The Delegation Gap: Where We Drew the Line Between Asking and Acting, and Why

Two surveys this year measured the same gap from different ends. StackAdapt's AI Delegation Gap report, built on 687 brand marketers and agencies across North America, Europe and APAC, asked how much decision-making authority people will hand to AI. IAB UK's State of AI in Advertising asked how far the industry has actually gone with agents. Both found the same shape: wide use, narrow trust.

The usual reading of numbers like these is that trust lags and will catch up. We read them differently. The drop from 90 to 50 is not people failing to keep up with the technology. It is people correctly noticing that recommending and acting are different jobs with different costs when they go wrong.

Why the caution is rational

StackAdapt's respondents named their barriers: brand risk (63 percent), data quality (56 percent) and lack of transparency (36 percent). Each one is a reason a wrong action costs more than a wrong recommendation.

• Data quality. An agent acting on a broken conversion feed acts with full confidence on a wrong number. A recommendation built on the same number at least passes in front of somebody who might notice the number looks odd.
• Brand risk. Some actions cannot be undone. An email sent to a customer list stays sent. A recommendation to send it can be ignored.
• Transparency. If you cannot see why a change was made and what it did, you cannot tell a good agent from a lucky one, and you cannot correct it.

None of these improves just because the agent gets better at its job. A more capable agent on bad data is still on bad data. That is why we did not design our levels around how capable the agents are. We designed them around what happens when they are wrong.

The industry's ladder, and why ours stops short of the top

StackAdapt's report lays out a ladder of authority: recommend, prepare for approval, act within rules, act autonomously. Cresva's levels map onto the first three rungs, and the fourth is deliberately missing.

The top rung needs three things before we would switch it on: a typed confirmation from the person enabling it, a check of the exact target of every change before it is written, and a track record of clean changes at Safe actions. The first two are engineering. The third is time, and it cannot be built faster than the changes happen.

What "safe actions" means, concretely

Safe is not a mood. It is a property of each kind of change: it can be undone, and it cannot raise spend. At Safe actions, once the brand's limits allow any change to run alone (they start at none), these run on their own:

• Pausing a Meta campaign, ad set or ad.
• Pausing a Google Ads campaign, ad group or keywords.
• Pausing a Klaviyo email flow.

These run on their own only inside a goal that sets a daily spend ceiling, because each one can raise spend:

• Resuming any of the above. A resume restarts spend at the budget it was paused on.
• Moving a budget, at most a fifth of the live budget per change, twice per campaign per day, never in a campaign's first day.

And these always ask, at every level: bid and bid strategy changes, new keywords, price changes. Anything that cannot be undone, such as creating a campaign, archiving an ad or sending an email, also needs the person approving it to type SEND. The full table, and the reasons for each row, are in the companion post on Meta's ad account connectors.

What builds trust, and where the product shows it

StackAdapt names six foundations for delegating responsibly: context, transparency, control, testability, accountability and connected infrastructure. We cannot claim all six. Here is what each maps to in the product today, and the gaps are in the next section.

What we still cannot show

• A track record at Safe actions. On 10 October, no brand on Cresva, our own included, ran above Ask first. The level works, and nobody has used it in production yet. Until somebody does, our claim about it is a design, not a result.
• What a bid change will do before it runs. That is why bid changes always ask. We would rather say so than put a number on it.
• What would have happened after a recommendation you declined. We record that it was declined and by whom. We do not measure an outcome for a change nobody made, because it has none of its own.
• That people read the proposals. Some proposals expire after seven days with nobody having answered them. An approval queue nobody opens is not oversight.

The last one is the uncomfortable one. Ask first only protects you if somebody asks back. The habit that fixes it is small: open Approvals once a day, and treat an expired proposal as a decision you did not make.

How to choose your own line

1. Start at Ask first, not at the level you think you will end up at. Watch a few weeks of proposals before deciding which kinds you would have approved every time.
2. Move one kind of change at a time. If every pause proposal for a month was approved, pauses are a candidate. Budget moves are a separate decision.
3. Put money behind a goal, not behind a level. A daily ceiling on a goal with an end date is a smaller promise than a level that applies to everything. Setting a goal agents can act on walks through what you sign.
4. Keep anything you cannot undo on ask. No level in Cresva changes that, and we would not build one that did.
5. Write the line down where your team can see it. The Autonomy page is that place in the product; a level nobody remembers choosing is not a decision.