I was often missing App Store reviews I wanted to reply to. I have a small iOS app on the App Store that I built in college. Between my daily work building software and running the business, checking for new app reviews was easy to put off. A customer would take the time to share feedback, and I wouldn’t always respond as quickly as I wanted to.
So I built a simple AI customer service agent for my app. It checks for new reviews, prepares replies based on how I’ve responded in the past, and brings the drafts to me for approval. Once I approve a reply, it can submit it to the App Store.
This simple agent only took me a Saturday afternoon to setup and now I’m on top of all of my reviews. I never miss a review and the responses to common questions or feedback are already written. I only have to dive in on responses for unique questions or helping the customer with something specific.
Here’s how I built it and why this small workflow has been so useful.
The basic customer service task
The workflow I wanted was straightforward:
- Regularly check for new App Store reviews.
- Find the ones that still need a response.
- Write a draft that sounds like me and addresses the actual feedback.
- Let me edit or approve it.
- Submit the approved reply to the App Store.
The biggest help is fetching the reviews that need to be replied to. I do this every week to stay on top of them. Having the replies pre-written in my voice made it easy to reply to each one since I don’t have to write them from scratch and a lot of them have recurring themes.
It was very important to me to keep an approval step in the design of the workflow. I want to see what we’re telling a customer, especially when they’re reporting a problem or asking for help.
While AI agents are very smart and do a good job, I wanted to make sure that responses to my customers were accurate and not AI slop. It is perfectly fine to use AI, but if I am going to put my name behind it, ultimately I am responsible for what the agent writes.
The architecture
The agent was built as a private plugin for Codex, with a small Python tool that connects to Apple’s App Store Connect API. A local SQLite database keeps track of the work between runs.
The flow is:
Scheduled check → fetch reviews → update the queue → draft replies → my approval → submit to Apple
Each piece has a clear job:
| Component | What it does |
|---|---|
| Scheduled task | Starts the recurring review check and brings actionable work back to me. |
| App Store Connect integration | Fetches reviews and existing responses, then submits replies after approval. |
| Local SQLite database | Stores review snapshots, drafts, approval versions, and submission history. |
| Codex and writing instructions | Use the review, previous replies, and app information to prepare a relevant response. |
| Approval and submission checks | Make sure the reply being sent is the version I approved and check for changes before sending. |
Codex handles interfacing with me over chat. The Python tool handles the API calls and the data that need to be consistent from one run to the next.
In this implementation, the integration and database live on my Mac, while Codex provides the AI drafting. We didn’t need to build a separate hosted application or connect a separate model API key for the plugin.
Collecting reviews without losing track of them
The tool connects to App Store Connect and fetches all available pages of reviews, along with responses that have already been submitted.
That matters because checking only the newest page could leave older unanswered reviews behind. The first import includes the backlog, and later checks update the same saved queue.
The database allows the agent to distinguish an unanswered review from one with a draft waiting for me, or one I’ve already responded to. Existing drafts stay in place instead of being rewritten every time the task runs.
Writing replies that sound like me
I gave the agent examples of my previous replies and a short set of writing instructions.
A positive review may only need a sentence or two saying thank you. A frustrated customer needs to be acknowledged and provided with next steps to resolve. When I’m personally offering help, the reply should sound like it comes from me.
The agent also needs boundaries around product information. It shouldn’t invent a fix, promise a release date, or assume that troubleshooting advice from an old reply still applies. Previous responses are writing examples; technical advice needs current, verified support information. If something is uncertain, the agent can leave an internal note for me to resolve.
Approving the exact reply that gets sent
The scheduled task collects reviews and writes drafts. It doesn’t approve or publish them on its own.
I see the customer review alongside the proposed response, then approve it or ask for an edit.
After Apple accepts the submission, the tool records that review as replied to.
Using AI to build the workflow
AI helped implement the agent as well as write the customer replies. I started with the process I wanted in plain language, then used Codex to build the App Store Connect integration, the saved queue, and the approval workflow around it.
Packaging the tool and instructions together means I can reuse the workflow without explaining the whole process again. This makes the process so much more efficient than writing a prompt to describe the task and copy and pasting data into a chatbot everytime I want to reply to reviews.
I now have a scheduled process that can run regularly and take care of these reviews in the background.
What changed for me and my customers
I know I’m not the first guy to build an AI customer service agent in 2026, but this small agent has been super practical and easy to implement for any business that needs to manage reviews.
The biggest improvement is consistency. Reviews used to depend on me remembering to go look for them. Now the workflow brings unanswered reviews to my attention with a good-to-go response already written 9/10 times.
I’m saving the time I used to spend finding reviews and writing the response.
My customers benefit because I see their feedback immediately now. It also makes it easier to respond promptly when a customer has a question, needs help, or simply took the time to say something nice about my app.