Every AI content tool on the market sells speed: generate a carousel in seconds; produce a month of social posts in an afternoon. But speed without compliance architecture means exposure at scale. Health claims trigger FTC enforcement. Financial guarantees invite scrutiny. A single hallucinated source turns a professional-looking post into a legal problem. And most AI-powered content pipelines have no structured safeguard between what the model produces and what goes out the door.
Arjun Jamwal has been working through that gap firsthand. A Lead Product Manager at Verizon with over a decade of product experience spanning OneTrust, KPMG, EY, and Barclays, Jamwal also runs Good Muncher, a wellness brand focused on reducing ultra-processed food consumption. Wellness is one of the categories where the compliance stakes are sharpest: say the wrong thing about a supplement and the FTC letter arrives before the engagement metrics do. To scale content without scaling legal risk, he built an automated carousel pipeline that handles roughly 90% of production while reserving the remaining 10% for curated inputs, compliance gates, and human approval.
He calls the approach "deterministic friction": intentional constraint points that force human judgment at the moments where automation creates the most exposure. As AI-powered creative production scales across the advertising industry, that kind of deliberate friction is becoming harder to treat as optional. "I really think that deterministic friction is the way to go. That's going to be key going forward," says Jamwal.
The pipeline is designed so that no generated content can reach review without passing through three layers of human-curated constraint: verified source data, persistent brand guidelines, and a compliance gate. Strip any one of those layers out and the system becomes what many AI content tools already are: a generator with no memory of what the brand is supposed to sound like and no awareness of what it is not allowed to say.
It starts with a curated fact database of roughly 200 entries, each manually verified against its source publication. When Jamwal runs the pipeline, he specifies which fact to reference. The system pulls that fact, checks it against the brand's voice and content guidelines, applies a design template, and outputs a carousel image ready for review.
The brand guidelines document was generated using Claude, connected to his existing website to extract voice, tone, and content boundaries. That document serves as a guardrail for every subsequent generation, which means the hundredth carousel carries the same positioning as the first. "The carousel gets automatically generated," says Jamwal. "But the branding guideline file gets plugged into Claude Code with Playwright every time."
The fact curation step is intentionally manual. Jamwal reads through each fact, verifies correctness, confirms the backing publication is legitimate, and skims the source material. "The input data has to be very clean. I read through each one to make sure the fact is correct and backed by the right publication. Not just some fly-by-night source."
Before any generated content reaches human review, it passes through what Jamwal calls the tripwire: a structured list of restricted terms and claim types. The list blocks language like "cure," "safe," unattributed health claims, and financial guarantees. The system scans each generated carousel against it and regenerates if a violation is detected. Nothing ships with a flagged term still in it.
The tripwire exists because generative models have no concept of legal risk. In wellness, where FTC and FDA enforcement around health and supplement advertising continues to tighten, the consequences extend beyond fines to reputational damage that compounds across every piece of content a brand publishes. But the same logic applies in financial services, pharmaceuticals, insurance, and any other category where specific language carries legal weight.
"I don't want to be held legally accountable, and I want to be compliant with all guidelines," says Jamwal. "So I curated a list of words and claim types that the system is not allowed to use. None of that will appear in my pipeline."
The split between human and machine is deliberate. The system handles layout, formatting, visual generation, and template application. The human handles fact selection, source verification, tripwire curation, and final approval. For brands and agencies navigating the trust implications of AI-generated content, the architecture behind the content matters at least as much as the content itself.
Jamwal frames deterministic friction as more than a compliance strategy. He sees it as the only durable defense against model collapse: the feedback loop where models trained on AI-generated data produce progressively narrower, less accurate outputs with each generation. The content gets blander, the claims get vaguer, and the model plays it safer every cycle. "That feedback loop is going to produce worse outputs eventually because the model is going to play it more safe," says Jamwal.
The risk compounds for content teams that rely on AI-generated outputs without embedding domain expertise and brand judgment into the workflow. When every brand draws from the same generative pool without differentiation layers, the result is content that converges toward a generic center, indistinguishable and increasingly unreliable.
Jamwal's pipeline resists that decay because the facts are human-curated, the brand guidelines are derived from an actual website rather than generated from scratch, and the tripwire enforces hard boundaries on what can ship. The 10% human layer is the part of the system that keeps the output defensible as the underlying models change.
Jamwal has launched the tool with a handful of early users alongside his own daily use for Good Muncher's Instagram. "You need to understand what the customer eventually wants. I can't make health claims because the customer is going to think this website is going to cure their disease." As AI-powered creative tools lower the barrier to professional-grade content production and content velocity accelerates across every format, the volume is only as valuable as the guardrails governing what ships.
"AI is going to rewrite a lot of content. And we're going to hit a wall where synthetic data just ruins output. The differentiator is the deterministic friction behind it."
