A seven-figure Shopify brand run by a team of four was spending roughly 14 hours every week on tasks that felt suspiciously like busy work: pulling order data, writing customer follow-up emails, updating inventory spreadsheets, and compiling a Monday morning report nobody actually enjoyed reading. Over six weeks, they rebuilt that entire workflow using Make.com, Claude by Anthropic, and Airtable. The result was not perfect—but it was dramatic enough to be worth documenting in detail.
The Problem Was Specific, Not Generic
Before anything else, the team did something most businesses skip: they tracked exactly where the hours were going. Not “roughly” or “around.” Exactly. They used a simple Toggl timer for two weeks and got a number that surprised even them.
- 4.5 hours/week writing post-purchase follow-up emails for orders flagged as delayed
- 3 hours/week manually copying Shopify order data into an Airtable base for the ops manager
- 2.5 hours/week pulling inventory numbers from three supplier portals and dumping them into a Google Sheet
- 4 hours/week building the Monday ops report, which was essentially a manual aggregation of data already sitting in other tools
Every one of these was automatable. None of them required a human making judgment calls. That distinction matters—these were data-moving and template-filling tasks, not creative problem-solving.
Step One: Shopify to Airtable, Automated
The first win came from stopping the manual copy-paste between Shopify and Airtable. This is a solved problem, and Make.com solves it cleanly. Their Shopify module triggers on new orders or order status changes, pulls the relevant fields, and pushes them into a specified Airtable base.
Setup took about 90 minutes the first time—mostly because the team had inconsistent field naming between the two platforms and had to do some mapping work. Make.com’s interface is visual and drag-and-drop, so you can see the data flowing from one node to the next. That said, Make.com is not as beginner-friendly as Zapier. If you have never used a webhook before, budget extra time. Pricing starts at around $9/month for the Core plan, which gives you 10,000 operations per month—enough for a brand doing a few hundred orders weekly.
The end result: Airtable updated automatically within 2 minutes of any Shopify order status change. Three hours a week, gone.
Step Two: AI-Written Delay Emails That Don’t Sound Like Robots
This one took longer to get right and is worth explaining in detail because the obvious approach fails.
The obvious approach is to write a single email template and auto-send it to every delayed order. Every customer gets “Hi [Name], we wanted to let you know your order is running behind…” That reads like a form letter, and customers know it. It also triggers more reply volume because it feels impersonal.
Instead, the team used Claude (via Anthropic’s API, accessed inside a Make.com HTTP module) to generate a short, personalized paragraph based on four variables pulled from the order: the customer’s first name, what they ordered, how many days delayed, and whether this was their first order or a repeat purchase. The prompt took about a week of iteration to nail down. First-time customers got a slightly warmer, more apologetic tone. Repeat customers got a tone that assumed familiarity, like a quick heads-up from a brand they already trusted.
Claude’s API pricing runs roughly $3 per million input tokens on the Haiku model, which is more than fast enough for email copy. For a brand doing maybe 60 delayed orders a week, the cost of generating those emails with Claude is essentially rounding error—under $2/month.
One real limitation of this setup: Claude occasionally writes emails that are a little too wordy. The team added a hard constraint in the system prompt—”respond in no more than three sentences”—and that fixed about 80% of the verbosity issues. The remaining 20% gets caught by a Slack notification that flags any email over 80 words for a quick human review before it sends. That human review takes maybe 10 minutes a week. They kept it intentionally. Some mistakes are cheap to catch manually and expensive to let through.
If you’re thinking about the broader picture of how AI agents handle tasks like this versus just triggering actions, the distinction between AI agents and chatbots is worth understanding before you build anything more complex.
Step Three: Supplier Inventory, Scraped and Structured
This was the messiest piece. Two of the three supplier portals had no API. One had a CSV export, one had a basic REST API, and one required a manual login to a web interface that looked like it was designed in 2008.
For the CSV portal, Make.com handled it via an email parsing trigger—the supplier sent a weekly inventory update by email, and Make grabbed the attachment, parsed the CSV, and wrote the relevant rows to a Google Sheet. Straightforward.
For the REST API, a direct Make.com HTTP module call did the job.
For the web portal with no export option: they used Bardeen, a browser automation tool that runs in Chrome. Bardeen can log into a web page, navigate to a specific table, extract data, and push it somewhere else. It is not cheap at scale—the Pro plan runs $10/month per user—and it is fragile when suppliers update their UI, which happened once during the six-week pilot and broke the automation for three days. That brittleness is real and worth planning for. They set a monthly calendar reminder to manually verify the Bardeen workflow is still running correctly.
Combined, the inventory piece now runs every Friday at 6 AM automatically. The Google Sheet is current before anyone starts work. The 2.5 hours disappeared.
Step Four: The Monday Report Writes Itself
The Monday ops report was a four-page Google Doc that someone assembled by hand every Sunday evening. Revenue for the week. Top SKUs. Delayed order count. Inventory flags. Customer service ticket volume. It was always late, always slightly wrong, and everyone complained about it while also not wanting to be the one who stopped doing it.
The new version runs as a Make.com scenario at 7 PM every Sunday. It pulls aggregated data from Airtable (orders), the Google Sheet (inventory), and Gorgias (their helpdesk, which has a Make module). Claude takes that structured data dump and generates a plain-English narrative—not bullet points, actual prose—summarizing what happened that week and flagging anything that looks anomalous. The output goes into a Google Doc automatically.
Is the narrative perfect? No. About once a month, Claude interprets a dip in ticket volume as “improvement in customer satisfaction” when it was actually a tracking system glitch. The team added a disclaimer at the top of every generated report: “Auto-generated. Review anomalies manually before acting.” That is not a failure of the system—it is honest documentation of what the system is. For a deeper look at how these kinds of reports can be structured, automating weekly business reports with AI covers the general architecture well.
What the Numbers Actually Looked Like After Six Weeks
Fourteen hours claimed back per week. In practice, the team found they were spending about two of those hours on maintenance, monitoring, and the occasional manual review. Net recovery: around 12 hours per week. At an average fully-loaded hourly cost of $35 for the team members involved, that is $420/week or roughly $21,840 per year in recovered time.
Total monthly tool cost for the automation stack: Make.com Core at $9, Claude API usage at approximately $4, Bardeen Pro at $10, Airtable Plus (already in use) at $10. About $33/month. The ROI math is embarrassingly lopsided.
The less quantifiable win: the team stopped dreading Sunday evenings. That matters. Burned-out teams in small businesses are an underrated operational risk.
What Did Not Work (and Why)
They tried automating customer service responses to common questions—refund requests, “where’s my order” messages—using a Claude-powered draft that would appear in Gorgias for an agent to approve. The agents hated it. Not because the responses were bad, but because reviewing AI drafts felt like more cognitive work than just writing a response from scratch. They killed that workflow after two weeks.
This is something the myths around AI customer service rarely address honestly: AI assistance does not always reduce workload. Sometimes it adds a new kind of friction. The tasks that automated cleanly were all pure data-moving tasks with no emotional or judgment component. The moment the work required reading a frustrated customer’s tone and responding appropriately, humans did it faster and better.
My Honest Take on Where to Start
If you are running a small e-commerce operation and considering something similar, do the Toggl exercise first. Two weeks, track every operational task, log the hours. You will almost certainly find two or three tasks that are genuinely automatable—pure data movement, templated outputs, scheduled aggregations. Start there. Do not start with customer-facing AI until you have built confidence with back-office automation that nobody sees if it goes slightly wrong.
Make.com is the right hub for orchestration if you are comfortable with slightly technical interfaces. If you want something simpler, comparing Zapier, Make, and n8n is worth reading before you commit. For LLM work inside workflows, Claude Haiku is fast and cheap enough that you should default to it for anything that needs to generate text at volume. GPT-4o is better for nuanced reasoning but noticeably more expensive per call at scale.
And build in the human checkpoints deliberately. The goal is not zero human involvement. The goal is to eliminate the tasks that require zero judgment from the humans who have plenty of it.
FAQ
Do you need a developer to build this kind of automation stack?
Not necessarily, but some technical comfort helps. Make.com and Airtable are genuinely no-code for most use cases. The Claude API integration via an HTTP module in Make does require you to understand how to format a JSON request, which is a small but real technical hurdle. A developer can set it up in an hour if you get stuck, and many freelancers on Upwork do this kind of work for $50–$150 per workflow.
What happens when the automation breaks?
It will break, and you need a plan before that happens. Make.com sends email alerts on failed scenario runs by default—turn those on immediately. For anything customer-facing, build in a Slack notification fallback so a human knows to intervene. Bardeen in particular requires monthly spot-checks because it depends on website structure that changes without notice.
Is the data flowing through Claude secure enough for business use?
Anthropic’s API does not use your submitted data to train models by default on paid plans, but you should read their data processing agreement directly rather than trusting a summary. For order data that includes customer names and email addresses, consider whether you need to anonymize personally identifiable information before it hits any external API. There is a lot of misinformation about this topic—understanding the real risks around AI data privacy will give you a clearer picture than most vendor documentation.
This article was produced with the assistance of AI, and its featured image was AI-generated. We review for accuracy, but please verify critical details.
