Faster, not smarter: the trouble with AI that only knows how to discount
The newest wave of conversion-stage AI is genuinely good at one thing.
It spots a moment of hesitation and reacts in milliseconds.
The reaction is almost always the same. A discount.
Faster discounting, cheaper discounting, and more precisely triggered discounting. The action never changes. And that’s the problem.
What the new tools actually do
Credit where it is due, the mechanics are a real step up from a sitewide promo code.
- Checkout stall detection. The shopper pauses for 60 seconds, so the system fires a message and an offer.
- Triggered, single-use codes. A cart over a threshold gets free shipping. A price objection in chat gets 10% off for 15 minutes. A first-time buyer gets a welcome code.
- Cross-channel recovery. The shopper leaves, and a follow-up arrives via email, WhatsApp, or DM within the hour, naming the exact product they looked at.
Single-use, time-limited, triggered by behaviour. All better than blasting one code to everyone. None of it is the problem.
But every path ends in a discount
Look at the decision tree. It has one leaf. One endpoint. One thing it can do.
Detect a signal, deploy a code. The intelligence is in the detection. There is none in the decision, because there is no decision. The answer is always yes.
And that can be an expensive yes.
Take the 60-second checkout stall. It gets treated as a problem to fix with money. Often, it is not a problem at all.
People pause at checkout for ordinary reasons.
- They’re finding their card.
- They’re reading the returns policy.
- They’re checking the delivery date.
- They’re replying to a text.
- They’re comparing two sizes in another tab.
A pause is not a cry for help, and plenty of those shoppers finish on their own.
Fire a code at all of them, and you have paid people to do the thing they were already doing.
The recovery problem is the same problem, bigger
Cross-channel recovery has the same blind spot, only across a much larger group.
Over 70% of carts get abandoned, and that number has barely moved in a decade. So the recovery pool is enormous, and the temptation is to chase the whole thing with an offer.
But a high abandonment rate does not tell you who left for good. Some shoppers abandon thinking. To check a price elsewhere. To wait for payday. Or simply because the kettle boiled. A share of them come back on their own, no prompt needed.
When the follow-up always carries a discount, every one of those self-returners now returns with money off. You did not recover them. You caught them on the way back and handed them a voucher.
And all of this assumes the message even lands. Mostly it does not. A recovery sequence has to survive deliverability, then an open, then a click, then a checkout, and people drop out at every step.
Even abandoned cart flows, the best-performing automated email there is, place an order only around 3% of the time on Klaviyo's own benchmark.
Open rates look healthier, but Apple's Mail Privacy Protection now logs an open whether anyone reads the message, so that number is mostly noise. Firing across email, WhatsApp and DM widens the top of the funnel without fixing the leaks underneath it.
The reach you picture is not the reach you get.
Speed makes the same mistake cheaper to repeat
Here is the number worth sitting with. Nielsen estimates that around 84% of price promotions are unprofitable.
Most of that is subsidized volume. Sales you would have made anyway, now made at a lower price. Some of it is time-shifted, sales you would have made next week pulled forward to today and discounted for the privilege.
Automating the trigger does not fix any of this. It scales it. You are now making the same unprofitable offer in milliseconds, across every channel, to everyone who slows down for a moment. The leak is identical. The tap is just wider open.
The number that flatters everything
There is a reason this keeps happening. Most teams measure the wrong thing.
The usual metric is redemption. The code went out, was used, and the dashboard turned green. It looks like the offer worked.
It tells you almost nothing. A used code does not prove the sale needed it. The shopper who would have bought anyway also redeems and looks identical in the report. Counting redemptions cannot distinguish between offers that changed a decision and those that just gave away margin.
The honest question is not "did the offer get used?" It is "Did the offer change what the shopper did?" Those are different questions, and only one of them protects profit.
The move these tools are missing
Suppression.
Take the exact same signals and ask a different question. Not "what discount fits this trigger." Instead, "Is this shopper likely to convert without one?"
If the answer is yes, do nothing. Hold the offer back. Let them buy at full price, because they were going to.
Picture two shoppers who both stall at checkout for a minute. One has visited three times this week, has the item saved in a basket, and has landed straight on the payment step. The other arrived via a voucher code site and is bouncing between tabs. The giveaway machine offers the same 10%. A model reading intent gives the first one nothing and holds the offer for the second, where it might actually change the outcome.
Same signals. Different decision. That decision is the part that the discount-only tools never make. They can read intent. They only know how to spend it.
How to tell if your AI is only discounting faster
Three quick checks.
- Does it ever decide not to show an offer? If every trigger ends in a code, that is an automation, not a decision.
- Do you measure whether behaviour changed, or only whether codes got used? Redemption rate flatters everything.
- Are high-intent shoppers protected from offers, or do they get the same code as the wobblers? Run a holdout. Show the offer to one group, withhold it from a matched group. The gap between them is the only number that counts.
The point
Faster is not the same as smarter.
Discounting at machine speed still hands a margin to people who did not need the nudge. You have just made the mistake quicker and in more places at once.
The upgrade worth wanting is not an AI that discounts the instant someone pauses. It is one that knows when to stay quiet.
