Giving the AI your target CPA up front, or letting it infer one from history?

If you give the assistant a target CPA, say 40 for leads, it judges every campaign against that line and flags problems fast. If you leave it out, it guesses a baseline from the last 30 days. That's fine when the data is clean, but a tracking gap or a holiday dip can make a decent campaign look like a loser, and any cuts it suggests come from a baseline that was never real.

A stated target has its own problem: it goes stale. If margins shifted in spring, the old number quietly steers every recommendation. Putting a date on the target in the brief helps, and so does asking the assistant which benchmark it's actually using, which is a cheap check. Has anyone been burned by the guessed baseline?

Agreed on the stale target. For e-commerce, the number people forget to update is contribution margin after shipping and returns, not CPA. A 40 CPA target set before returns went from 5% to 18% is a completely different target than it looks. So give the assistant the margin and the return rate, with a date, instead of handing it a single CPA.

Iris is right that margin beats CPA for e-commerce, and I'd go one step further: give the assistant your break-even CPA, not your target CPA. Break-even is what you can pay before a sale starts losing money, so it's a hard ceiling. Target CPA is more of a wish, and it often sits above that ceiling. Once the assistant knows the ceiling, it can tell you when a campaign is actually unprofitable instead of just behind goal. Then set your target below break-even, so the buffer is something you chose on purpose.

I don't think a target below break-even is a safe default. It works when your margin is wide and stable, but for a lot of businesses the break-even number is itself a moving estimate, because returns, shipping and discounts all shift month to month. If you set the target a fixed 20% under a break-even that's already off, you end up optimizing for a buffer that isn't really there. I'd rather see both numbers dated, with the gap written down as a decision you revisit instead of a rule that stays put.

Writing the gap down as a decision makes sense, and returns add a wrinkle: they land weeks after the sale. If a campaign ran in March and the refund window closes in April, a break-even calculated in March is too optimistic, because the refunds aren't in yet. So date the return lag too, something like "returns settle about 30 days after the order, so only use orders older than 45 days for margin." That tells the assistant which period it can trust, and it won't call a campaign profitable on a month whose refunds haven't arrived. For a store where returns are a big share of orders, that lag can distort the number more than the 20% buffer does.

The 45-day cutoff is safe, but it throws away your newest weeks, and those are exactly where a change you made would show up first. Whether that costs you depends on order volume. At a few hundred orders a week, 45 days still leaves plenty of data to judge a campaign on, so keep it strict. If you're smaller, waiting that long means you're judging campaigns on old data. In that case it's usually better to take the return rate from your last mature cohort, apply it to recent orders, and label the margin as estimated.