The Guardrails Of Imagination

Watch the session

The setup

AI arrived inside production pipelines faster than the paperwork that governs it. Agencies are shipping synthetic voices, generated environments and machine-assisted edits into paid media right now, while the consent forms, the union addenda, the indemnity clauses and the disclosure rules are still being drafted behind the work. The guardrails are being built after the car left the garage.

That is the problem this session was assembled to answer, and the panel did not pretend otherwise. One of them told a story about needing a SAG addendum for an AI voice model that did not exist yet, so the agency sat down with SAG and wrote one. Another argued that the question itself was wrong, that you do not need guardrails for AI so much as you need the ones you already have for everything else. Nobody on stage claimed to have a finished policy.

What they had instead was five people who have already put AI work into market at national scale and can describe, in specifics, what it cost, who signed off, and where it nearly went wrong. For an independent shop being asked by a client this quarter to "use AI on this," that is the more useful hour.

Who was on stage, and why they're worth listening to

Part One — The cost fiction

The commercial pressure on every independent in the room is the same. A client reads that AI collapses production cost, and arrives with a budget that assumes it. Melanie put the question to the full panel: when clients assume production costs need to drop to near zero, what is the counter-offer that proves high-craft AI work requires a premium?

Miguel had already answered it an hour earlier, unprompted, when asked what gets lost. His answer was that the premise of effortlessness is wrong.

That is a positioning statement, not a complaint, and it is the sharpest pricing move on the panel. He then named what the money is buying, which is the part most agencies fail to articulate on an estimate.

He also described a process change that follows from it: "In our agency, we say we don't pitch anymore; we do prototypes. Our pitch is something that you can work with, watch it, play with it." Note what that does to a scope of work. If the pitch is a working artifact, the cost has moved forward in the timeline rather than disappeared.

Jen's version of the same argument comes from inside a production department, and it is the line most likely to survive this session.

She was specific about which people. The director, the production designer, the cinematographer. Those roles do not vanish in an AI pipeline; they become the thing that distinguishes the output from everything else being generated that week. "That's the human element that gives the AI taste. Because we all know the AI slop we're seeing out there is those off-the-shelf models, and it is a lack of taste."

Miguel described the same staffing pattern from the other side of the Atlantic. On AI films his teams bring in an AI DOP, someone trained in cinematography rather than in prompting, who understands framing, composition and lighting, and pairs with a stronger prompter. Two skills, two people, one shot.

Then Jen gave the number that makes the pricing argument land, and it is the number clients do not expect.

Highdive's Jeep work is the proof case. Stellantis knows its own vehicles to the millimetre, and off-the-shelf models could not hold a Jeep consistent from scene to scene. So Highdive built a proprietary system, referred to on stage as Digit, to create controlled AI models of client products. Jen's account of what happened when the client saw it: "It's game over for them. There's no more convincing to be done." The tool was not the sale. The control was.

Johnny added the strategic-foundation argument, and he brought the only efficiency figure quoted in the session: an average 60% efficiency in cost per asset across Silverside's clients. His point was that the number is meaningless on its own.

There is a second cost most estimates miss, and he named it plainly: tuning. An untuned model migrates toward the median, which is where every competitor's output already lives. Fine-tuning a system around a brand's point of view is R&D, it takes time, and somebody pays for it.

Miguel's contribution to that thread was the most useful thing an agency can say to a client who wants a guarantee.

He anchored it in a partnership he says was announced days earlier, between Google and A24, structured around R&D for filmmaking tools rather than around a film. No deliverable, no campaign. Just tooling. If the largest players are funding discovery, an independent charging for discovery is on defensible ground.

The tooling guardrail: do not marry a model

The single most repeated operational instruction of the hour came from Johnny, and it has a compliance dimension the room did not immediately see.

His reasoning was twofold. First, coverage. In his experience a single platform solves 50 to 70% of a given problem, which leaves practitioners going elsewhere for the remaining 30 to 50% and bending their workflow around a partial fit. Second, decay: "Every 30, 60, 90 days it'll be leapfrogged. Something else will be better."

Silverside's stated position is model agnosticism and modular builds, plugging APIs and MCPs into an environment built around the practitioners who will use it. This is a company position as much as a personal one, and he says so. But the adoption logic underneath it is portable to any shop, and it is the strongest labor argument in the first half:

Melanie then asked the question that separates a conference panel from a procurement meeting: how do you get legal and compliance to sign off on a modular, multi-model stack when AI policies are strict and cybersecurity is a live issue?

Johnny's answer was the most concrete guardrail material in the session. When Silverside works with model providers as partners, he described negotiating a set of protections, and two of them are the ones every independent should be asking about before they put a client's material into a system.

Data protections. "Making sure that what you're putting into the models is not going to train the underlying model; it's only going to make your instance better within a walled garden."

Indemnities. "No one wants to get sued. As long as you're not taking someone else's IP and putting it into the system as your own, then we're able to build those indemnities in a really robust way that allows you to operate as a brand with a lot more creative freedom."

The conditional in that sentence is the guardrail. The indemnity holds because of a hygiene rule about inputs. Break the rule and the protection goes with it.

He then made a claim about leverage that most independents will find counterintuitive, and it is worth testing:

His explanation is that the model market is a land grab, so providers will trade early access, roadmap influence and sometimes project funding for reference use cases. He is candid that he has a commercial interest in the conclusion. The underlying observation, that the negotiating position is better than agencies assume, is checkable and free to test.

Part Two — Who does the work now

Melanie's hardest question was aimed at Lautaro and Miguel together, and it had two halves. If automated pipelines can spin up millions of customized assets faster than anyone can track, how do you audit for ethical bias and brand safety? And if machines absorb high-volume execution, are you destroying the apprentice-style playground where junior creators build foundational skills, especially when those juniors are the ones losing jobs?

The panel answered the second half thoroughly and never came back to the first. That is worth naming. Nobody on stage described an audit process for bias or brand safety at scale, and the question stayed on the table for the rest of the hour.

On labor, Lautaro started by conceding the loss before making his case. He said the job displacement is coming and he thinks about it constantly. His counterweight is that production used to be gated by access to equipment, crews and budget, and that gate has moved. He described a fully AI-executed campaign where the casting expertise came from someone out of the fashion industry with no advertising background, and a separate hire where a candidate with a degree in art history turned out to be an exceptional prompter. His framing: the expert categories are being redefined, and the redefinition creates entry points that did not exist.

Miguel took the same position further, and past comfortable:

That is an optimistic reading of a compressed career ladder, and it is fair to note the ladder compresses in both directions.

The counter-argument arrived from the floor. Baptiste, an ECD at Real Chemistry, made the sharpest structural observation of the session and he made it as a question. Senior talent, he said, is picking AI up quickly, plausibly because those people lived through the arrival of the web and watched colleagues who refused to adapt get replaced. The problem is at the other end.

The barrier he identified is $20 a month. Not talent, not curiosity, not aptitude. A subscription.

Johnny's answer inverted the premise and gave the session its most quotable line on labor.

Then the claim underneath it:

His prescription had three parts, and the first one answers Baptiste directly. Buy the subscriptions, and treat significantly more than $20 as worth it, because there is no age or seniority profile for taste. Second, build rigorous training pipelines so juniors learn under someone with creative excellence rather than learning alone. Third, give them room to put ideas in front of people. His observed result at Silverside is that young hires without advertising backgrounds contribute out of proportion to their experience, because "they don't yet know the rules of what you're allowed or not allowed to do."

Jen answered with a policy and a number, which is what most readers will want.

Her diagnosis of the problem that solved: juniors were using AI, getting good results, and were embarrassed to admit it. The permission mattered as much as the tooling. Alongside it, Highdive gives everyone a $100 per month credit to experiment, deliberately month-to-month, because the agency has not committed to a single model. A creative director leads AI internally and runs the training.

Her advice on guardrails followed from the same place, and it is a resourcing instruction rather than a policy one: find the people already staying up late building their own agents, sit with them, and give them the floor.

One panelist added the policy ask nobody in the industry is making loudly enough: the model providers should be issuing student licenses, the way software companies historically have, and the absence of an accessible entry point is a structural problem rather than an individual one. The observation that followed drew the biggest laugh of the session, that the previous generation solved this with cracked Photoshop, a key generator and three viruses on the machine.

The scale question the panel would not concede

Running underneath the labor thread was a disagreement about whether hyper-personalized, machine-generated volume is worth having at all. Miguel argued against his own capability, and flagged that he was doing it.

His example was a football match watched simultaneously by millions, and the belonging that produces, set against an AI recommending you a restaurant. "There is something greater when we can all watch the same."

Johnny arrived at a related conclusion by a different route, preferring localization over personalization as the higher-leverage use, and describing the failure state as brand fatigue: the same ad in front of the same person often enough that consideration drops. Two people whose businesses depend on generative scale, both arguing that volume is the wrong target.

Johnny went further and set a limit on his own product, noting as he did that it was a strange thing for someone at an AI lab to say out loud.

That is the sentence to hand a client who has decided the technology is the brief. Knowing where the tools do not belong is the same judgment the panel spent the hour trying to price, and it carries more weight coming from the person with something to lose by saying no.

Lautaro reframed the whole thing around scarcity. Photography and every making discipline after it were economies of scarcity, gated by a limited supply of people who could do the work well. Generated content removes the ceiling, which relocates the job.

That sentence is the closest the session came to defining the craft guardrail. When anything can exist, the discipline is deciding what should.

The uncomfortable truth

The session's centerpiece is a consent story, and Jen told it as an example of using AI for good. It should be read as both a model and a warning.

Highdive was days from launching a national brand campaign for Jeep. The same voiceover artist had been the brand voice for 10 years. Shortly before the session with him, he had a stroke. He survived, and at the time was unable to perform.

What happened next is the whole argument of this chapter compressed into one production.

Consent first, from him and with his family. Then the part that has no precedent:

The performer approved the output, not only the process. And the compensation held:

Jen reports he has since recovered his speech and mobility, is not as strong as he was, and that the arrangement continues on the same terms.

Read the sequence again, because it is a usable checklist. Consent from the performer. Consent involving the family, at a moment when he was not in a position to negotiate. A union instrument that did not exist, built rather than skipped. Performer approval of the final asset. Pay parity with the in-person rate.

Here is the uncomfortable part. Every one of those steps was optional. The addendum did not exist, which means nothing formally required Highdive to create one. They did it because the people involved decided to, and because SAG moved fast on an unusual request from a shop it presumably trusts. That is a relationship-dependent outcome, and most independents do not have that relationship or that lead time.

The guardrail worked. It worked because individuals chose to build one under deadline pressure, not because a system caught anything. That is a fragile foundation, and it is the honest state of the industry rather than a criticism of the agency that handled it well.

The panel's disagreement about whether that gap needs its own rulebook was the last substantive exchange of the hour. Asked what advice they would give teams setting guardrails this year, Johnny declined the framing and offered a philosophy instead:

His position is that the regulatory environment is catching up on its own, and he cited a New York requirement, dated on stage as June 9, obliging disclosure to consumers when a synthetic performer is used. His instruction to brands is to be as transparent as possible and to build strong internal perspectives on where AI gets used, where it does not, how it gets used, and what gets said to consumers, while starting before any of it is settled. "Consumers are really smart, and starting from a place of honesty is going to make you stand out really well."

A different panelist rejected the question outright. The position, and this attribution could not be confirmed from the recording, was that guardrails specific to AI are a category error:

Both positions were left standing. One says the rules are arriving and you should move before they land. The other says the rules you already hold are sufficient and AI does not deserve a separate ethics. The Jeep voiceover story sits awkwardly between them, since the existing rulebook had no page for what Highdive needed, and somebody had to write one.

What you can use