Forward Deployed Creative

Every June, ad executives descend on Cannes with the shared goal of getting overserved and selling the future of advertising, ideally at the same time.

I spent most of this Cannes posted up in the lukewarm lobby of the Mondrian doing just that, but this year felt different. 

Whether the person across from me came from a DSP, CDP, DCO, or some other acronym in the advertising industrial complex, the conversation moved quickly to creative: why it matters more, why it is harder to understand, and what to do about it.

That consensus is more important than it may sound: the cottage industry of modern advertising was built around the assumption that creative, audience, and media were separate inputs. 

The creative agency produced creative, the data platform provided audiences, the media agencies ran campaigns, and everyone was one big happy dysfunctional family.

Today, creative influences the platform’s prediction of who is likely to respond, co-producing the audience alongside media and audience decisions. 

TL;DR: the double-espresso/rosé conversations mostly talked about one thing: audiences haven’t disappeared, but creative has an increasingly important seat at the targeting table whether we like it or not.

Our thesis is that whoever makes the interaction between creative and performance outcomes legible will define creative intelligence, much as the last generation of ad-tech companies defined audience and media intelligence.

We will get to what this means for your Meta ads. But first, a history lesson.

Pepper, silver, and the FDE

In April 1601, the East India Company (EIC) sent its first fleet into the unknown in search of pepper. The fleet had some silver and a plan formed thousands of miles from the market the EIC hoped to enter.

More than a year later, the ships reached the Strait of Malacca, but the pepper never came. Historical accounts differ on why. Who cares.

The point is that a plan years in the making ran into local, illegible reality almost immediately. The East Indies already had a massive trade network, and the Portuguese had spent the prior century proving there was money there. But London could not read that opportunity from London.

It wasn’t that trade was simple in Europe. Goods already moved through merchant houses, ports, fairs, guilds, insurers, and overlapping legal regimes across Europe, the Mediterranean, and the Levant. But those corridors had built up enough trust and structure to make the market legible. Routes, counterparties, and rules were understood well enough for specialization.

In the East Indies, instructions took months to arrive, and by then, prices, rulers, supply, and demand could all be different. A ship could reach the market, but each voyage still basically started from zero.

So the EIC built a forward-deployed system. Its “factories” were permanent trading posts staffed by factors: part merchant and part operator. They negotiated with local merchant communities, maintained relationships, and sent intel back to London. London was still responsible for broad strategy, but the factors applied it to local realities.

What one factor learned did not disappear when the ship left port; it moved through the network, informed the next decision, and gradually made the EIC smarter. The market remained messy, but the EIC built a system capable of operating inside what it could not initially understand.

Four centuries later, Alex Karp’s most important contribution to enterprise software was reaching a similar conclusion: frontier systems cannot be built entirely from point-solutions because defining the work is part of the work.

Palantir’s answer was the forward-deployed engineer. FDEs embed with customers, learn the mess of the actual operating environment, and configure the software around the objects, relationships, and decisions that matter there. The field work discovers the right nouns and verbs, and the software makes them reusable.

Like the EIC’s factors, the FDE sits between a centralized system and a unique local one it can’t fully understand from a distance. Software can solve pieces of the problem, and consultants can point out what’s broken (expensively), but neither closes the loop. The FDE does the unscalable work of getting in the weeds directly with the client, then applies the right mix of software, process, and past learnings to actually solve problems end to end. Those new learnings then feed back into the system and improve future deployments.

Without the software, the field work does not compound and without field work, the software codifies around the wrong thing.

For most of modern advertising, life looked a lot more like European trade: complicated, yes, but legible enough for everyone to stay in their lane.

Advertising’s legible era

The first major digital-advertising advantage came from making audiences legible. Cookies, IDFAs, third-party data, your mother-in-law’s Facebook profile, etc turned groups of people into selectable objects. An advertiser could decide whom it wanted to reach, then buy access to that audience.

Programmatic buying, campaign structures, bid strategies, and media-mix tools then made media buying similarly legible. Meta ads manager has never been simple, but historically the things determining the outcome were visible. You’d select an audience, choose an objective and bid type, and throw in some creative. This is an oversimplification, but you get the point.

That structure supported a cottage industry of point solutions. A media agency could manage campaigns, a creative agency the creative, the data provider could supply audience intelligence, an attribution platform could measure conversions, and a SaaS could produce DCO variations.

The proliferation of interest graphs (read: TikTok) marked the beginning of the end(ish) for prescriptive audience selection. Most walled gardens today index on broad targeting, automated placements, and campaign solutions that hand more of the decision-making to black box models. We write more on that shift here.

This is increasingly the standard: advertisers supply an objective, a budget, and creative. That’s it. The platform watches how users respond to that creative, predicts who else is likely to respond, and allocates impressions accordingly.

Effectively, creative is now part message, part targeting input, and part experiment.

“Creative is the new targeting” is decent shorthand (we were among the first to say it) but it’s inconclusive. The creative changes the platform’s prediction of whom to reach; the audience it finds then produces performance signals which are used to change targeting iteratively, and this data is then used to target future creative. Two ads in the same campaign can reach meaningfully different populations, even with the same targeting settings.

For a long time, brands attempted to define the minutia of brand identity because they had to. But today, the brand is defined in large part by the consumer. We still see resistance to this, but think about it: do you really think you know your customers better than the platform watching them react in real time? As Eugene Wei put it, “The algorithm knows.” 

The brands that understand today’s attention economy will ultimately win. For-you style algorithms are the most liquid attention markets there are. The right message will reach the right users, but first, you need to make the right creative.

Creative data is illegible

Advertisers are actaully drowning in creative data: platform performance metrics, competitor ads, scripts, useless third-party dashboards, comments, etc. What is scarce, however, is the ability to connect that data to its impact and then execute on those learnings.

This is because creative performance is downstream of so many things:

1.     The message a customer sees.

2.     The audience the platform picks based on the creative

3.     The creative quality itself

4.     The market demand at that level of scale

5.     So much more

Ads Manager collapses all of that into a CPA or ROAS and asks everyone to carry on. Broadly, this CPA is the “creative-market” fit at a certain scale but the components are illegible.

So when an ad “wins,” what exactly won? The message may have been more persuasive. The platform may have found a warmer audience. The auction may have been cheaper. The conversion may have happened anyway. The customer may have converted and then churned, refunded, defaulted, or otherwise sucked.

Even looking at the creative only and holding everything else equal, creative data remains illegible to most teams. Take a street-interview ad for an AI search product. The interviewer asks three people to choose between Google and ChatGPT. The first two pick one or the other. The third says, “Neither, I use [redacted]” and explains why.

“Street interview” describes the format, but basically none of the substance. The asset is also a challenger narrative, comparison, pattern interrupt, product testimonial, and compressed argument about how search should work. Maybe it won because of the first two answers, the credibility of the third person, the framing, the screen-recorded demo, or the creator looking trustworthy.

Now change the hook, creator, pacing, value props, and CTA at once (as one does). If the new version wins, good luck assigning credit. Unless you have a structured testing system and a clear creative ontology around it, good luck knowing anything other than one ad beat another.

Most teams define creative strategy as seeing what the platform showed love to early on and calling everything else a loser. That’s a good start, and right ~60% of the time in our data. But that gives almost no insight into why something worked, how much better it is than something else, or what to make next. 

None of this makes testing useless. But most paid social “A/B testing” is not really A/B testing. There is no control audience. So there is no real A/B test. It’s apples and oranges. The platform decides who sees what, how much each ad gets, and when. So the dashboard can tell you which ad won, but not what creative decisions drove that win.

That is the creative-intelligence problem: more observations (fancy dashboards) than ever, less confidence about what any one of them means, with control scattered across briefs, platforms, creators and agencies.

The testing-shaped object

Not all things that look useful are useful. Often, they just produce the feeling of work. Will goes much deeper on the “tool-shaped object” here. AI gen content (AIGC)  is a particularly shiny testing-shaped object.

If creative data is illegible, what happens if you add hundreds of generated variants on top? More hooks, more creators, more CTAs, more messages, more things to worry about in ads manager. The result is mostly more noise.

Ten hooks × five creators × three CTAs × four concepts = few hundred ads. Congrats, you have a testing-shaped object. Oh and by the way, each asset needs to reach some stat-sig level of testing spend regardless. So you are out 100k in just testing spend to get maybe a couple of marginal winners.

Volume alone isn’t the answer.

We’re not saying AI cannot make ads. It obviously can. The problem is that creative volume is not a replacement for creative strategy. Each creative is an experiment, and the only way to understand it legibly is to start with feature extraction.

What problem are we testing? What claim? What proof point? What creator? What format? Like an ML model, you need to understand feature importance: which inputs actually explain performance, and which ones are just noise. You need some ontology around those inputs, even if the first version is imperfect. Maybe you start with personas. Personas are usually too vague to matter as much as brands think, but hey, it’s a starting point.

The idea is to test the most obvious creative hypotheses, understanding which features seem to matter, and refining the ontology from there. You cannot really do this with a purely generative process because most brands do not have enough good creative hypotheses.

You need qualitative creative strategy to start somewhere, then patterns emerge, and only then can you build something like creative intelligence. This is why A->Z AI ad workflows fail.

The bottleneck is not production or complexity: we’ve seen very simple ads work very well, rather intelligence.

What the f*ck is Forward Deployed Creative?

Forward Deployed Creative is the operating model for a world where creative, audience, and media are continuously shifting inside platforms the advertiser cannot fully inspect.

Like the factor, the FDC is a local interpretation layer: The factor had letters, ledgers, samples, and merchants. The FDC has Ads Manager, a measurement layer, creative insights, audience intel, and whatever cursed BI stack the company has assembled.

The idea is to go beyond either 1) making ads or 2) handing over ideas for what to make. The FDC goes and does the damn thing, while turning the local mess into something useful at the same time. An FDC does five things:

1. Embed. Learn the product, customer, unit economics, compliance constraints, brand voice, and internal business objectives. What does the product actually do? What can the brand credibly say? What outcome is this business looking to achieve?

2. Represent. Define the local nouns and verbs: concepts, hooks, claims, objections, creators, audiences, offers, and customer quality. Those definitions cannot be dropped in generically. A qualified customer means something very different to a GLP-1 provider vs a neobank.

3. Produce. Turn hypotheses into creative and put them into market. Creative intelligence has a slightly annoying problem: the data does not exist until someone makes the idea and collides it with reality. 
   
   Producing the creative and producing the data are often the same act. And no, panels are not the key to scoring creative. You have a $100B attention machine called the Meta algorithm that will score your creative way better.

4. Interpret. Work out whether an apparent winner is actually statistically significant (we built a full statistical model in the backend that is doing tens of thousands of computations a month). Then connect the platform result back to actual business truth. Each creative is a test of how the world is, so treat it like a scientist would treat their experimental results.

5. Act. Make the next concept, adjusting to the data accordingly. If it worked, double down. If not, refine the hypothesis and try again. 

6. Do it all it again

The challenge, of course, is that there is too much signal for any one human to parse at scale. The top echelon of performance advertisers need thousands of ads a month, and even newcomers are making hundreds. That requires a second brain: a creative operating system that turns thousands of signals into the few that matter.

The FDC’s job is thus not only to take the existing local signal and execute on it, but to identify new alpha that can be applied globally. Local context improves the broader system without pretending every advertiser has the same customer or market. 

That is what makes the model forward-deployed rather than services with the same Claude model everyone else has. FDCs use judgment to find the real opportunities; software makes those learnings reusable. 

Without software, the field work stays bespoke. Without field work, the software is useless.

TLDR: The legacy media stack assumed creative, audience, and media could be managed separately. The platforms combined them. Creative intelligence now requires rebuilding that understanding inside the advertiser, close enough to the market to see what is actually happening.

The idea is to put capable operators close to the market (in this case, creative), give them software that remembers, and let every client engagement make the system smarter.

Welcome to Forward Deployed Creative.