BIONIC BUSINESS BY SAMUEL WOODS
An online store gets the most from AI agents when each rule it runs on is a file the agent reads and a check the agent can't get past: what you may claim, what's in stock, and what you'll offer to keep a customer. With those in place, an agent can write product pages, answer shoppers and handle cancellations.
I'm Sam Woods, and I publish Bionic Business, a newsletter that thousands of online entrepreneurs read. This page is for people who run an online store. It has seven cases from stores I work with, and the three agents I'd build first. Last updated November 2026.
Every figure on this page came from the store it describes. The last column of each table says what the figure leaves open.
Three that worked:
| Store | What the agent does | Build cost | Result | What the figure leaves open |
|---|---|---|---|---|
| Skincare store, $620K a year | Writes product descriptions and checks each one against a list of claims the store may not make | Under a week, plus about $2,800 of legal work | 19 of 212 descriptions were blocked before anyone saw them, and conversion on 38 rewritten pages rose 14% | The 14% is a before-and-after, and three pages also got new photos |
| Coffee subscription, $890K, 3,100 subscribers | Answers each cancellation request with a change before any discount | About two weeks | Saves went from 11% of cancellation requests to 27% | Saved subscribers may leave sooner than earlier ones did |
| Store, $1.8M | Recommends products on the site, and checks live stock before it names one | About four days of a developer's time | None of 2,640 recommendations named an unavailable product | Sessions that used it converted at 3.4 times the rate, and those shoppers were already closer to buying |
Four that went wrong, or nearly did:
| Store | What the agent did | What it cost | What the figure leaves open |
|---|---|---|---|
| Supplements brand, $3.1M | Wrote product and ad copy with no list of prohibited claims to check against | One prohibited claim ran in paid social for 11 days with $6,200 behind it | The regulatory question is still open |
| Outdoor-gear store, $4.3M | A gift finder read a stock file exported once a night | Nothing, because a one-day test came first. 9 of 40 recommendations named a product that was sold out or in stock only in the wrong size | A normal September weekday, before the busy season |
| Home goods seller on Amazon and Etsy, $1.9M | Wrote listings and buyer messages with none of the platforms' rules | A policy warning from Amazon, and three listings lost placement | Three listings is too few to prove the titles caused it |
| Ecommerce brand, $2.6M | Sent order and cart texts over Black Friday weekend before its test was finished | An estimated $52,000 to $71,000 in lost text revenue | The estimate is modeled against the year before |
A skincare store at $620K a year has an agent write its product descriptions. Every draft runs against a list of claims the store may not make, a blocklist of 140 terms and a sheet of approved benefits written with the store's lawyer.
Across 212 descriptions and variants, 19 were blocked before any person saw them, and no prohibited claim reached the site. The founder's copy review went from about five hours a week to under one. Conversion on the 38 rewritten pages rose 14% against their own previous six weeks. That's a before-and-after on the same pages, and three of them also got new photography.
The legal work cost about $2,800, and the build took under a week.
A supplements brand at $3.1M a year, in the same kind of regulated category, had an agent writing product descriptions and ad variants. A review of 48 variants found nine with claims its lawyer had prohibited. Four were already published, and one ran in paid social for 11 days with $6,200 behind it.
The list of prohibited claims was a PDF in a 2024 legal folder, and the person approving the copy had never seen it. The agent had learned from the brand's old best-performing copy, which included those claims. The regulatory question is still open.
A coffee subscription at $890K a year, with 3,100 subscribers, had 412 cancellation requests in a little over four months. An agent reads each one, sorts the reason and offers a change first: a different roast, a longer interval, a paused month or a smaller bag. It may offer a discount for one reason only, price, and at one level only.
The share of cancellations saved went from 11% to 27%, which is 66 more subscribers kept. The store puts that at roughly $16,600 in retained revenue, using how long earlier subscribers stayed, and saved subscribers may leave sooner than that.
If a subscriber asks to cancel a second time, the cancellation goes through at once. The store had 14 complaints about save attempts the year before and none in this period. The build took about two weeks.
A store at $1.8M a year built its on-site recommender so that it has to check live stock before it names a product. If the check returns nothing within two seconds, or the stock data is more than 15 minutes old, the agent gives category advice and offers a handoff to customer service. Across 2,640 product recommendations it named no unavailable item, and support tickets about unavailable items went from 30 to 40 a month to none.
Sessions that used the recommender converted at 3.4 times the rate of those that didn't, with a 22% higher order value. People who use a recommender are already closer to buying, so the store says that figure overstates the lift. The stock check needed a developer for about four days.
An outdoor-gear store at $4.3M a year tested a similar gift finder on 40 staged requests in one day. Seven recommendations named a product that had sold out that morning, and two more named one in stock only in the wrong size. Its agent read a catalog file exported at 3:10 each night, and on a normal weekday six to eleven of the store's top 30 products sell out.
Nothing reached a customer. The test ran on an ordinary September weekday, and the store sells out several times more often in its fourth quarter.
A home goods seller at $1.9M a year, on Amazon and Etsy, had an agent write listings and answer buyers. Nine of 612 buyer messages told the buyer to email the business for a faster replacement, which Amazon doesn't allow, and the account got a policy warning. A second warning in the same category could suspend it.
Fourteen of 38 rewritten titles also ran past the category's character limit and were cut off in search, and three of those listings lost placement over the next two weeks. Those three had brought in about $41,000 over the previous 90 days. The seller blames the cut-off titles, and three listings is too few to prove it. The agent had the product specs and none of the platforms' rules.
An ecommerce brand at $2.6M a year tested an agent that writes abandoned-cart and post-purchase texts. A review of 40 messages found three that named an item the customer hadn't ordered, because the agent read the latest cart instead of the completed order. The store set a rule that nothing would send live until 100 messages in a row were correct.
It was at 61 that week, and the founder approved live sending, in writing, two days before Black Friday 2025. Her reasoning was that the error rate already beat the manual process and the revenue window was worth the risk. Over four days the agent sent 18,400 messages, and 1,290 named the wrong item.
Opt-outs ran at 4.9% against a usual 1.2%, about 790 subscribers. A carrier began filtering the brand's number, and delivery fell to 71% for the rest of the weekend. The brand estimates $52,000 to $71,000 in lost text revenue that weekend, modeled against the year before.
The rule was right, and nothing enforced it. The decision to go live was argued on message accuracy, and what the rule protected was the subscriber list and the carrier relationship.
Numbers. It watches the figures that protect your cash and tells you when one moves. The build is in Run This Numbers Agent (For Protecting Cash).
The next sale in your customer list. The build is in Run This Customer Intelligence Agent (For Your Next Sale).
Your email list. An agent runs the weekly work on your list inside permissions you set. The build is in Let Claude (or ChatGPT) Run Your Email List (Full Playbook).
For tracking other stores' prices and claims, see How To Build A Competitor Watch Agent (Cowork Or Codex).
The skincare store and the coffee subscription each keep their rule in a file the agent reads on every run, with a check that blocks anything outside it. The supplements brand had its rule in a PDF, and its agent was told to follow advertising guidance. That instruction stopped none of the nine claims.
Writing the file is the hard part. The coffee founder spent a morning on hers, and it was the first time she had decided what she would and wouldn't pay to keep a subscriber.
In the cases on this page, agents write product descriptions, recommend products to shoppers, answer cancellation requests, reply to buyers and send order texts. Each of the three that worked reads its rule from a file and has a check it can't get past.
It can once the claims you may not make are in a list it checks every draft against. The skincare store's copywriter still reviews each draft, and gets the blocked ones with the matched term marked.
It won't if it has to check live stock before it names a product. The $1.8M store above built its recommender that way and named no unavailable item in 2,640 recommendations. The outdoor-gear store's version read a nightly file, and 9 of 40 test recommendations named a product that was sold out or in the wrong size.
It can once the platform's rules are in a file it checks every message against. The home goods seller's agent had none of them, and nine messages sent buyers to email, which Amazon prohibits. Platform rules change, so someone has to own that file and re-read the policy.
It can, with a list of what it may offer for each reason and a rule that a second request to cancel goes through at once. The coffee subscription kept 66 more subscribers that way and had no complaints about it.
I'd wait until it has passed the test you set for it. The $2.6M brand set a bar of 100 correct messages in a row, went live before reaching it and sent 1,290 wrong texts in four days. A carrier filtered its number until January.
The skincare store spent about $2,800 on legal work and under a week on the build. The recommender's stock check took a developer about four days. The coffee subscription's agent took about two weeks, including a morning for the founder to decide what she would offer.
Use Cowork or Codex for an agent that runs when something arrives, such as a new product or a cancellation request. Use Claude or ChatGPT for drafting and for Skills you run by hand. Everything else is the store platform, email tool and helpdesk you already pay for.
The Pocket Guide To Your First Working Agent walks through one agent, starting with your inbox.
Bionic Business is a newsletter by Samuel Woods. Three free issues go out each month, including The Scan, my round-up of AI news with a verdict on each item, and issues that walk through one system you can run that week.
Cortex is the paid side, at $50 a month. Members get a strategy issue and a build guide each month, with the files to install the system, plus Skill and agent packs. You can cancel at any time, but there are no refunds, for any reason. All sales are final. See Cortex.
Subscribe free: