AI can safely run the jobs that only read and report: the weekly report, stock run-out warnings, payout matching and win-back lists. Jobs that change money or what customers see, like pausing a losing ad set, should wait for a person's yes. Jobs that are hard to undo, like big stock orders, prices and product launches, should stay human. Sort jobs by risk and how easy they are to undo.
Meet Dana. She runs a made-up coffee brand.
It is Monday, 7 a.m. She has not had her own coffee yet.
She is already three spreadsheets deep. Amazon orders in one. Store orders in the next. Meta spend in a third.
By 10 a.m. the Monday report is done. It is last week’s report with new numbers.
Then she spots it. Her best blend runs out in 11 days. The next batch is six weeks away.
The report took three hours. The warning took one glance, and it came too late.
Most of Dana’s morning was work AI can do now. Some of it needs her yes first. A few calls should stay hers.
AI can run the jobs that only read and report. It can prepare money changes for a person to approve. Jobs that are hard to undo should stay with people.
The trick is not finding jobs AI can do. It is sorting them by what a wrong move would cost, and how easy it is to take back.
Two questions sort every job
Before you hand over any job, ask two things.
- What does a wrong move cost? A late report costs an hour. A paused ad set costs a day of sales at most. A wrong stock order can tie up cash for six months.
- How easy is it to undo? A report can be re-run. An ad set can be turned back on. An email sent to 20,000 people cannot be unsent.
Put those together and you get three groups.
- Let it run. Low cost, easy to undo. The AI does the job and tells you what it did.
- Needs a yes. It changes money or what customers see, but it can be undone. The AI warns you, writes the change, and makes it only when a person approves. Then it checks the change landed.
- Stay human. Hard to undo, or it sets direction for the brand. The AI can bring the numbers. A person makes the call.
That is the same idea as permissions in an AI harness: read freely, change very little, and make every change wait for a person.
The table: brand jobs sorted by risk and undo
Here are common jobs for a brand that sells on Amazon, its own store and Meta ads. They run from lowest risk at the top to highest at the bottom.
| Job | What the AI does | Cost if wrong | Undo | Who decides |
|---|---|---|---|---|
| Weekly report | Pulls every channel, writes what moved and why | An hour of confusion | Re-run it | Let it run |
| Data health check | Flags a source that stopped syncing or a login that expired | A stale number goes unnoticed | Nothing to undo | Let it run |
| Payout matching | Ties each Amazon and store payout to the orders and fees behind it | A fee error goes unfound | Nothing to undo | Let it run |
| Stock run-out warning | Says which product runs out, and when, at today’s sales pace | A late reorder | Nothing to undo | Let it run |
| Fee and cost alerts | Flags a fee per unit that jumped since last month | Margin leaks for weeks | Nothing to undo | Let it run |
| Win-back list | Builds a list of buyers who are past their usual reorder time | A wasted list | Delete the tag | Let it run |
| Pause a losing ad set | Writes the pause for ad sets that lose money after all costs | A day of lost sales | Turn it back on | Needs a yes |
| Block wasted searches | Lists Amazon searches that cost money and never sell | Some lost traffic | Remove the block | Needs a yes |
| Send the win-back email | Drafts the email for the list above | An awkward email | Cannot unsend | Needs a yes |
| Shift ad budget | Proposes moving spend toward what pays back | Weeks of wasted spend | Slow to read | Needs a yes, with care |
| Draft a reorder | Writes a purchase order from run-out dates and lead times | Cash tied up in stock | Hard once placed | Stay human |
| Change a best seller’s price | Models what a price change does to margin | Lost rank or lost margin | Slow and public | Stay human |
| Launch or cut a product | Brings the margin and repeat numbers | Months of work | Very hard | Stay human |
| New product claims | Can check wording against your rules | Legal and trust risk | Very hard | Stay human |
Notice the pattern. The top third only reads. The middle writes a change you can take back. The bottom spends cash or sets direction.
Jobs AI can run today
These jobs read your data and tell you something. They change nothing. That is why they are safe to hand over first.
The weekly report
Someone on most teams spends a few hours every Monday pulling exports. Amazon, the store, Meta, email. Then they paste it into a sheet and write a few lines.
An AI can do all of that before you wake up. The hard part is not the writing. It is making sure each channel is counted on the same basis, so the totals tie. Amazon orders and Amazon payouts are two different numbers. Our guide to settlements vs orders explains why.
Stock run-out warnings
A run-out warning is simple math. Units on hand, plus units on the way, divided by units sold per day. The AI runs it every day for every product and tells you which one runs out first.
The value is in the timing. A warning 40 days out is a reorder. A warning 9 days out is a scramble.
On Amazon, running out costs you twice. Amazon says an ad will not display if the product is out of stock (Amazon Ads FAQ). So you lose the sales and the ad traffic at the same time.
Payout matching
A payout is the cash a store or marketplace sends to your bank. Each one bundles sales, refunds, fees and ad charges. Matching it to the orders behind it is slow, dull work. It is also where fee errors hide.
The timing makes it harder. Amazon settles seller accounts about every two weeks, and it holds part of the balance back as a reserve for refunds and chargebacks (Amazon’s guide to seller payments). A Shopify payout can come in lower because of refunds, chargebacks, fees or a reserve (Shopify Help Center).
An AI can match every payout and list what does not tie. A person only looks at the gaps.
Win-back lists
Most repeat buyers come back at a regular gap. Say your coffee buyers reorder every 35 days. Anyone at day 60 with no order is drifting away.
Building that list each week is pure reading. The AI can tag those buyers in your store or email tool. Sending them anything is a separate step, and it needs a yes.
Try it: days of cover and your reorder date
This is the math behind a run-out warning. The defaults are a made-up example.
Those were the safe ones. Here is where it gets interesting.
Jobs that need a person’s yes
These jobs change money or what a customer sees. They can be undone, but not for free. So the AI should work in four steps. It warns you. It writes the exact change. It makes the change when you say yes. Then it checks the change landed.
Pausing a losing ad set
This is the best example of a job that needs a yes.
Here is a made-up case, with illustrative numbers. A travel mug brand runs a Meta ad set that sends shoppers to Amazon. Meta’s own report says the ad set returns $2.10 for every $1. That looks fine.
But Meta cannot see sales that happen on Amazon. Its number is a guess. Once you use Amazon’s own sales data, then take off Amazon’s fees and the mug’s landed cost, each sale loses about $4. At 30 sales a week, that is $120 a week, or over $6,000 a year, on one ad set.
No single system could show this. Meta does not know the fees. Amazon does not know the ad cost. It only shows up when the two are read together, with your product cost. The owner did not miss it. The data hid it. More on that in do Meta ads drive Amazon sales.
Meta already has a tool for part of this. Its automated rules check your campaigns, ad sets and ads, and can update them or notify you (Meta Business Help). One action turns an ad set off when it meets your conditions (Meta’s list of rule actions). But a Meta rule only sees Meta’s numbers. It cannot see Amazon’s fees or the mug’s cost.
So the AI should find this and write the pause. A person should say yes, because they may know something the data does not. Maybe that ad set is building an audience for a launch next month.
Blocking searches that never sell
On Amazon, some searches cost money every week and never turn into a sale. The AI can list them, with spend and clicks. A person approves the block. If a blocked search was worth keeping, it takes one click to put it back.
Sending the win-back email
The list was safe to build. The email is not safe to send without a look. It goes to real people, and it cannot be unsent. Have the AI draft it in your brand’s voice. Then read it.
Moving ad budget
Pausing one bad ad set is easy to judge. Moving $20,000 a month from one channel to another is not. Payback can take weeks to show. So this one needs a yes and a careful read of the numbers behind it.
Jobs that should stay human
These are hard to undo, or they set the direction of the brand. The AI can bring every number to the table. The decision belongs to a person.
- Big stock orders. Once a purchase order is placed, cash is gone for months. Have the AI draft it and show the run-out dates. A person signs it.
- Prices on your best sellers. A price change is public. On Amazon it can move your rank, and rank is slow to win back.
- Launching or cutting a product. This is strategy. The numbers help. They do not decide.
- Supplier and agency deals. Contracts, terms and people.
- What you say your product does. Any new claim carries legal risk. Keep a person on it.
- What counts as a cost. The AI should never decide on its own whether something like a creator fee belongs in your margin. Ask, write it down once, and use it every time.
This next part is the one most brands skip.
The real risk: confident answers on wrong inputs
People worry that AI will make things up. For a brand, the bigger risk is quieter. The AI answers confidently from the wrong inputs.
It mixes Amazon orders with Amazon payouts. It uses a product cost from last spring. It reads Meta’s guess about Amazon sales as fact. Each answer sounds right. Each one is wrong.
So before you hand over any job, get your numbers to mean one thing. Each figure needs a clear basis and period. Contribution margin, meaning what a sale leaves after the costs that come with it, should be worked out per channel. Our contribution margin guide shows how. A job built on a shaky number is shaky, however well it runs.
Start with these three
If you hand over nothing else this month, hand over these.
- The Monday report. It saves a few hours every week. It changes nothing, so a mistake costs little. It also shows you fast whether the AI’s numbers tie to your own.
- Stock run-out warnings. Daily, for every product, with a date. The value of an early warning is high. The cost of a wrong one is a second look.
- A weekly list of ad sets that lose money after all costs. Have the AI write the pause for each one. You approve or skip. Check a week later that the pause held.
Run these three for a month. Then check them against what you already know. Did the report tie to your books? Did a warning come early enough to matter? Did the paused ad sets really lose money?
If the answers are yes, move the next job up a row.
Where Synthesis fits, and where it stops
Synthesis connects a brand’s stores, ads, email and books, and checks each number against its source. Its Monday report and stock warnings run on their own. Its automations warn you first, write the change, make it in Meta, Amazon or Shopify when you say yes, and check it landed.
It does not place stock orders, set prices or decide what to launch. Those stay with you. And it has no chat of its own. You ask your questions inside Claude, ChatGPT or another AI app you already use.
Wrap-up
Sort every job by two questions: what does a wrong move cost, and how easy is it to undo?
Let the AI run the jobs that only read. Make it ask before it changes money. Keep the hard-to-undo calls for people.
Then start small. Three jobs, one month, checked against what you already know.
Questions people ask
What can AI automate for an ecommerce brand?
Start with jobs that read and report. The weekly report, stock run-out warnings, matching payouts to orders, and building win-back lists are all safe to hand over. They change nothing, so a mistake costs a little time, not money.
Should AI be allowed to change my ad budgets?
It can write the change, but a person should approve it. Pausing a losing ad set is easy to undo, so a quick yes is enough. Moving large amounts of budget between channels is harder to judge and should get a closer look.
What should never be automated in a consumer brand?
Keep anything that is hard to undo or that sets direction. That means big stock orders, price changes on your best sellers, product launches, supplier deals, and new claims about what your product does.
How do I decide which jobs to give to AI?
Score each job on two things. How much could a wrong move cost, and how easy is it to undo? Low cost and easy to undo means let it run. Either one high means it waits for a yes. Both high means a person does it.
Is AI automation safe if my data is messy?
Not yet. The biggest risk is not a strange answer. It is a confident answer built on the wrong inputs, like mixing Amazon orders with Amazon payouts. Get every number to mean one thing first, then hand over jobs.
What are the first jobs a brand should hand to AI?
Three good first jobs are the Monday report, stock run-out warnings, and a weekly list of ad sets that lose money after all costs. Each one saves hours, and none of them changes anything without a person.