Toronto-headquartered. Owned by the Teacher’s Union. Very polite. Puts "please" in front of API error messages. A company so Canadian that when their own survey data came back and basically told them to apologize, they said sorry and published it anyway.

I say that with love. Canada gave us poutine, Drake, and a metric system that actually works. It also apparently gave us a forty-page market research report that reads, top to bottom, like a hostage tape written by the hostage.

Here's the setup. StackAdapt commissioned a report. Real survey firm, NewtonX, 500 marketers, the whole legitimate-looking apparatus.

Point of the report: prove the industry is finally ready to let AI touch the budget. CMO letter up top, the kind that opens with "AI already has a seat at the marketing table," like it's a toast at a wedding nobody invited the groom to.

Then you turn the page. And the data does not agree with the toast. At all. It's the polite Canadian equivalent of your best man standing up at the reception and reading the bride's old diary entries about somebody else.

And before we even get into the numbers, a quick shame roll call, because it matters. MarTech Series, Advanced Television, PPC Land, and MediaPost all covered this report in the last 24 hours. Every single one of them ran it as news. CMO quote up top, headline stats repeated like established fact, zero questions asked about who paid for the survey, how the questions were worded, or whether the "50%" printed in size-72 font actually means what it claims to mean. Four newsrooms, one press release, and not a single follow-up question among them.

Shame on all four.

More on exactly how badly they whiffed further down, but it needed saying up front, because this is the part of the ad tech ecosystem nobody likes to talk about: the trade press doesn't fact-check vendor research anymore, it just republishes it with a byline attached.

There's a chart. StackAdapt calls it the "authority ladder." Cute name. Sounds aspirational, like something you'd climb toward enlightenment. It is instead a horror movie with a very orderly beginning and a floor that gives out.

Rung one: 90% of marketers are fine with AI recommending things. Sure. Recommend away, little robot.

Rung two: 89% are okay with AI prepping a change for a human to approve. Also fine. That's basically a very fast, very tireless intern who never asks for a raise.

Rung three: 78% will let AI act inside rules a human already wrote. Getting warmer. Still tethered.

Rung four, the one that actually matters, the one where AI gets to act on its own, after it has already proven it works, the one where the whole "delegation" promise lives or dies: 50%.

Wait, what?

A coin flip. That's the number. Proven performance. Documented results sitting right there. And half the industry still looks at it and says no, thank you, eh.

That's not a gap. That's a cliff with a hand-lettered sign reading "Delegation This Way," nailed to a guardrail that gave out three feet back.

Here's where the actual science comes in, and where this chart stops being impressive and starts being suspicious. StackAdapt has never released the exact question wording, the response scale, or the order those four questions were shown in. That's not a technicality, it's the whole ballgame. This is a textbook setup for what survey researchers call acquiescence bias: when you show respondents a series of increasingly permissive statements in ascending order, people tend to keep agreeing out of momentum, right up until the final, most extreme statement, where the psychological cost of agreeing finally outweighs the social cost of disagreeing. Stack four "how comfortable are you with AI doing X" questions from mild to wild, and a linear-looking staircase with one big drop at the end is exactly the shape you'd expect to see whether or not the underlying attitude is real. We don't know if that's what happened here, because StackAdapt won't show us the instrument. That's the point. A company that built an entire report around the importance of transparency and auditability declined to be transparent or auditable about the one chart the whole report hangs on.

Only 6% of marketers act on in-platform AI recommendations almost always. Six percent. You have better odds correctly guessing a stranger's birthday.

And when the survey asked marketers what they actually need before trusting AI with anything, the category that scored dead last, at 31%, was explanation and reasoning.

Read that twice. The single thing marketers said they need most is the exact thing AI is worst at, according to StackAdapt's own respondents, printed in StackAdapt's own report, under StackAdapt's own logo.

There's also a quieter admission sitting a few pages later that deserves way more attention than it got from literally anyone who covered this.

StackAdapt's own report says, in its own words, that comparisons between the 500-person external survey and the 187-person client survey should be treated as "directional" because "question wording, response formats, and sample composition may differ." Translate that out of PR-speak: StackAdapt is telling you, in writing, that you cannot rigorously compare its two headline data sets to each other. And then the report spends the next thirty pages doing exactly that comparison, over and over, brand versus agency, external versus client, region versus region, as if the caveat two paragraphs earlier didn't exist.

You don't get to disclaim your own methodology and then build your entire narrative arc on the thing you just disclaimed.

StackAdapt lists the features supposedly one product roadmap away from fixing all of this. Predicted impact before you apply a change: 37%. Limited-budget testing: 36%. A clear explanation of why a change was made: 35%. A full audit trail: 31%. Easy undo: 31%.

Not one clears 40%. Not one.

If any single feature were the magic key, you'd expect the number to move somewhere. It doesn't. Every proposed fix lands in the same flat, shrugging middle, which tells you the thing StackAdapt never quite says out loud: nobody thinks a button fixes this. They think a person needs to be on the hook when something breaks. And right now, nobody is.

Worth sitting with why this pattern matters statistically, not just rhetorically. In a well-designed survey, if trust really were being blocked by one specific missing feature, you'd expect to see separation in the data, one option clearly outperforming the rest, a genuine signal poking up out of the noise. Instead you get five options bunched within six points of each other, all hovering in the mid-thirties. That flat clustering is itself evidence. It suggests respondents weren't discriminating between meaningfully different fixes, they were just distributing a vague, generalized unease across whatever boxes the survey gave them to check. That's not a market waiting for the right feature. That's a market that doesn't trust the category yet, full stop, and no single checkbox item is going to fix a trust problem that's actually about accountability.

Seventeen percent of respondents said, flatly, there is no clear accountability when AI makes a bad call. The other 83% scattered across team, leadership, manager, and individual operator, and none of those categories broke a third.

Picture the room. Campaign tanks. Budget's gone. Everyone's pointing at everyone else, very calmly, very reasonably, the way people do at a meeting where nobody wants to be the villain. And in the middle of the circle sits the AI, unbothered, already three optimization cycles into the next campaign, having never once been asked to explain itself.

Nobody in this survey can agree on who's holding the bag when the bag catches fire.

Here's the methodological wrinkle nobody covering this bothered to mention, and it's a real one. This was a multiple-response question. StackAdapt's own footnote admits it: "Multiple responses were allowed; percentages may exceed 100%." That means a single respondent could check "the team," "the manager," and "leadership" all at once, and the survey would count that as three separate data points reinforcing a sense of shared accountability, when what actually happened is one confused person hedged their bets across every box on the page. A multi-select question cannot tell you whether an organization has one clear accountable owner. It can only tell you how many boxes people were willing to tick. Those are different questions with different answers, and StackAdapt's chart quietly answers the easier one while implying it answered the harder one.

Here's the part that should have gone in the executive summary, and conspicuously didn't.

StackAdapt surveyed its own paying clients too. 187 of them. Real customers. People who already bought the product, sat through the demo, believe in the mission enough to sign the invoice every month.

Where do these true believers want AI's role to grow? Not bidding. Not budget. Reporting. The single lowest-stakes, lowest-authority category on the entire ladder.

The people with the most reason to trust the machine looked at the machine and said, in effect: you can write us a nice little summary. That's the ceiling. From the customers. From the people already paying for the thing.

And this is the group you'd most expect to be inflated in StackAdapt's favor. In survey research this is called social desirability bias combined with sunk-cost self-selection: people who already spent money on a product have a documented tendency to rate it more favorably than a neutral population would, partly to avoid the discomfort of admitting they made a bad call, partly because they're more exposed to the vendor's own marketing narrative in the first place. If anything, the client survey should have produced the rosiest numbers in the entire report. Instead, even this group, tilted in StackAdapt's favor by every incentive a behavioral scientist could name, still capped their enthusiasm at "let it write us a summary." When your most favorably biased sample still won't back your own thesis, that's not a delegation gap. That's the ceiling.

Look at what the rest of the industry is doing while StackAdapt was busy commissioning a document that quietly argues against its own product roadmap.

The Trade Desk is on every single earnings call narrating an agentic future like it already signed the lease, hung the art, and started forwarding its mail there. Every DSP with a Series B and a Figma file is slapping "AI-native" across a landing page as though it were a Michelin star instead of a font decision made by someone in brand at 4pm on a Thursday. Everyone is performing readiness right now. It's the industry's current group costume.

StackAdapt did something that should have been braver, and in a narrow sense actually was. It commissioned real survey data instead of just vibes and a stock photo of a glowing brain. And then the data came back and said, in language polite enough to bury in a chart legend but completely unmistakable once you sit with it for more than ninety seconds: not yet. Not like this. Not without receipts.

StackAdapt published all of it anyway. Then it put a "book a meeting" button at the bottom. Read that sequence again slowly, because it's the whole comedy in miniature: your own customers just told you, in writing, in a survey you paid for, that they do not trust the machine with their money. Your response is to immediately try to sell them the machine again, harder, with a nicer PDF this time and a CTA at the end.

At least four trade outlets covered this report inside the first 24 hours. Not one of them opened the hood.

MarTech Series ran it as a straight news release. CMO quote up top, framed uncritically, the 90/89/78/50 ladder repeated like four independently verified facts about the physical world rather than four numbers from an unreleased survey instrument commissioned by the company the numbers happen to flatter. No mention of methodology. No mention of who funded it, beyond the obvious.

Advanced Television did essentially the same thing, dutifully walking through the regional breakdowns, quoting the CMO twice, treating "only 50 per cent comfortable using Autonomous AI" as a headline stat worth repeating rather than a five-alarm fire worth investigating. The word "commissioned" doesn't appear. NewtonX barely registers as more than a passing credit line. This is the trade press equivalent of reading a company's own earnings call transcript and calling it an independent audit.

PPC Land gets a little more credit, and only a little. Their piece actually noticed the shape of the collapse, describing the pattern as close to linear until the final step, where it collapses, which is at least an observation rather than a stenography job.

They quoted the EMARKETER analyst's sharpest line in the whole report, the one about trust being the real obstacle and automation pulling back the moment nobody can answer for it. Real instinct.

But even PPC Land stopped short of the obvious follow-up: if trust collapses the moment nobody can answer for a decision, who exactly is answering for StackAdapt's own decision to publish this without releasing the instrument behind it? Nobody asked.

MediaPost, for its part, put the whole thing behind a paywall and led with "Advertisers Love AI Until It Reaches For The Budget," which, credit where it's due, is at least an honest headline about the shape of the collapse. But the visible portion of the piece is two sentences and a subscribe wall. Whatever scrutiny exists behind that login, and based on the opening, there's no strong reason to assume there's much, it's not doing the industry any favors sitting locked away while the press-release version circulates everywhere for free.

None of these four outlets asked for the actual question wording. None of them asked what "proven performance" meant to the people answering. None of them asked what share of the 187 "customers" were customers customers, versus prospects, versus agency partners whose account rep happened to be in the room. This is what happens when a company hands the trade press a PDF with charts, a methodology page that sounds rigorous enough to skim past, and a quote from an actual outside analyst, and everyone downstream assumes someone upstream already did the fact-checking. Nobody did. A press release wearing a lab coat got treated like peer review.

Nobody at StackAdapt stood up in a meeting and said "we are not ready." No confession exists, and I went looking specifically for one, because that would have been the easier story to write. What exists instead is forty pages where nearly every meaningful number, the ladder's 28-point cliff, the 6% action rate, the last-place explainability score, the flat unlock percentages, the accountability free-for-all, the client vote for the least ambitious use case on the menu, all point the same direction, independently, without needing to be forced into alignment. You don't need a confession when the evidence agrees with itself this consistently.

But here's where I have to slow down and actually explain the mechanics, because "this is bullshit" isn't an argument, it's a conclusion, and a real one needs the receipts.

Problem one: undisclosed instrument. StackAdapt has not publicly released the exact question wording, the full response scale, the scale anchors, or whether "agree" and "comfortable" meant top-two-box responses or something looser. Without that, you cannot distinguish a genuine 50% comfort level from an artifact of vague, agreeable phrasing. What does "after performance is proven" even mean to a respondent answering a web survey in ninety seconds? An incrementality test? A week of decent ROAS? A vibe? Nobody knows, because the definition was never published. A number without a defined denominator isn't data, it's decoration.

Problem two: no sampling or weighting disclosure. We know the industry survey had 500 respondents across six countries. We don't know the response rate, the exclusion criteria, or whether NewtonX weighted the sample to any target population. Standard practice for a defensible public survey is to publish all of this. StackAdapt published none of it.

Problem three: the "global" framing doesn't survive a second look. The external sample skews harder than the word "global" suggests. The EMEA bucket is 72% UK respondents. The APAC bucket is 84% Australian. That's not a scandal, it's a sample size wearing a trench coat labeled "worldwide." A report that wants to claim regional insight about EMEA is, in practice, mostly describing British marketers, and calling that EMEA is like calling a ham sandwich a continental breakfast.

Problem four: correlation dressed up as causation. The report explicitly argues that accountability, not accuracy, is the primary barrier to AI delegation. That's a causal claim. Survey data of this kind, without experimental design, control groups, or even published cross-tabs, can only show association, not causation. Brand risk, data quality, transparency, and performance volatility all score as major barriers too, at comparable magnitudes. The report picked the storyline that flattered its own "responsible delegation" product pitch and downplayed the others. That's not analysis. That's marketing wearing analysis as a costume.

Problem five: the client survey isn't independent, and StackAdapt admits as much itself, tucked into a footnote about "directional" comparisons that the report otherwise ignores for thirty straight pages.

Taken together, none of this means the underlying intuition is false. It's probably true, broadly, that marketers are more comfortable with guardrails than with autonomy, and that's a completely reasonable finding. What's not true, or at minimum not demonstrated, is that these specific numbers mean what StackAdapt is telling you they mean, with the precision StackAdapt is implying by printing them in size-72 font on a slide. The finding is plausible. The certainty is manufactured.

Here, finally, is the reporting question that matters more than any single statistic in the report, and the one nobody downstream bothered to ask.

There is no publicly available independent evaluation of StackAdapt's own platform. None. No holdout-controlled test on bid accuracy. No verified incremental ROAS study with real attribution behind it. No published data on how often the platform's own automated changes get reversed, or why. No third-party look at brand safety or supply-path quality when the system runs without a human watching over its shoulder.

StackAdapt spent forty pages building a genuinely persuasive argument that trust has to be earned through transparency, testability, and audit trails before AI gets more authority. It has not, anywhere in the public record, applied that same standard to itself.

So here's the sentence StackAdapt never printed in its own report, and probably should have: if the bar for trusting AI with a budget is proof, this report is the clearest evidence yet that nobody, including the company that commissioned it, has cleared that bar.

How we reported this: This piece draws on StackAdapt's publicly released report, "The AI Delegation Gap" (2026), commissioned by StackAdapt and fielded externally by NewtonX among 500 marketing and advertising professionals from May 19 to June 8, 2026, plus a separate 187-respondent survey of StackAdapt's own clients, along with trade coverage of the report from MarTech Series, Advanced Television, PPC Land, and MediaPost. All percentages cited are pulled directly from the published report, including its own footnoted disclosure that the external and client surveys allow multiple responses and should be read as directional rather than precisely comparable. StackAdapt has not released the underlying questionnaire, response scale, weighting methodology, or individual base sizes for its headline figures; where a claim depends on that missing detail, it is flagged as such above. ADOTAT contacted StackAdapt for comment on the report's methodology and on whether any independent evaluation of its own platform exists; this piece will be updated with any response received. ADOTAT has no financial relationship with StackAdapt, The Trade Desk, or NewtonX, and no company named in this piece reviewed it in advance.