a fallback to a person, step by step in RoutMe Flows.

VR

Vinita Rathi

Product and conversation design

A where-is-my-order automation in RoutMe has four parts: a message received trigger, an agent step that calls a read-only order lookup tool, a reply that states the real status, and a fallback that hands the conversation to a person when the order cannot be found. There is no Zap, no webhook relay and no glue code between them.

Order status is the most common question a store receives, which is why it is the first thing most teams automate and the first Zap they regret maintaining. This post walks through building the flow, naming the decisions that matter, and being straight about what the flow will not do. The conceptual background is in WISMO is a lookup problem.

What the old Zap looked like

A common version of this automation catches an incoming message through a WhatsApp provider, extracts an order number with a model call, queries the store, formats a reply, and posts it back through the provider. Each hop is a separate step with its own failure modes. If the store lookup times out, the customer receives nothing, and the only evidence is a failed run in a dashboard you check when someone complains.

Step 1: Connect the data

The flow needs one capability: read an order and its tracking state. Expose it as a tool. For a Shopify store the pattern is a small MCP server that calls the store's own APIs with a read-only token, as described on the Shopify page. RoutMe has no Shopify app and no native connector, so say so to yourself before you plan. For other systems the tool can wrap an API or a read-only database view.

In RoutMe you add the server URL and an optional bearer token, press Test connection, and see the tools it reports. Name the tool for what it does, such as get order status, and give it a description the model can understand. Keep the returned fields to the minimum a customer needs: status, items, carrier, tracking reference.

Step 2: Start the flow

Create a flow. You can begin from the ready-made order status flow, describe it in plain words and review the draft, or build it step by step. New flows are created paused, so nothing sees live traffic until you activate it. Use the message received trigger. If you want only order questions, set a keyword such as order, although an agent that can recognise the intent is usually better than a keyword.

Step 3: Add the agent step

Add an agent step and write its instructions in plain language. Be specific about three things.

  • What it may do: look up an order with the attached tool and reply with the status.
  • What it must not do: promise delivery dates the data does not contain, discuss refunds or change anything.
  • How to identify the customer: ask for the order number or the email used on the order, and look up only that.

Define outcomes: found, not found and needs a person. Attach the MCP server. The agent works for up to six rounds, picks an outcome and the run follows that edge. Wire the default edge to a handoff, since it is where the run goes if the model ends without choosing.

Step 4: Reply and branch

The agent can send the answer itself with send text, which is allowed because the customer's message opened the 24-hour window. Alternatively route the found outcome to a send text step that uses placeholders from earlier data. For not found, add a send text asking for the order number again, and after a second miss route to the assign step so a teammate takes over. Two failed identification attempts is a sensible limit.

Step 5: Test it

  1. 1Connect a test number and activate the flow.
  2. 2Send a valid order number and check the reply against the real order.
  3. 3Send an unknown number and check that it asks again rather than guessing.
  4. 4Send something unrelated, such as a refund request, and check that it hands off.
  5. 5Open the run trace for each message and read which nodes ran.

Step 6: Add the proactive version

The reactive flow answers when asked. The proactive one tells customers before they ask. Add a second flow with an external event trigger named order shipped. Your store or your own back-office tooling posts that event with the customer's WhatsApp number and a data object containing the order number and carrier, and the flow sends an approved utility template. Because the customer has not written to you recently, it has to be a template, and it should be categorised as utility, since miscategorising it as marketing is the usual cause of rejections. See WhatsApp message templates and approval.

What this flow does not do

  • It does not edit, cancel or refund orders. Those need a person or a deliberate, separate design.
  • It is not a connector to thousands of apps. It reads the systems you expose as tools.
  • It does not run on a schedule. It starts from a message or an event.
  • It is only as correct as the system it reads. If the tracking data is wrong, so is the answer.

For the measurement side, how to reduce WISMO tickets explains what to track, and how an AI agent looks up order status safely covers the safety side. If you are leaving a Zapier setup, start with do you still need Zapier. See pricing on the pricing page, the security page for data handling, or try it in a demo.

Writing the instructions

Here is the shape of instructions that works well, described rather than quoted. Tell the agent its role in one sentence. Tell it the single lookup tool it can use and when to use it. Tell it to answer in the customer's language and keep replies short. Tell it what to do when the tool returns nothing, which is to ask once more and then hand off. Finally tell it the things it must never do, such as inventing a delivery date or discussing a refund.

Resist the urge to paste your whole policy manual in. Long instructions make behaviour harder to predict and harder to test. Keep policy that must be exact, such as a return window, in a deterministic If step, and keep judgment in the agent.

What to watch after launch

  • The share of order questions that end in a handoff, and why.
  • Runs where the lookup tool failed, since those are customers who got an apology instead of an answer.
  • Questions the flow did not expect, which tell you what to add next.
  • Whether replies match the source data, checked by sampling a few each week.
  • Migration
  • Zapier
  • WISMO
  • Flows
VR

Vinita Rathi

Product and conversation design

Writes about where automation belongs in a support conversation, and where it does not.

Frequently asked questions

Does the flow need a template?

Not for replies. When the customer messages you, a service window opens and the agent can answer with free-form text. A template is needed only to message a customer first, such as a proactive shipping update.

What if the order is not found?

Give the agent an outcome for not found and route it to a clarifying question or a person. The agent should never guess an order number.

Can it change or cancel the order?

Not unless you deliberately build that in. The lookup tool should be read-only, which means the flow answers questions and hands changes to a person.

See it on your own data

Connect a system, drop an agent node into a flow, and watch it answer from what is actually true in your systems right now.