Free every Thursday · 59,694 Readers
Fastlane Insider Newsletter: What's working for Shopify in the AI era?
Free every Thursday · 59,694 Readers
Fastlane Insider Newsletter: What's working for Shopify in the AI era?
The expensive part of a hand built Amazon report was never the 90 minutes. It was the stockout nobody saw for two days because the person building it was busy packing orders.
For two years, our daily Amazon report was the worst part of the morning. Someone logged into Seller Central, clicked through a handful of tabs, and rebuilt the same spreadsheet as the day before. It took up to 90 minutes and told us almost nothing we couldn’t have guessed.
So we handed the whole job to an AI skill. That is a less exciting sentence than it sounds, and the useful part of the story is everything around it: what the old report was really costing, what we changed about what it measures, the parts that broke, and the few jobs we refuse to automate. It started on Amazon, but if you sell on Shopify or across a few channels, you will recognize the feeling.
The daily Amazon report was costing us time, our best analyst’s attention, and at least one weekend of sales, and none of it showed up as a line item. Before getting to the fix, it is worth being honest about how bad the old way was.
Ninety minutes is the honest high end, and it is an estimate, because we only timed it properly once. Even on a slow day it rarely dipped below seventy. The work was pure transcription: read yesterday’s sales off one screen and type them into another so a human could finally look at them. By the time the report was readable, the sharpest part of the morning had gone and we still hadn’t decided a single thing.
For most of those two years the job fell to one of our analysts, who was far too sharp to spend an hour a day moving numbers around. We eventually handed it to a VA, which felt clever until we realized we had just made the same waste cheaper. The real loss was never the wage. It was an hour of genuine thinking traded away every morning, and on the days that analyst was out, the report simply didn’t get made.
Our worst miss came during a promo week when orders roughly doubled. The hand built report fell behind because the person running it was buried in packing and customer replies, and we didn’t spot a hero product running low until it had been gone for two days. A decent chunk of that weekend went with it.
That gap is easy to shrug off until you look at the figures. Across ecommerce, out of stock rates tend to run near 8%, and about 69% of shoppers will walk straight to a competitor when the thing they came for is gone. Katana’s study of 375 brands put the median cost of stockouts at roughly $21,000 a year, and peak weeks only sharpen it, which is what the numbers on how unpredictable peak demand has become keep showing.
A spreadsheet is a perfectly good tool right up until it isn’t. Ours held together when we shipped a few dozen orders a day from one channel. Once we added SKUs and a second storefront, it cracked in small, annoying ways: one stray row broke a formula nobody noticed for a week, and two people editing at once overwrote each other without warning. We were asking a frozen document to do a living job, and every manual paste was another chance to slip in a wrong number.
There was a timing tax on top of all that. Because the report only arrived late morning, anything it told us to do happened after lunch at best, so we lost most of a day of reaction time before anyone had read it.
We turned the Amazon report into a scheduled AI skill that pulls every channel at the variant level, adds two outside demand signals, and reports only the numbers that could change a decision that day. Each of those choices came from something the old report got wrong.
For us, an AI skill is a saved set of instructions that does the whole pull and writes the report up the way a person would. Think reliable assistant, not robot. The funny part is how ordinary this has become: by mid 2026, roughly 96% of online retailers were using AI somewhere, yet far fewer had built it into their daily work. The gap almost always comes down to setup, which is why it pays to treat the skill as one piece of a lean AI stack for a growing ecommerce team rather than a standalone gadget. If nobody on your team writes code, hiring someone to build a small tool around the data you already own beats bending a generic app until it sort of fits.
Fine detail is worthless if the detail is measuring the wrong thing. A report can split a product into ten neat lines and still mislead you, because what each line claims to measure isn’t what is actually being counted. In catalogs where two items look nearly identical on the shelf, that mismatch gets brutal, and supplement sellers tend to hit it early.
Take this Lion’s Mane collection by Nootropics Depot. The retailer sells an 8:1 dual extract and a 1:1 extract and positions them for different reasons to buy. Same mushroom on the shelf, two different shoppers. If your Amazon report rolls both into one “Lion’s Mane sold” line, you can’t tell which audience is showing up, and you won’t notice when only one of the two variants runs out.
The fix is to report at the variant level from the start, not the parent product level, so each version carries its own sales line and its own stock status. Any multichannel seller with ratios, sizes or flavors needs the same discipline, and the same unique per variant identifiers that make barcode inventory management on Shopify reliable.
For a long time the report only read Amazon, so every call we made about stock or pricing rested on maybe a third of the real picture. We taught the skill to read every channel we sell on and fold them into one view. That reconciling is genuinely fiddly: a 2026 inventory study pegged 72% of multichannel sellers as struggling to keep stock synced across channels, and we sat right in that group. Cross border orders made it harder still, since duties and taxes at checkout can turn a profitable looking order into a loss. The report now flags those gaps instead of hiding them inside one healthy revenue figure.
The old report could tell us what sold but never why demand moved, so we added two signals from outside any marketplace dashboard. One is how often AI assistants name our products when a shopper asks for a recommendation; you can check how visible your store is inside AI answers in about half an hour. The other is plain organic search. Watching where your products rank explains a sales wobble days before Amazon’s own charts admit it, and a hero product sliding down page one reads as a soft dip long before it turns into a cliff.
The trap with automation is thinking that because you can report everything, you should. Our first version carried something like forty metrics, and nobody read past the third chart. So we cut it back to the numbers that might change a decision that day. If no realistic value of a metric would make us act differently, it doesn’t belong in the daily report; it can live in a monthly review where it does no harm.
The newest line in our report barely existed a year ago: shopping that runs through AI agents that search, pick and pay for a customer, with the person only approving the final spend. Meta Muse AI agent payments arrived in September 2026, when Shopify switched on Shop Pay agentic checkout for Muse across Shopify stores and PayPal followed a day later. Fastlane’s breakdown of what Meta Muse means for Shopify merchants covers the store side of that change.
The detail that made us sit up: Amazon blocked Muse from shopping on its site, so an agent acting for a customer can’t complete that order on Amazon and has to buy somewhere else. That is a channel shift we now watch for. We flag any sign that agent driven demand is moving off our Amazon listings toward our Shopify store, where an agent can actually finish the job.
Once the Amazon report ran itself, the morning shrank to about five minutes of reading, our analyst moved from assembling numbers to acting on them, and a few urgent problems stopped waiting for the next day’s report.
A report that used to take up to ninety minutes now runs on a timer and is waiting before anyone opens a laptop. Human time dropped to roughly five minutes of reading, and yes, that five is an estimate too. The better change was what we did with the time we got back. Our analyst now reads the report and acts on it, and we started feeding its history into reorder planning, the same reorder point and safety stock math that breaks down when brands keep eyeballing inventory. Reordering turned from a panicked reaction into something closer to a plan.
What genuinely scares people about automating a report is the worry that the software will start doing things, not just noticing them. It is a fair worry, and the field that has thought hardest about where to draw that line is security, because a wrong automated move there can wreck a company in minutes.
You can see the model clearly in how a modern agentic security operations centre is built. The AI agents there do a lot: they enrich alerts with context, write up case summaries, map suspicious activity against known attack patterns, and run constant health checks on sensors. What they are not allowed to do is pull the trigger. Isolating a machine or shutting down an account stays with a human analyst, and the work shifts from reacting to every alert toward ranking everything by actual risk first.
We copied that split almost exactly. Our skill reads the data, then flags and ranks whatever looks off. A person still decides whether to drop a price, place a reorder or lean on a supplier. Drawing that line on purpose is the reason we trust the report instead of nervously rechecking it.
Most reporting runs on a comfortable daily beat, and for most numbers that is fine. A small number of problems get worse by the hour, though, and a tidy daily digest is useless for those. A suppressed listing or a price that glitches to a loss can’t wait for 8 am tomorrow. So we lifted those few urgent events out of the daily batch and set the skill to ping us the moment it spots them. The daily report keeps everything that can safely wait; the urgent handful never waits in it.
A report only earns its keep when every finding becomes a tracked job with an owner, and getting there meant fixing three failures and deciding which jobs stay human. A report nobody acts on is just a prettier spreadsheet.
The usual way a finding dies is depressingly familiar: someone drops it into a chat channel, three people react with a thumbs up, and it is forgotten by lunch. Teams that avoid this treat every finding as a tracked job with a name and a deadline, never as a passing message. In our setup, a late supplier becomes a procurement ticket with an owner, and a suppressed listing becomes a job for whoever runs that account. Nobody has to remember anything from a morning message, because the work already exists in the system with someone’s name on it.
Plenty broke early, and some of it rattled us. A bundle SKU got counted twice for a week and turned an ordinary product into a fake bestseller. A timezone slip stamped a day’s numbers onto the wrong date, so a busy Monday read back as a dead Sunday. The one that scared us most was when the skill stated a number with total confidence and it was simply false. A blank, you question. A confident wrong number, you act on.
We fixed all three, and the lessons matter more than the bugs. Check every total against one trusted source for the first two weeks. Lock the whole report to a single timezone before you trust any date. Treat a confident wrong number as more dangerous than a missing one. And keep a human eye on anything that makes real money move.
Automating a report is easy to feel good about; automating a human moment is where we would ask you to stop. The skill will happily tell us a supplier is late or a listing has been pulled. It will never write the message to a supplier we have worked with for a decade, and it never replies to a customer who is angry about a late parcel. Those moments stay with a person, and so does any call where a wrong automated decision carries real financial or legal weight. The report’s job is to get a person to the right conversation faster, not to fake the conversation for them.
A $500K seller should automate the one report they already build by hand, while a $5M seller should first find and close the most expensive data blind spot across channels. None of this needs doing all at once.
At this size you don’t need an operations overhaul, and chasing one will hurt you. You need the single report you already live by to stop eating your mornings. One report you trust completely beats five half finished ones. The order that worked for the smaller sellers we have compared notes with looks like this:
At this size the daily report is rarely the real problem; data scattered across channels almost always is. Your first move isn’t automation for its own sake. It is finding the single most expensive blind spot and sealing it before you build anything shiny on top, which looks less like tidying and more like triage:
Whatever your revenue, the first step is the same, and it is gloriously low tech. Spend a week writing down exactly what the person making the report does: every click, glance and small judgment call they make on autopilot. That document basically is the skill, because automation only works once you can describe the job precisely, and almost nobody has written theirs down. It is the same habit as honestly reading your store’s commerce readiness score and fixing what matters first. Get the description right and the technology becomes the easy part. Skip it, and you will automate a process nobody fully understood, which is just a faster way to be wrong.
Handing our Amazon report to an AI skill didn’t make us a tech company. It gave us back a saner morning and freed our sharpest people to think instead of type. The real win had little to do with the AI itself. It came from understanding our own process well enough to explain it, reporting at the unit where stockouts actually happen, and keeping human hands on the decisions that move money or relationships. If your team is still rebuilding yesterday’s numbers every morning, the week spent writing down how that report gets made is the place to begin.
Automate your daily Amazon report by documenting the manual process first, then turning those steps into saved instructions that run on a schedule. Spend a week writing down every click, metric and judgment call the person building the report makes. Then set up an AI tool or a small custom script to pull the same data from Seller Central and your other channels, report it at the variant level, and deliver it before the team starts work. For the first two weeks, check every total against one trusted source and lock the report to a single timezone, because double counted bundles and date slips are common early failures.
A daily Amazon report should include only the numbers that could change a decision that day. For most sellers that means sales and stock status for each variant, any listing that has been suppressed, prices that have moved outside an expected range, and items close to running out. Multichannel sellers should fold in every channel so stock and revenue reflect the whole business. Vanity metrics such as long trend charts belong in a weekly or monthly review instead. A first version with dozens of metrics usually goes unread past the first few charts.
It is safest to let AI monitor and flag problems in Seller Central while a person makes the changes. A reporting skill can read data, rank issues by risk and alert you instantly to a suppressed listing or a broken price. Decisions that move money, such as price changes, reorders and supplier messages, should stay with a human who reviews the flag first. This mirrors how agentic security teams work, where AI agents enrich and rank alerts but a human analyst decides on any action. It also protects you from an AI stating a wrong number with confidence.
Report at the variant level because a product total can hide a stockout or a demand shift in one version of the product. If two sizes, flavors or formulations roll into one sales line, you cannot see which one shoppers are buying, and you will not notice when only one of them runs out while the total still looks healthy. Giving each variant its own sales and stock line lets you reorder the right item, spot which audience is growing, and catch problems days earlier. This matters most for catalogs where similar items look almost identical on the shelf.
