Not in a keynote. Not in some sanitized LinkedIn carousel with a stock photo of a handshake. He said it on our podcast, during Cyber Week, while the rest of the industry was mainlining espresso and refreshing dashboards like nervous day traders. "I generally don't fear the big, big agencies," Bloom told me. "I actually fear the two gals or the two guys in the garage building something from the ground up. Scarcity creates a lot of creativity."

That's the Head of Partnerships at PMG — an agency with Apple on the roster — telling you the biggest threat to his industry isn't Publicis or Omnicom or whatever WPP is calling itself this quarter. It's two people with a laptop and an API key.

He's right. And in 2026, that's not a metaphor. It's an architecture. This article is the blueprint.

Let's start with the number that should make every agency CEO sweat through their Patagonia vest. "We shouldn't spend 70% of our time on planning, buying, trafficking, reporting, billing," Sam said. "All that stuff sucks and nobody wants to do it. I really worry about the mental health of people in our business."

Seven-zero. That's not an inefficiency. That's a tax. An agency employing 200 people effectively has 140 of them doing work that a well-architected system of agents and APIs could handle before anyone finishes their morning coffee.

And here's what nobody wants to say at the Cannes rosé tent: the tools exist right now. You can wire an LLM into a DSP API, build a trafficking agent that reads an IO and sets up campaigns while you sleep, and automate the pacing-check-and-screenshot-and-email-to-client cycle that is slowly draining the life force out of every 27-year-old media planner in America.

The reason it hasn't happened at scale isn't technical. It's structural. Agencies bill hours. Hours require humans. Automating that work doesn't streamline the agency — it lobotomizes the revenue model. The garage shop doesn't have this problem. It never had the legacy model to protect.

Trafficking and tag QA — every major platform has an API; the reason agencies still do this manually is that the people who could build the automation are in a different department than the people who need it. Pacing and anomaly detection — should have been automated in 2019; an agent can poll APIs every fifteen minutes and escalate only when something actually requires a human decision. Reporting — an LLM that reads platform data and generates a narrative summary with anomaly callouts can do this in seconds; the human value is in the interpretation, not the assembly. Billing and reconciliation — spreadsheet work that's error-prone, time-intensive, and makes people miserable. Automate it.

The garage shop starts with zero legacy workflow. It automates everything automatable on day one. Its three humans deploy against the 30% that machines genuinely can't do yet. That's not a disadvantage. That's a superpower.

Before you run off and tell your board you're building an "AI-native agency," consider the other thing Sam said that should be tattooed on the forearm of every CTO in ad tech. "Everything coming at us, they say they're AI," he remarked. "It's not clear who's really reinventing or who's just pretending. It takes a lot of energy to sort the wheat from the chaff."

He's being diplomatic. Roughly 90% of what's being pitched as "AI" in advertising right now is the same product it was two years ago with a ChatGPT wrapper and a new domain name. The DSP that was "machine-learning-powered" in 2023 is "agentic" in 2026. The logo changed. The product didn't.

Sam has an elegant test for this. "I love it when a company says, here's the problem we're trying to solve," he told me. "When I see that AI can do something you couldn't do before, or make something way easier — that catches my eye. When it's not doing either of those things, generally I'm unimpressed." PMG-polite for "this is garbage."

So if you're building or buying an AI-native stack, here's the five-question filter:

Problem legitimacy. What specific problem does this solve that couldn't be solved before? If the answer is vague ("we optimize performance"), run. If it's specific ("we reduce trafficking setup from four hours to twelve minutes by reading the IO and pushing directly to platform APIs"), keep listening.

Architecture depth. Proprietary model trained on real data, or a generic LLM with a nice interface? Nothing wrong with a good wrapper — but the pricing should reflect it. If someone is charging enterprise SaaS rates for a GPT-4 prompt with their logo on it, that's not a product. That's a markup.

Measurable workflow delta. Show me the before and after. Not a case study written by marketing. Time saved. Error rate reduced. Cost per campaign changed.

Human-override design. When the AI makes a decision you disagree with, what happens? Audit log? Escalation path? Kill switch? "I do not want to remove that human control," Sam said flatly. Any system without explicit human checkpoints isn't AI-native. It's AI-reckless.

Pricing alignment. Seats or outcomes? Outcome-based pricing signals confidence. Seat-based pricing signals a subscription with a chatbot.

The AI-native garage agency isn't a smaller version of a holding company. It's a fundamentally different architecture. You don't get here by firing people from a traditional agency and hoping the survivors figure it out. You start from the workflow and design backward.

Layer 1 — Execution. Containerized microservices, not a monolithic platform with a chatbot bolted on. A trafficking agent that reads an IO, maps line items to platform-specific campaigns, pushes to APIs, and flags anomalies for human review. A pacing agent that polls every fifteen minutes, auto-adjusts within pre-approved guardrails, and escalates when it can't. A reporting agent that generates narrative summaries, not data dumps. A reconciliation agent that matches delivery to billing and only surfaces problems. Each service ships via CI/CD and iterates in days, not months. The garage shop iterates on its own infrastructure the way a software company iterates on product — because it is a software company that happens to buy media.

Layer 2 — Identity and Data Spine. This is the unfair advantage. A central metadata store — the campaign graph — holds structured relationships between accounts, campaigns, creatives, audiences, and outcomes. Every agent reads from and writes to it. Connectors pipe into retail media clean rooms (Amazon Marketing Cloud, Walmart Luminate) where actual purchase data lives, into CTV device graphs for household-level cross-screen stitching, and into first-party brand data for segments built on real behavior. "Retail media — these are signals we just haven't been able to see," Sam said. "Consumption, life changes, seasonality, habitual behaviors. We're just scratching the surface." The holding companies have been trying to build this for years. Most of it is PowerPoint. The garage shop builds it on day one because it doesn't have forty legacy systems to integrate. It has a clean sheet and a cloud account. (We go deep on the identity architecture and how it connects to measurement in our Retail Media Report 2026 — more on that below.)

Layer 3 — Measurement. Sam has explained this approximately nine thousand times: "The biggest challenge we have is winning clients off of last click. It forces clients down a certain path and they miss the narrative." Last-click attribution is a lie. By its logic, your waiter made your dinner. The garage shop can't afford measurement laziness, so its stack has three tiers from day one. Strategic: a lightweight, frequently refreshed MMM that treats retail media as its own channel, fed by real purchase data from Layer 2. Tactical: geo-holdouts and audience splits that validate MMM coefficients before you scale new investments. Continuous: an always-on holdout group so you always know whether the platform's reported 6x ROAS is real or fiction. (It's fiction. It's always fiction.) (Part 2 of this series builds the full measurement operating system. The data behind it comes from our Retail Media Report 2026 — and the numbers should terrify you.)

Layer 4 — Transparency and Governance. Not sexy. Essential. Supply-path audits that restrict buying to direct, verifiable paths. Suppression lists and brand-safety rules the team controls, not the platform. Audit logs for every agent decision — every bid adjustment, every pacing change — timestamped and human-reviewable. And defined human escalation triggers configurable by client, campaign, and risk level. "How do you build systems and processes with humans in the loop?" Sam kept asking. He's right. Today the human-in-the-loop isn't optional — not because the AI can't handle it, but because clients need to trust the system, and trust is built by demonstrating that a person they know is watching the machines. (Part 3 takes this into CTV specifically — where supply-chain opacity and signal degradation make governance the difference between premium inventory and expensive garbage.)

I asked Sam what AI would fail at first if it replaced him tomorrow. He didn't hesitate. "The relationship side, number one." Then he went deeper: "The stories in people's heads — client side, publisher side, agency side — that's what we're all battling. What's between our ears."

This is what every "AI will replace agencies" think piece gets wrong. The hard part of media isn't buying media. It's understanding why a CMO is terrified about their board meeting next week and how that fear will shape every decision for 30 days. It's knowing that a client's CFO doesn't actually care about incrementality — they care about not being embarrassed in front of the CEO. No agent does that. No agent is close.

The garage shop's three people aren't doing less work. They're doing different work: client relationships and strategic counsel, creative direction (not the versioning — the thinking), and judgment calls when the data says something's broken and someone has to decide what it means before the client's morning standup.

The traditional agency has three structural advantages: scale, access, and talent. The garage shop neutralizes all three. Platform APIs don't care how big you are — Amazon doesn't give you a better CPM because you're Dentsu. The walled gardens are walled for everyone, and what matters isn't a "Google relationship" but the technical competence to use the APIs. And talent? The best young people in advertising don't want to spend 70% of their time on trafficking. "The creativity — particularly the technology — is just wild," Sam said of his younger colleagues. "I get a lot of energy from young people in this business."

Those young people are going to build the garage shops. Some already are.

Sam called advertising "the longest-running sitcom." He's right. But every sitcom gets cancelled eventually. The question isn't whether the AI-native garage shop is coming. It's whether the incumbents will adapt before it arrives.

Given that most of them still can't automate a trafficking workflow, I wouldn't bet the ranch.

This is the free preview. The blueprint is one thing. The operating system underneath it is another.

Part 2, "The Measurement & Finance Knife Fight," goes deep on the structural dishonesty baked into retail media measurement — why 53% of executives don't trust their own incrementality numbers yet 74% are increasing budgets anyway, how to build the three-tier measurement OS that replaces last-click with actual math, and the CFO-grade financial narratives that get budget committees nto say yes. It publishes alongside key findings from our Retail Media Report 2026: The $70 Billion Lie — 48,000 words of original research, 52 executive interviews, 100-person survey data, and the most comprehensive analysis of what's broken, who's winning, and what happens when AI shopping assistants skip the whole apparatus entirely.

Part 3, "The CTV Quality & Signals Crisis," unpacks why CTV's supply chain is developing the same MFA-style rot that nearly destroyed the open web — and what an attention-as-diagnostic framework, a quality tiering rubric, and retail-media identity plumbing actually look like when you spec them into IOs and PMP deals.

The full series, the complete report, and the operational frameworks are available to ADOTAT subscribers. The brands building independent measurement, the agencies rearchitecting their stacks, and the RMNs that want to survive consolidation — they're already reading this. The question is whether you're in the room or reading about it afterward.

[Subscribe to ADOTAT →]