I've spent the past months running a large, multi-market transformation of how campaigns are created. The kind of project where the ambition on the table is genuinely exciting. Faster content, more variants, more markets served at once, less manual production drag. The pull is real: efficiency gains that enable cost reductions, revenue upside from faster, more personalised campaigns, and that ever-present promise of being “future ready”. The technology to do this is, for the first time, actually good enough. And that's exactly where things get tricky.
Because the moment "what we could build" becomes almost limitless, it quietly crowds out something much less glamorous: the business work of deciding what this actually needs to do, and how it should run once it's live. Skip that, and it doesn't cause a problem right away, but further down the line, the conclusion becomes "the technology isn't doing what it was supposed to". Now you've got a costly, resource-draining problem on your hands, and what seems like the obvious fix, better technology, won't get you back to the part you skipped.
A business initiative, run like a tech project
AI content scaling is a business initiative, but by default, it rarely gets treated as one. It gets framed, funded, and staffed as a technology and production capability project instead: build the engine, plug it into the workflow, measure throughput. That's the more concrete problem, and concrete problems are easier to greenlight than ambiguous ones.
The less concrete problem - what do we actually need this to do, for whom, and how that gets run day to day, gets deferred. It isn't negligence, it genuinely can't be answered cleanly up front, because both the technology and the business need to keep moving. So teams do the reasonable-seeming (and arguably more exciting) thing: start building capability now, and figure out the rest as they go.
Why that deferral turns into a “tech constraint” that isn't one
Those deferred questions don't stay quiet forever. They resurface later, once you try to scale or speed up, and by then it looks nothing like what it is. "The output isn't good enough" becomes round after round of revisions. "The workflow is too slow" becomes content stuck for days before it goes live. "The tool can't keep up" becomes a growing backlog nobody's getting through. It gets diagnosed as a technical limitation, and someone goes looking for a better model or a faster pipeline.
But the tool is rarely the constraint. Endless revisions usually mean nobody agreed on what good enough looks like. Content stuck for days usually means nobody knows whose sign-off it needs. A growing backlog usually means there's no one with the mandate to clear it. All of it is revealing a decision layer that was never built, because nobody stopped early enough to ask the business-need question long enough to notice governance was part of the answer. When content moved slowly, that gap didn't hurt - there was always time to catch things downstream. Speed the process up without ever having built that layer, and every unclear ownership line becomes a visible failure, fast.
The missing decision layer
While knowing what the business needs from the tech is the starting point, turning that into something that actually works day to day means making four more calls.
The first is criteria: What does "good enough" actually mean, for this market, this content type, this level of risk. This needs to be a moving bar, because a routine variant and a first-time claim in a new market don't deserve the same scrutiny, and treating them the same is exactly what turns approval into a bottleneck.
The second is rights: Given those criteria, who holds the authority to sign off without escalating it further up the chain. If every piece still needs the same three people regardless of risk, you haven't built a decision layer, you've just given it a name.
The third is accountability: Who answers for it if a "good enough" call turns out to be wrong. This is easy to skip, because it feels obvious - whoever has the right to approve it must also own it. But in practice the two get separated more often than you'd expect, and once they are, people quietly stop using the authority they've been given and escalate anyway, just to be safe. A decision layer only works if the person with the right to decide is also the person who carries the outcome.
The fourth is speed: How fast does all of this have to move to keep pace with what the technology can now produce. This one gets forgotten often, because it was never a live question when content moved slower. If not addressed, you’re left with a decision process sized for a world that no longer exists.
Skip any one of the four, and the others fail. Get criteria right with no one empowered to apply them, and you've written a document nobody uses. Get rights and accountability right with no criteria, and you've just moved the guessing downstream, from the tool to the approver. Get all three right but too slow, and you've relocated the bottleneck instead of removing it. This is why organisations can do this work in good faith, one piece at a time, and still end up back where they started.
A gut-check before you scale
Before greenlighting AI-driven content production at scale, it's worth forcing these in order, and revisiting them, because they won't stay answered:
What do we actually need this to do, and for whom? Based on what the business needs AI to deliver, not what AI is capable of.
What does "good enough" mean, for what? With different criteria for a routine variant versus a first-time claim in a new market.
Who holds the authority to sign off without escalating? For each level of risk you defined in question 2.
Who answers for it if a "good enough" call turns out to be wrong? If the answer is different from question 3, you don't have accountability, you have exposure.
How fast does all of this need to move? Mapped to the speed the technology now makes possible.
Bonus question if you've already started:
If something isn't working, is it really the technology? Before reaching for a better model or pipeline, check whether the real gap is one of the five questions above that was never properly answered.
The technology will keep expanding what's possible, that's not going to slow down. The organisations that get real value from it won't be the ones with the best generation capability alone. They'll be the ones who did the less exciting work of deliberately deciding what they actually needed it for and built the decision layer to match - before the gap they left behind came back looking like a tech problem.
