AI in campaign management: what to hand to a tool, and what stays with a person
Every ad platform now ships automation, and every tool promises an AI agent that will run your campaigns. Whoever actually runs them needs one answer: which work can be handed off tomorrow morning, and which decision has to stay with a person. We looked at 1,499 store-days to see why an automatic alert is not enough.
The algorithm points a direction. A person decides where to go.
32.5%
of days in an online store's Meta account were 30% or more below what the week before predicted. One day in 3. 1,499 store-days
50%
of those weak days were back to a normal pace within three days. After an ordinary day: 61%
11
alerts a month per store, median, if you set an alert on a 30% one-day drop. Between 6 and 13 across stores
AI in campaign management: what can be handed off, and what cannot
AI in campaign management means using a tool that works on a business's ad account to prepare, analyse and alert, while a person stays the one who decides how much money goes out, on what, and when to stop. You can hand it ad copy variants, a first pass over reports, alerts on anomalies and a weekly summary. Budget, business goal, promotions, message and brand, and the decision to stop stay with a person. That is the same logic as implementing AI in a business: a task that repeats, whose output can be checked and where a mistake is cheap goes to the tool. Everything else stays.
I lead campaign management and the team that carries it out, and every week someone asks me whether we can let AI run it yet. My answer is "part of it, and you have to choose which part". This article splits the work in three: what the platforms already do on their own, what is worth handing to a tool, and what must stay with a person. In the middle there is a finding about automatic alerts, because it showed us why "tell me when something breaks" is harder than it sounds.
What the platforms already do on their own
Before adding a new tool, know how much automation is already running in the account. On Meta that is Advantage+ campaigns, automatic placements and a broad audience the system widens by itself. On Google it is automated bidding, Performance Max and asset groups the system mixes. In both cases the decision of who to show an ad to, and how much to bid on each impression, is no longer with whoever manages the account. What is left is the budget, the definition of what counts as a conversion, the creative and the products that go in.
That means an agent that will "catch the auction" adds little. The platform does that job every hour, against data no outside tool can see. Where AI really helps is the work around it. The platform optimises against what it can see. It does not know there is a promotion next week, that the lead product is out of stock or that a shipment is delayed, and it does not know what actually reached the bank account, as we laid out in the article on reported versus real ROAS. Performance Max in Google Shopping is a good example: the system decides by itself where to show, and whoever manages it has to know what it was given to work with.
What to hand to AI
Four jobs pass the three tests from the pillar article: they repeat, the result can be checked, and a mistake is cheap.
Ad copy variants. Headlines, short and long versions, different wordings of the same offer. A person picks and approves, and the ad gets measured in practice, so a mistake shows up fast. This pairs with the refresh pace of ads: most new ads never catch budget, and anyone who wants to test many versions can produce them cheaply.
A first pass over a report. The tool goes over the week's spend, purchases and clicks and returns a list of what moved and what did not. It does not decide what to do. It saves the first fifteen minutes of staring at numbers.
Alerts on anomalies. Money spent with no purchase, a campaign that stopped, an ad that was rejected. An alert on something technically broken is always worth having. An alert on a "weak day" is a different story, and the data below is about it.
A weekly summary. Three lines on what happened, what changed and what looks odd, every Sunday morning. Whoever reads it adds what the tool does not know.
An AI agent for a business that advertises: how the split looks
This table is the decision tool. Copy it, fill in the jobs your business has, and decide tomorrow morning. "Check" means the tool prepares and a person approves every time, until it is proven safe to release.
Job
Decision
Why
Ad, headline and email copy variants
Hand to AI, with approval
A mistake is cheap, and the ad is measured in practice
Weekly summary of the numbers
Hand to AI
Repeats every week, and a person reads it before it goes out
First pass over a report
Hand off, and check the conclusion
The tool finds what moved. What it means needs context
Alert on money spent with no active ad or no tracking
Hand to AI
A technical fault, nothing to argue about
Alert on a "weak day"
Check
Half the cases are back to a normal pace within three days, as the data below shows
What to duplicate and what to switch off in a campaign
Check
Depends on stock, margin and promotions, which the tool cannot see
Budget and business goal
Keep with a person
How much money goes out is the owner's decision
Promotions and timing
Keep with a person
A promotion changes every number, and the tool does not know when it starts
Message, brand and the promise made in an ad
Keep with a person
What goes out in the business's name stays with whoever signs for it
Checking against store revenue
Keep with a person
The platform reports what it saw, not what came in
Deciding to stop a campaign
Keep with a person
Stopping can send it back to learning, and it is hard to undo
The decision tool: which campaign management work goes to AI, which is checked, and which stays with a person. Our experience managing campaigns for online stores, 2026. Not a sample
A "weak day" alert: fires a lot, says little
An alert is the first job people want to hand to a tool, and it sounds simple: "tell me when purchases drop". We tested how simple it really is. We took the Meta accounts of Israeli online stores, May to October 2026, and defined an alert the way any tool defines it: a day on which purchases are 30% or more below what the pace of the previous seven days predicted for that day's spend. Sample size: 9 stores, 1,499 counted store-days, and 1,473 thousand ILS of spend by these stores over the whole period.
That alert fired on 32.5% of days. One in 3. For the median store that is 11 alerts a month, and across stores the range is 6 to 13. An alert every three days or so is not monitoring. It is noise that teaches whoever receives it to stop opening it.
How many days an alert fires, by depth of the drop
Share of store-days on which purchases were below what the previous seven days predicted, at each threshold. Even at a 50% drop, an alert fires about one day in five.
Preneuriat data, 1,499 store-days across 9 stores, May to October 2026
The more important question is what happens next. We took every flagged day and looked at the three days after it: did the pace return to within 15% of the store's normal rate. In 50% of cases it did. But after days that were not flagged, in 61% of cases the next three days were also at a normal pace. The difference between "a day the alert caught" and "an ordinary day" is just 11 percentage points.
How many flagged days returned to a normal pace within three days, by depth of the drop
The dashed line is an ordinary day, after which 61% of cases look normal. After a flagged day, at every depth of drop, about half the cases returned to normal.
After an ordinary day
Preneuriat data, 1,499 store-days across 9 stores, May to October 2026
A weak day does not tell you whether something broke. Half of weak days are back to a normal pace within three days, and after an ordinary day it happens only a little more often.
Two things follow. First, depth of the drop does not help. When the threshold goes from 20% to 50% there are fewer alerts, but the return rate stays between 47% and 50%. A deeper one-day drop is not a sign the situation is more real, because purchases are a small number that moves for no reason. Second, a daily alert is good for faults, not for decisions. Whoever stops a campaign on one weak day stops, half the time, a campaign that would have come back by itself. And stopping and restarting can send a campaign back into a learning period, as we described in the article on Meta's learning phase.
At the week level the picture is different. A week in which purchases were 30% or more below the previous four weeks showed up in 15.4% of store-weeks, and the following week was still weak in 59% of cases. That is a small sample, 29 weeks whose follow-up we could check, and it should be read as a direction, not a law. But it fits what managers see in practice: the useful alert is on a week, or on several days in a row, and after it a person looks at the context.
What must stay with a person
Five things you do not hand over, even when the tool offers to take them.
Budget and business goal. How much money the store is willing to spend, and what its monthly goal is, is set by cash flow, stock and margin. The platform will take any budget it is given, and the tool will suggest raising it when the numbers look good. Whether the store has more money to spend is the owner's decision.
Promotions and timing. A promotion next week changes the pace, the conversion rate and what is fair to compare. A tool that looks at the previous weeks will say something "dropped" exactly when the store lowered a price, or "rose" exactly when it raised one. Whoever decides must know about the promotion.
Checking against store revenue. What the platform reports and the revenue actually received are not always the same, and without comparing them you cannot know whether a campaign really works. This check is done against the store's orders, at least once a week.
Message and brand. What you say to the customer, how, and what promise you make. A tool can suggest ten versions. It does not know what the business is willing to promise, so it does not know what it must not write.
The decision to stop. Stopping and restarting can send the campaign back to a learning period, so a person decides, usually after a week and not after a day, once they can see the context: a promotion, stock, the season.
What to do tomorrow morning
List the repeating jobs in the account. What happens every week, and where someone copies numbers from one table to another. That is usually the weekly summary and the report.
Start with copy variants and a weekly summary. The two jobs where a mistake is cheapest. Every result goes past a person for the first weeks.
Set alerts on faults, not on weak days. Money going out with no tracking, a campaign that stopped, an ad that was rejected. Anyone who wants a performance alert sets it on several days in a row, or on a week.
Keep the money in a person's hands. Budget changes, stopping and starting a promotion go through approval, even when the tool suggests them well.
Check every week against the store's orders. Not against what the platform reports.
This is the second article in the series on implementing AI in businesses. For the full order of work, from choosing a task to measuring savings, read the opening article on implementing AI in a business. And if you want to go through the process with someone who has done it before, without changing systems, you can read about AI implementation with Preneuriat.
Questions we get
What can be handed to AI in campaign management?
Ad and email copy variants, a first pass over a weekly report, alerts on technical faults and a weekly summary. All of them are jobs that repeat, whose output can be checked, and where a mistake is cheap. Budget, promotions, message and the decision to stop stay with a person.
Can an AI agent run campaigns without a person?
Not all of the work. The platforms already make the bidding and placement decisions themselves, and an outside agent adds little there. Decisions about money, promotions, message and stopping need a person, because they depend on information that is not in the ad account: stock, margin and the store's promotion calendar.
An AI agent for a business: where to start in advertising?
With a weekly summary of the numbers and copy variants for ads. Two jobs that repeat, where a mistake is cheap, and that a person reads before they go out. After a month, count how many corrections were made, and only then expand to alerts and report analysis.
When should you alert on a drop in performance?
Immediately for faults, and for performance only after several days or in a weekly summary. In the data we examined, an alert on a 30% one-day drop fired on 32.5% of days, and in 50% of those the pace returned to normal within three days. After an ordinary day it happened in 61%, so one weak day is not a reason to act.
Should you stop a campaign that dropped for one day?
Not on one day. Daily purchases are a small number that moves for no reason, and stopping can send the campaign back to learning. Look at several days in a row or a week, and check whether there is an explanation: a promotion, stock, a tracking fault.
How do you check that AI is not misleading you in campaign management?
Compare what it reports to the store's orders, not to what the platform shows, and once a week count how many corrections a person made to its output. When there is a gap between the report and real revenue, start with real ROAS.
Everyone talks about AI, and most businesses that tried it are left with a subscription nobody uses. This is the opening article of a series on AI in small businesses and online stores: not which tool to buy, but which task to take first, how to do it in stages, and how to know after a month whether it really saved time or just moved work around.
Every month the same question sits on the design schedule: how many new ads to make. We opened 55 store-months of Israeli online stores, counted what went up each month, how much of it caught budget, and what that did to purchases per shekel.
Every online store offers free shipping over a certain amount, and almost none checks what that threshold does. We opened 33.4 thousand orders, shekel by shekel around each store's threshold, to see whether customers really top up an order to reach it, and what the store pays for that.
Daniel Hirsh1 Nov 202612 min read
Fit check
Want to see your real ROAS?
Every Preneuriat client gets a live dashboard: store revenue, all media spend, real ROAS, updated all day. The first call goes both ways.