SmartData Deploys AI Agents to Replace Ad Testing Loops, Lower CAC

The Foresight Era: AI Agents and the Future of Marketing Decisions With SmartData — Ai4

Ai4

Autonomous AI agents shift marketing from passive analytics to direct generative execution, eliminating multi-department testing bottlenecks to scale ad spend while cutting customer acquisition costs.

AI-assisted summary of The Foresight Era: AI Agents and the Future of Marketing Decisions With SmartData — original source. Check the original for full context.

Source published: 2025-08-26

Detailed notes

## Background Published on YouTube by the Ai4 channel on August 26, 2025, this presentation features Armen Rostamian, President of Marketing Analytics at SmartData. According to the session description, the talk examines how AI agents accelerate marketing insights to create predictive growth engines and power foresight-driven decisions across media investment, creative testing, and lifecycle strategy.

## TL;DR Armen Rostamian argues that agentic AI fundamentally compresses the distance between marketing vision and execution by removing corporate friction and technical bottlenecks. He introduces a five-stage analytics maturity curve culminating in generative execution, where diagnostic models directly trigger automated creative production and testing. Through a real-world cybersecurity case study, Rostamian shows how breaking ad creative down into structured metadata tables allowed a brand to break through audience fatigue, expanding media spend by 50% while cutting customer acquisition costs by 10%. He stresses that sustainable AI transformation must build upon solid data foundations and human intuition, multiplying human effectiveness rather than seeking blind automation.

## Atmosphere & style The presentation maintains an authoritative, polished, and pragmatic tone, blending high-level strategic forecasts with rigorous quantitative frameworks. Rostamian presents with steady clarity and executive confidence, systematically dismantling common corporate inefficiencies and replacing AI hype with concrete operational blueprints. Listening to the talk feels like attending a focused enterprise masterclass, transitioning smoothly from big-picture economic shifts to granular, data-driven marketing workflows.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=4s) AI agents enable an exponential productivity revolution in marketing, shifting the enterprise focus from simple cost reduction to organizational alignment and massive operational leverage. - [02:07](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=127s) Agentic workflows compress the gap between strategic vision and execution, removing bureaucratic red tape and the need for rare, polymath 'data unicorns'. - [04:21](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=261s) Marketing analytics is advancing past descriptive, diagnostic, predictive, and prescriptive stages into a fifth frontier: generative operational execution. - [10:12](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=612s) Traditional creative testing traps enterprises in an expensive, multi-departmental 'hamster wheel' that wastes budget and delays campaign iteration. - [12:16](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=736s) AI agents deconstruct qualitative creative assets into structured metadata, enabling marketers to isolate and optimize discrete variables like opening hooks and call-to-actions. - [12:16](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=736s) Deploying agentic creative synthesis to break narrative saturation allowed a cybersecurity brand to scale advertising spend by 50% while decreasing CAC by 10%. - [16:28](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=988s) Autonomous testing engines will eliminate budget waste by pre-screening tests with low mathematical probability of success and customizing core concepts across multi-channel algorithms. - [16:28](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=988s) High-performing AI workflows must be anchored in human intuition and deep customer empathy rather than unguided, blind automation. - [18:40](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=1120s) Achieving true agentic decision-making requires sequentially mastering foundational data infrastructure, process automation, and machine learning without taking shortcuts. - [18:40](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=1120s) Enterprise AI architecture must be modular and future-proofed so that new foundational model releases increase leverage rather than breaking existing workflows. - [19:50](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=1190s) Unlike deterministic channels such as direct mail, algorithmic ad platforms rely on opaque distribution systems where AI optimizes campaign targeting rather than guaranteeing deterministic delivery.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=4s) The Exponential Productivity Revolution in Marketing

Armen Rostamian, President of Marketing Analytics at SmartData, opens by outlining his firm’s core mission: enabling organizations to make sharper, more grounded marketing decisions by meeting them wherever they happen to be on their data journey. The central thesis of the talk, however, focuses on how artificial intelligence use cases and autonomous AI agents can unify fragmented teams and unlock superior business outcomes across enterprise marketing departments. To frame the scale of this shift, Rostamian highlights statements from two tech leaders: Sam Altman’s prediction that the tech ecosystem will soon witness a one-person organization achieve a billion-dollar valuation, and Jensen Huang’s forecast that AI will mint more millionaires over the next five years than the first two decades of the commercial internet combined.

In Rostamian's view, these two observations converge on a single fundamental truth: individual and organizational productivity can now scale exponentially. This transition is not merely about mechanical cost-cutting or appearing tech-savvy; it is about leveraging AI to forge organizational alignment. Looking at current benchmarks, industry leaders like Nvidia and Netflix represent the highest productivity per head in corporate history, generating roughly $2.5 million in revenue per employee. As marketing organizations incorporate agentic frameworks, they enter a similar structural revolution where individual decision-makers can multiply their output many times over.

- **Keynote Quotes**: Sam Altman projects the rise of a billion-dollar solo company; Jensen Huang predicts unprecedented wealth generation outpacing the internet's early era. - **Top Productivity Benchmarks**: Nvidia and Netflix currently lead corporate efficiency at approximately $2.5 million in revenue per employee. - **Primary Objective**: Shifting organizational AI adoption from simple cost reduction to strategic leverage and cross-functional alignment.

---

[02:07](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=127s) Closing the Distance Between Strategic Vision and Execution

When marketing organizations embrace advanced AI, the traditional catalog of corporate excuses disappears. The success of an initiative shifts entirely to the novelty and inherent strength of the strategic concept itself. In standard enterprise environments, great ideas often "die on the vine" due to layered corporate politics, interdepartmental red tape, and the sheer number of people chiming in at every juncture. Agentic AI drastically compresses the distance between initial strategic vision and final execution, cutting away unnecessary process so teams can focus purely on high-level strategy.

```mermaid flowchart LR accTitle: Vision to Execution Compression accDescription: Flow showing how AI agents eliminate corporate intermediaries between customer-centric strategy and deployment.

subgraph Traditional["Traditional Bureaucratic Pipeline"] A["Marketer Vision"] --> B["Multiple Departments & Politics"] B --> C["Technical Intermediaries / Data Unicorns"] C --> D["Diluted Customer Output"] end

subgraph Agentic["Agentic Streamlined Pipeline"] E["Customer-Centric Marketer"] --> F["AI Agent Network"] F --> G["Direct High-Velocity Execution"] end ```

Simultaneously, agentic systems eliminate steep technical barriers. Historically, marketing departments had to hunt for rare "unicorns"—individuals who possessed deep customer empathy while simultaneously mastering statistical regressions and software development. With autonomous agents handling data parsing, predictive modeling, and execution, the person sitting closest to the customer can make direct, high-impact decisions. Rather than demanding technical polymaths, organizations can empower customer-focused marketers to operate with exactness, freeing up cognitive headspace to improve quality rather than merely chasing operational efficiency.

---

[04:21](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=261s) Moving from Analytics to Operations: The Five Frontiers

To understand where AI agents fit, Rostamian reviews the classic analytics growth curve, which organizations typically climb across four historical stages: descriptive, diagnostic, predictive, and prescriptive. In descriptive analytics, an organization simply tallies metrics, such as reporting total daily sales revenue. Moving into diagnostic analytics, the team isolates the underlying mechanics, discovering that sales revenue is a direct function of website traffic and conversion rates. With those mathematical relationships understood, the predictive phase forecasts future outcomes, projecting tomorrow’s revenue from expected traffic volumes and conversion rates. Finally, prescriptive analytics recommends resource allocation, advising teams to invest in optimizing the sales funnel to lift conversion rates from 2% to 3% for a calculated return.

While organizations cannot bypass these foundational steps, AI provides a competitive edge by shifting the entire curve both upward and to the left—accelerating how rapidly teams advance and lowering execution friction at each step. Yet the most profound transition is that AI turns analytics into pure operations through a fifth frontier: generative execution. Instead of ending with a static prescriptive recommendation to overhaul a sales funnel, generative AI actively produces and deploys the solution, translating analytical intelligence directly into automated operational output with fewer resources.

```mermaid flowchart TD accTitle: The Extended Analytics Growth Curve accDescription: Five stages of marketing analytics maturity ending in generative autonomous operations.

subgraph Analytics["Traditional Analytics Maturity"] D["1. Descriptive: What happened? (Count sales & visits)"] DI["2. Diagnostic: Why did it happen? (Conversion & traffic drivers)"] P["3. Predictive: What will happen? (Forecast revenue & CAC)"] PR["4. Prescriptive: What should we do? (Invest in funnel upgrades)"] end

subgraph Operations["Autonomous Execution"] G["5. Generative: Direct automated execution & creative synthesis"] end

D --> DI --> P --> PR --> G ```

---

[07:15](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=435s) Creative Performance: Tackling the Qualitative Bastion of Marketing

To demonstrate how generative operations work in practice, Rostamian turns to the most qualitative, subjective domain of marketing: creative performance. While quantitative platforms and delivery algorithms mediate the space between brands and consumers, evaluating whether an ad creative resonates has traditionally relied on subjective opinion. Creative departments and external influencer networks bring vital creative instinct, but their workflows rarely operate with scientific or methodical precision.

Rostamian walks through a representative scenario: an enterprise spending $1,000,000 across its media portfolio with a six-month return horizon. Through descriptive tracking and machine-learning attribution models, the growth team diagnoses that click-through and conversion rates are dropping sharply on Meta ads. Feeding this data into a Media Mix Model (MMM)—a statistical predictive engine—reveals that the account has hit diminishing marginal returns on its current ad formats, causing customer acquisition costs (CAC) to spike. While the mathematical model might prescribe slashing ad spend to protect margins, business growth demands a different path: diversifying ad themes and refreshing creative concepts. In the generative era, these predictive signals no longer trigger slow agency brainstorms; they feed directly into generative systems that turn prescriptive diagnostic data into fresh creative inputs.

- **Scenario Profile**: Enterprise spending $1,000,000 with a six-month payback window facing audience fatigue. - **Diagnostic Signal**: Declining click-through and conversion performance isolated to Meta paid campaigns via ML attribution. - **MMM Finding**: Ad spend is past the optimal efficiency point on the degradation curve, driving CAC upward. - **Strategic Fix**: Creative diversification, converting predictive recommendations into generative AI prompts.

---

[10:12](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=612s) Breaking the Traditional Creative Hamster Wheel

In an ordinary enterprise, attempting to solve a creative fatigue problem triggers a bloated, multi-layered workflow fraught with delay and defensive corporate posturing. The analytics team hands down their performance findings, and the brand team often responds defensively, arguing their current concepts are sound. Once convinced to run a test, the brand team drafts a brief, introducing the first layer of human interpretation. That brief is passed to outside creative agencies or user-generated content (UGC) networks, who interpret the brief through their own subjective lenses and produce a batch of ad assets.

```mermaid flowchart TD accTitle: The Six-Layer Traditional Creative Hamster Wheel accDescription: Sequential handoffs showing failure points across five departments in legacy creative testing.

A["Analytics Team: Delivers diagnostic fatigue insight"] --> B["Brand Team: Defensive review, conceptualizes test brief"] B --> C["Creative Agency / UGC: Subjective interpretation & asset production"] C --> D["Internal Reviewers: Selects 'best' creatives subjectively"] D --> E["Media Buying Team: Configures audiences & runs campaigns"] E --> F["Finance Team: Reviews & approves incremental testing budgets"] F -.->|"Months later / $50k invested / 6 failure points"| A ```

Those assets return to internal committees, where individuals pick the "best" variants based on personal taste. The selected assets move to media buyers, who attempt to match them to target audiences, and finally to finance to secure incremental test budgets. This legacy loop spans five corporate departments, navigates six distinct communication layers with six potential points of failure, consumes upward of $50,000 in capital, and takes months to run. By the time results trickle back, the team is stuck on an exhausting operational "hamster wheel," burning capital on incremental tests while leaving real performance to chance.

---

[12:16](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=736s) Standardizing Creative Assets: The Cybersecurity Case Study

AI agents disrupt this fragmented loop by placing management directly into the hands of a lean, focused growth team. The first critical layer of agentic intervention involves standardizing the creative asset itself—converting what was once a purely qualitative review into structured, granular data. Instead of relying on a human reviewer’s impression that an influencer fits a specific demographic, computer vision and language agents catalog every micro-component of an ad into an exhaustive metadata table. The agent records precisely what occurs in the first three seconds (for instance, a female creator in a domestic setting introducing a cybersecurity hook), details how the product is pitched, and evaluates whether the closing call to action is explicit.

By deconstructing videos into discrete, quantifiable variables, marketers can isolate what truly drives performance:

$$\text{Creative Performance} = f(\text{Hook}_{0\text{-}3s}, \text{Narrative Framework}, \text{Audio Tone}, \text{CTA Format})$$

If the opening three-second hook performs exceptionally well but conversions stall due to a muddled five-second conclusion, the team can hold the working variables constant while isolating and replacing only the defective elements.

Rostamian illustrates this through a cybersecurity client experiencing plateauing growth. SmartData's agentic analysis showed that nearly all of the brand's ad spend was concentrated on a single repetitive narrative: "robocalls are annoying, and we can stop them." While initially successful, continuing to pour capital into this single concept pushed the company down a steep efficiency degradation curve.

Instead of waiting for an agency brainstorm, an AI agent identified the saturation point and synthesized a net-new thematic concept: the "vanish from the internet" angle. This theme repositioned the product around complete digital footprint removal rather than everyday call filtering, appealing to an entirely untapped audience segment. The brand deployed the agent-derived creative, scaling total advertising spend by 50% while simultaneously reducing customer acquisition cost (CAC) by 10%. The system bypassed the corporate hamster wheel entirely, validating and scaling a novel concept with minimal friction and zero organizational debt.

- **Cybersecurity Scaling Results**: Spend increased by 50% alongside a 10% CAC reduction. - **Narrative Shift**: Transitioned from saturated "stop robocalls" ads to an expansive "vanish from the internet" positioning. - **Analytical Mechanics**: Quantified qualitative video elements into a structured metadata table, isolating successful variables from failing ones.

---

[16:28](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=988s) Rewriting Campaign Testing Operations and Preserving Human Foundations

Looking ahead, Rostamian emphasizes that agentic workflows will fundamentally reshape the operational mechanics of campaign and testing strategies. Modern multi-channel marketing demands that core conceptual themes be customized for distinct platform algorithms; an ad concept must be expressed differently on TikTok than on Facebook while maintaining narrative cohesion. Generative engines excel here by taking a single validated concept, such as the "vanish from the internet" theme, and adapting it into platform-native variants. Soon, individual creative components produced by human talent will feed directly into generative assembly engines rather than being shipped straight to the end viewer, allowing AI to stitch together optimal, bespoke combinations dynamically.

On testing strategy, Rostamian notes that the single largest drain on corporate marketing budgets is running tests that have zero mathematical probability of success. AI agents eliminate these bad tests before they launch by assessing risk, audience overlap, and historical variance. In the near future, organizations will rely on an autonomous, continuously updated testing calendar that orchestrates experimentation risk and opportunity without manual upkeep.

Rostamian closes by grounding this technological shift in reality: every layer of AI capability is fundamentally built on top of non-AI knowledge. True competitive advantage does not stem from blind automation, but from anchoring agents in human understanding, domain intuition, and an empathetic grasp of the customer experience. Marketers must first understand what resonates and why; only then can AI agents scale that intuition into high-velocity operational execution.

[18:40](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=1120s) Laying the Foundation for Agentic AI and Future-Proof Systems

To interact with AI agents in a genuinely productive way, organizations must recognize that there are no shortcuts to success. The progression toward high-impact agentic systems requires completing each prerequisite layer methodically: building strong foundational data infrastructure, establishing robust process automation, deploying machine learning, and only then advancing toward autonomous AI decision-making. Skipping any of these intermediate steps undermines the reliability of the entire system.

```mermaid flowchart LR accTitle: The AI Decision-Making Maturity Curve A["Foundational Data"] --> B["Automation"] B["Automation"] --> C["Machine Learning (ML)"] C["Machine Learning (ML)"] --> D["Agentic AI Decision-Making"] ```

Forward-looking teams must design their architectures so that they "look forward to the next model release, not dread it." Solutions must remain flexible and modular, allowing organizations to seamlessly plug in newer, more capable foundational models to increase operating leverage rather than fearing that updates will break fragile workflows. Ultimately, the true goal of enterprise AI is not to replace people, but to "multiply the human effect," bringing brands closer to a genuine human-to-human connection with their customers while stripping out the operational noise that burdens internal teams.

- **Non-Negotiable Progression**: True AI decision-making rests on the sequential execution of data foundations, automation, and machine learning. - **Architectural Adaptability**: Systems should be built to gain leverage from future frontier models rather than becoming obsolete when new iterations drop. - **Human-Centric Amplification**: AI agents serve to remove procedural friction, resulting in improved experiences for both internal marketing teams and end customers.

---

[19:50](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=1190s) Q&A: Precision Targeting and Algorithmic Ad Delivery

Addressing an audience question on marketing delivery and prevention, the discussion turns to the difference between direct channels and algorithm-driven advertising platforms. In direct marketing channels, such as physical mailers, brands have precise, deterministic control over audience targeting and delivery.

In contrast, digital platforms like Meta operate on opaque internal bidding and distribution algorithms that make their own real-time delivery choices. While predictive AI agents and data signals can significantly guide and optimize campaigns to make them far more targeted, advertisers must acknowledge that algorithmic ad delivery cannot guarantee an absolute one-to-one match in the way deterministic channels do.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=4s) meeting them wherever they happen to be on their data journey — Armen Rostamian

> [02:07](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=127s) die on the vine — Armen Rostamian

> [04:21](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=261s) shifting the entire curve both upward and to the left — Armen Rostamian

> [10:12](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=612s) hamster wheel — Armen Rostamian

> [12:16](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=736s) robocalls are annoying, and we can stop them — Armen Rostamian

> [12:16](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=736s) vanish from the internet — Armen Rostamian

> [16:28](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=988s) built on top of non-AI knowledge — Armen Rostamian

> [18:40](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=1120s) look forward to the next model release, not dread it — Armen Rostamian

> [18:40](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=1120s) multiply the human effect — Armen Rostamian

## Mentioned Armen Rostamian, SmartData, Sam Altman, Jensen Huang, Nvidia, Netflix, Meta, TikTok, Facebook, Ai4

## Video description Armen Rostamian, President, Marketing Analytics, SmartData

Abstracts

What if you could act on insights before the signal even hits your dashboard?

AI agents can accelerate marketing insights—shifting us to predictive growth engines powered by collective intuition.

I’ll share real-world examples of how forward-thinking teams are deploying agents for media investment, creative testing, and lifecycle strategy—grounded in foresight and driven by automation.

Join me to discover how the next era of marketing is being built on decisions made ahead of the curve.

[Source](https://www.youtube.com/watch?v=dJs2Dk7ARwk)

中文简报

SmartData部署AI Agent替代广告测试循环，降低获客成本自主AI Agent推动营销从被动分析转向直接的生成式执行，消除跨部门测试瓶颈，在降低获客成本的同时扩大广告支出规模。 Armen Rostamian认为，智能体AI（Agentic AI）通过消除企业内部摩擦与技术瓶颈，从根本上缩短了营销构想与执行之间的距离。他提出了一个以生成式执行为顶峰的五阶段分析成熟度曲线，在此阶段，诊断模型可以直接触发自动化的创意制作与测试。通过一个真实的网络安全案例，Rostamian展示了将广告创意拆解为结构化元数据表如何帮助品牌打破受众疲劳，在广告支出增长50%的同时将客户获取成本降低10%。他强调，可持续的AI转型必须建立在坚实的数据基础和人类直觉之上，其核心是成倍提升人的效能，而非盲目追求完全自动化。 AI Agent 正在推动营销领域的指数级生产力革命，促使企业将重心从机械降本转向战略杠杆与跨部门协同，对标英伟达与 Netflix 等人均年创收约 250 万美元的高效组织模式。 (0:04) Agent 工作流大幅压缩了战略构想到最终执行的路径，消除了对复合型“数据全才”的依赖，并将营销分析推向第五阶段——直接输出并部署解决方案的生成式自主运营执行。 (4:21) 通过将定性创意视频拆解为结构化元数据，AI Agent 打破了耗资数万美元的跨部门测试死循环；某网络安全品牌借此打破叙事饱和，在广告投放增加 50% 的同时实现获客成本（CAC）降低 10%。 (12:16) 未来的自主测试引擎可提前拦截数学胜率极低的无效测试并跨渠道自适应生成创意，但高绩效的 AI 系统必须扎根于坚实的数据基建与人类对客户的同理心洞察，而非盲目自动化。 (16:28)

Reading summary

Armen Rostamian, President of Marketing Analytics at SmartData, argues that autonomous AI agents enable marketing departments to compress the gap between strategic vision and execution by transforming traditional analytics into automated, generative operations. Compressing Strategy and Democratizing Execution (2:07) Rostamian frames the current wave of enterprise artificial intelligence as a structural productivity revolution rather than a simple cost-cutting exercise. Referencing forecasts by Sam Altman regarding future billion-dollar solo enterprises and Jensen Huang on unprecedented wealth creation, he highlights current corporate productivity benchmarks set by companies like Nvidia and Netflix, which generate roughly $2.5 million in revenue per employee. In marketing organizations, agentic workflows are intended to deliver comparable operational leverage. Historically, campaign execution stalled across layers of corporate politics, interdepartmental approvals, and dependencies on rare cross-functional employees who possessed both customer empathy and advanced technical or statistical skills. Rostamian asserts that agentic AI compresses this pipeline by delegating data processing, modeling, and operational mechanics to autonomous agent networks. This shift allows marketers sitting closest to the customer to implement high-level strategy directly without navigating corporate intermediaries or relying on specialized technical polymaths. The Shift to Generative Operations (4:21) Traditional marketing analytics maturity follows four progressive stages: descriptive (tallying past performance, such as sales revenue), diagnostic (identifying functional drivers, such as conversion rates and site visits), predictive (forecasting future outcomes and acquisition costs), and prescriptive (recommending budget or funnel adjustments). Rostamian emphasizes that organizations cannot skip these foundational layers, but AI accelerates progression across the curve. The critical transition introduced by AI agents is the fifth frontier: generative execution. Rather than concluding with a static prescriptive recommendation—such as advising an optimization of a sales funnel to lift conversion rates from 2% to 3%—generative systems directly produce and deploy the operational solution. In scenarios where Media Mix Models diagnose diminishing marginal returns and creative fatigue across digital ad platforms, predictive signals can be translated immediately into generative prompts and active campaign adjustments rather than slow, manual agency reviews. Deconstructing Creative Assets with Structured Data (12:16) Rostamian identifies creative performance as marketing's most subjective operational bottleneck. Conventional enterprise creative refreshes operate in a fragmented loop spanning five departments, six handoff failure points, upwards of $50,000 in testing capital, and months of administrative delay. This process often leaves performance gains to chance while internal reviewers evaluate concepts based on personal taste. AI agents address this inefficiency by converting qualitative creative assets into structured metadata tables. Computer vision and natural language models deconstruct video advertisements into quantifiable variables, including opening three-second hooks, narrative structures, audio tones, and closing calls to action. By evaluating these elements independently, teams can isolate and iterate on defective segments while keeping effective components intact. In a cited case study for a cybersecurity client, this metadata analysis revealed creative saturation in a legacy 'stop robocalls' campaign. An AI agent synthesized an alternative 'vanish from the internet' positioning, allowing the brand to scale advertising spend by 50% while reducing customer acquisition costs by 10%. Foundational Prerequisites and Algorithmic Realities (18:40) Deploying agentic systems requires completing non-negotiable intermediate milestones: establishing clean data infrastructure, building workflow automation, deploying machine learning, and finally introducing autonomous agentic decision-making. Skipping foundational stages compromises system reliability. Furthermore, Rostamian notes that agents can eliminate corporate budget waste by evaluating audience overlap and historical variance to block testing campaigns that lack statistical probability of success. Addressing audience questions on campaign delivery, Rostamian clarifies the distinction between deterministic channels and modern algorithmic advertising platforms. While direct marketing channels like physical mailers permit absolute, one-to-one delivery control, digital platforms such as Meta rely on proprietary internal auction algorithms. While predictive agents and structured data can guide and optimize campaign inputs, advertisers cannot force deterministic delivery on algorithmic ad networks. Rostamian concludes that enterprise AI systems should be built modularly to leverage future model iterations without disruption, reiterating that durable operational leverage remains fundamentally dependent on non-AI human intuition, domain expertise, and customer empathy.

精炼阅读摘要

SmartData营销分析总裁Armen Rostamian指出，企业营销正在跨越传统的分析咨询阶段，迈入由AI代理驱动的“生成式执行”时代；通过将定性创意资产拆解为结构化数据并压缩跨部门流转链路，组织能够在保留人类直觉与同理心的前提下，实现决策与执行效能的指数级倍增。压缩执行链路与人效规模化跃升 (0:04) Rostamian援引OpenAI首席执行官Sam Altman关于未来将诞生“十亿美元级单人公司”的预测，以及英伟达首席执行官黄仁勋关于“未来五年AI创造的百万富翁将超过商业互联网前二十年总和”的判断，认为个体与组织的生产力正在迎来指数级扩张。当前英伟达与奈飞以人均约250万美元的营收领跑企业效率基准，而营销部门引入代理型AI框架的核心价值，不在于机械式的削减开支或技术表象，而在于通过工具重构实现跨部门协同并提供战略杠杆。在传统企业环境中，优质策略常因部门政治、繁杂审批以及各环节的冗余介入而流产。以往营销部门不得不依赖兼具客户共情、统计回归与软件开发能力的复合型“全能人才”（unicorns）；而自主代理能够直接承担数据解析、预测建模与具体执行任务。这大幅压缩了战略构想到终端落地之间的路径，让最贴近客户的营销决策者能够敏捷、精准地采取行动，将团队认知资源从繁琐的流程内耗转移到提升战略本身的质量上。从分析进阶到生成式自主运营 (4:21) 传统营销分析通常沿袭四个阶梯式演进阶段：描述性分析（统计过去发生的事实，如汇总销售额与访问量）、诊断性分析（解析深层因果关系，如明确转化率与流量对销售额的驱动机制）、预测性分析（推演未来趋势，如根据流量与转化率预测明日营收）、以及规范性分析（提出资源配置建议，如建议优化漏斗将转化率由2%提升至3%）。Rostamian指出，企业必须夯实这些基础，而AI的真正价值在于推动分析曲线向“左上方”位移，并开创了第五个前沿——生成式执行（generative execution）。生成式执行将静态的分析建议直接转化为自动化的运营产出。例如，当一家营销预算为100万美元、投资回报周期为六个月的企业，通过机器学习归因与媒体组合模型（MMM）诊断出Meta广告渠道出现受众疲劳、边际收益递减且获客成本（CAC）上升时，传统流程往往停留在削减预算或召开漫长的策略研讨会；在生成式时代，预测模型给出的诊断与应对建议能够直接输入生成系统，自动生成并部署差异化的创意素材，完成从分析智能到运营动作的无缝闭环。创意资产微观量化与破除仓鼠轮 (10:12) 在传统企业运作中，应对广告疲劳往往会陷入低效的“仓鼠轮”：分析团队发现效果衰减后，品牌团队容易产生防御抵触，随后拟定主观测试简报并交付外部代理商或UGC网络制作；素材产出后需经内部委员会按个人喜好挑选、媒体采购团队匹配人群，最后由财务部门审批追加测试预算。该流程横跨五个职能部门、经历六道人为转接，不仅存在多处沟通断裂隐患，通常还需耗费数月时间与5万美元以上的测试成本，且容易将实际成效诉诸于运气。 AI代理通过将定性创意资产拆解为微观结构化数据打破了这一死循环。借助计算机视觉与自然语言处理，系统能够将视频广告解构为前3秒抓手（hook）、叙事框架、音频基调与行动号召（CTA）等可量化变量并生成元数据表，使团队能够精确锁定衰减环节并实施局部替换。在SmartData服务的一家网络安全客户案例中，该品牌长期局限于“拦截骚扰电话”的单一饱和叙事，导致边际效益急剧恶化；AI代理检测到饱和拐点后，自主构思出“在互联网上隐形”（vanish from the internet）的数字足迹消除新主题，最终推动该品牌在广告支出增加50%的同时将获客成本降低了10%。渐进式技术底座与认知直觉边界 (18:40) Rostamian强调，构建高效的AI代理决策体系不存在任何捷径，组织必须严格按照先后顺序循序渐进：夯实底层数据基础设施、建立流程自动化、部署机器学习模型，最后方可迈向自主代理决策。跨越任何中间阶段都会动摇整个系统的可靠性。此外，企业架构必须具备前瞻性与模块化特征，使团队能够灵活接入未来更先进的基础模型以获取杠杆倍数，而非因模型升级而面临系统瘫痪。在实验层面，AI代理还能结合受众重叠度与历史方差，提前筛除统计学上成功概率为零的低效测试，由自动化测试日历统筹风险与收益。尽管技术演进迅猛，Rostamian提醒营销决策者注意其边界与基石：所有AI能力的核心底层均建立在“非AI知识”之上，竞争优势并非源自盲目自动化，而是取决于人对行业逻辑的深度认知与对客户心理的同理心。此外在分发环节，数字平台（如Meta）的黑箱竞价与算法机制决定了其实时投放无法像物理直邮等确定性渠道那样保证绝对的一对一精准定向。因此，企业AI的最终归宿并非替代人类，而是剥离机械化琐碎流程以“放大人类效应”，促成更本真的人际连接。 Rostamian总结指出，营销智能化的核心在于以坚实的数据基础设施与算法模型为支撑，将原本主观模糊的定性资产转化为可操作的运算输入；但该体系的长效运转仍受制于数字媒体算法分发的非确定性现实，且高度依赖人类对客户需求的底层直觉，盲目跳过基础阶段或剥离人工洞察均会带来系统性操作风险。

Detailed notes — zh

## Background 本视频由 Ai4 频道于 2025 年 8 月 26 日在 YouTube 平台发布，主讲人为 SmartData 营销分析负责人 Armen Rostamian。视频简介指出，内容主要探讨营销团队如何借助 AI Agent（智能体）加速数据洞察，在媒体投放、创意测试与生命周期策略中构建由前瞻性驱动的预测型增长引擎。 ## TL;DR 演讲深入剖析了 AI 与智能体如何打破企业部门壁垒、消除内部沟通损耗，推动营销数据分析从描述、诊断、预测、处方阶段跨越至“生成式”自动化落地。针对数字广告中最难量化的创意素材优化瓶颈，讲者展示了 AI Agent 将视频拆解为精细元数据并生成全新创意主题的实战案例。他强调，企业必须按部就班夯实数据与自动化底座，让 AI 充当过滤噪音的放大器，使营销人员得以回归以客户为中心的战略思考。 ## Atmosphere & style 整场演讲洋溢着严谨专业且极具战略前瞻性的行业峰会氛围。讲者条理清晰、娓娓道来，将宏观的 AI 生产力叙事与落地执行中的精细案例紧密结合，既有一针见血剖析传统组织内耗的敏锐，又有勾勒人机协同未来的笃定。观看体验清晰务实，让听众在逻辑递进的分析框架中深切感受到营销决策模式转型的紧迫感与巨大潜力。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=4s) AI 将呈指数级放大个人与组织的生产力，显著缩短从战略愿景到业务执行的流转路径，降低内部沟通与政治损耗。 - [00:04](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=4s) AI Agent 抹平了统计建模与编程的技术门槛，赋能最贴近客户的一线业务人员直接做出精准决策，推动团队回归以客户为中心。 - [04:21](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=261s) 企业数据成熟度遵循描述性、诊断性、预测性到处方式的经典演进，而生成式 AI（Generative AI）作为第五前沿正将分析洞察直接转化为自动化运营落地。 - [07:15](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=435s) 在数字广告投放触及收益递减拐点时，直接削减预算会导致获客成本飙升，唯有实现广告创意主题与素材的多样化才能实现破局。 - [10:12](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=612s) 传统创意制作与评估流程存在多部门协作壁垒、耗时长、成本高且主观性强的“组织负债”，阻碍了营销迭代效率。 - [10:12](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=612s) AI Agent 通过逐秒拆解视频并建立颗粒度精细的素材元数据表（Meta-table），将定性创意转化为可量化、可进行严格变量控制的标准测试体系。 - [10:12](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=612s) 生成式 AI 能主动诊断素材疲态并拓展全新的创意维度，帮助企业在降低获客成本（CAC）的同时大幅放大切入新受众的投放规模。 - [16:28](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=988s) 核心创意概念确立后，生成式 AI 能够自动将其自适应重组为契合 TikTok、Facebook 等不同渠道特性的原生营销素材。 - [16:28](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=988s) 由 AI 动态驱动的自动化测试日历能够综合评估风险与业务收益，直接在源头剔除低价值测试，避免企业资源的无效空耗。 - [16:28](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=988s) AI 的卓越效能建立在人类非 AI 领域知识与商业逻辑之上，唯有具备识别真实信号的能力，才能最大化释放 AI 的自动化杠杆价值。 - [18:40](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=1120s) 企业构建 AI 决策体系绝不能跳步，必须扎实经历“底层数据治理 → 流程自动化 → 机器学习预测 → AI Agent 智能决策”的演进阶梯。 - [18:40](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=1120s) AI 的核心使命并非取代人类，而是成倍放大人的影响力（Multiply human effect），通过过滤冗余噪音让营销回归有温度的人际连接与深度洞察。 - [19:49](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=1189s) 面对现代数字广告平台的黑盒推荐算法，AI 虽无法实现确定性的一对一强行分发，但能在输入端优化策略，有效引导算法更精准地捕捉目标受众。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=4s) AI 驱动的生产力革命与组织对齐 SmartData 营销分析与 AI 业务负责人 Armen Rostamian 开场介绍了公司作为商业咨询机构的核心定位：帮助客户在现有发展阶段做出更精准的营销决策。他指出，本次分享聚焦于 AI 用例与 AI Agent（智能体）如何帮助组织实现更高维度的对齐（Alignment），从而为大型企业内部的营销团队带来实质性的业务突破。 ```mermaid flowchart LR accTitle: AI 缩短从愿景到执行的路径 subgraph Traditional["传统模式"] V1["战略愿景"] --> P1["部门墙与内部博弈"] --> P2["冗长流程与妥协"] --> E1["失真执行"] end subgraph Agentic["AI Agent 赋能模式"] V2["战略愿景"] --> A1["AI Agent 协同与标准化"] --> E2["精准直接落地"] end ``` 为了说明这一趋势的深远影响，Armen 引用了两位行业领袖的前瞻观点：Sam Altman 预言在不久的将来会出现“一人实现 10 亿美元估值的超级企业”；黄仁勋（Jensen Huang）则预测未来 5 年内 AI 催生的百万富翁数量将超过互联网诞生前 20 年的总和。在 Armen 看来，这两句话揭示了同一个本质逻辑——AI 将呈指数级放大个人与组织的生产力。当前英伟达（NVIDIA）和奈飞（Netflix）等高人效标杆企业的人均创收已达到约 250 万美元，而 AI 的普及将彻底重塑生产力上限。 - **消除借口与降低内部损耗**：AI 缩短了从业务“愿景”到“执行”的距离，减少了部门摩擦、政治博弈与反复传话带来的信息失真，让竞争回归到“点子是否新颖、力量是否强大”本身。 - **决策重心转向战略**：过去大量优秀想法因复杂的组织流程而胎死腹中，AI 赋能让团队将精力从“跑流程”转向“定战略”。 - **打破技术门槛，回归以客户为中心**：以往企业苦苦寻找既懂客户心理、又精通统计回归和编程的“独角兽”全才；如今借助 Agent 解决方案，最贴近客户的业务人员即可直接做出精准决策，真正释放以客户为核心的思考空间。 --- ### [04:21](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=261s) 从经典分析成长曲线到生成式 AI 的跨越企业数据成熟度的演进通常遵循经典的分析成长曲线，包含四个必经阶段：描述性分析（Descriptive，搞清发生了什么，如统计销售额）、诊断性分析（Diagnostic，查明影响因素，如将销售额拆解为访客数与转化率）、预测性分析（Predictive，基于转化率与访客趋势预测明日销售）以及处方式分析（Prescriptive，指导资源投入，如优化销售漏斗以将转化率从 2% 提升至 3%）。 ```mermaid flowchart TD accTitle: 数据分析成熟度演进与生成式前沿 subgraph Classical["经典分析四阶段（由左至右逐步演进）"] D["1. 描述性 (Descriptive) 事实统计（如：销售额）"] DIA["2. 诊断性 (Diagnostic) 归因拆解（如：流量/转化率）"] P["3. 预测性 (Predictive) 趋势预判（如：MMM模型）"] PRE["4. 处方式 (Prescriptive) 策略推荐（如：调整漏斗/预算）"] D --> DIA --> P --> PRE end subgraph Frontier["第五前沿"] GEN["5. 生成式 (Generative AI) 将分析直接转化为运营落地"] end PRE --> GEN ``` 企业无法跳过这四步中的任何一环，但 AI 能够起到“使曲线向上平移并向左移动”的作用，既加速演进节奏，又显著提升各个节点的产出质量与效率。Armen 强调，AI 正在将“数据分析”直接转化为“自动化运营”，而曲线的第五个前沿正是“生成式（Generative）能力”。生成式 AI 不仅能指出该做什么，更能以极低的人力资源消耗直接生成业务交付物，完成从洞察到落地的闭环。 --- ### [07:15](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=435s) 创意素材分析的定性与定量困境在营销领域中，最偏向定性、最难被量化衡量的环节通常是创意素材（Creative Assets）的表现。传统的创意产出高度依赖创意团队、自由设计师或网红（KOL/Influencer），充满主观色彩。而在现代数字广告体系中，创意与用户之间隔着复杂的算法平台，必须建立一套严谨的方法论来量化评估创意要素。以一家月广告预算达 100 万美元、要求 6 个月投资回收期的典型企业为例： 1. **统计与诊断**：通过机器学习归因模型，团队发现 Meta 广告上的点击率与转化率出现持续下滑。 2. **预测与瓶颈识别**：借助营销组合模型（Media Mix Model, MMM）进行统计预测，系统发现 Meta 渠道的投放已达到收益递减曲线的拐点，边际效益下降。 3. **处方决策**：若直接削减预算会导致获客成本（CAC）飙升，因此唯一的破局方案是“实现创意主题与广告素材的多样化”。 --- ### [10:12](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=612s) 传统创意流程的组织负债 vs. AI Agent 标准化突破当面临“素材缺乏多样性”的挑战时，传统组织协作模式往往会陷入沉重的“组织负债”与“仓鼠轮（Hamster Wheel）”式的低效循环。 | 对比维度 | 传统人工推进流程 | AI Agent 赋能模式 | |:--- |:--- |:--- | | **涉及部门与节点** | 涉及数据、品牌、创意/UGC、媒介采买、财务等 5 个部门，包含 6 个层级的沟通流转与潜在故障点 | 核心增长团队直接与 AI Agent 协作管理，大幅缩减协作链条 | | **研发耗时与成本** | 周期长达数月，单轮素材制作耗资可达约 50,000 美元 | 快速迭代产出，测试成本与试错门槛极低 | | **素材评估方式** | 高度主观（如“这个网红更契合客群”），评价模糊且失真 | 颗粒度精细的元数据表格（Meta-table），定量拆解前 3 秒钩子、人物设定、文案与 CTA | | **归因与优化精度** | 难以剥离变量，好坏全凭主观猜测 | 固定表现优异的变量（如前 3 秒画面），仅修改失效变量（如下单提示 CTA） | ```mermaid flowchart TD accTitle: 传统创意流程的六大断点 A["1. 数据团队产出洞察"] --> B["2. 品牌团队被动防守并拟定Brief"] B --> C["3. 创意/UGC团队主观解读并制作"] C --> D["4. 内部主观评审筛选素材"] D --> E["5. 媒介采买团队定向投放"] E --> F["6. 财务部门审批测试预算"] ``` AI Agent 的介入主要带来两层颠覆： 1. **定性素材的精细化与定量标准化**：Agent 将视频内容逐秒拆解。例如系统会记录“前 3 秒钩子为女性博主在居家环境中探讨网络安全”，建立详尽的素材元数据表。明确哪些变量在促成转化、哪些在拉低效果，从而实现精准的变量控制测试。 2. **生成式 AI 主动拓展全新创意维度**：以一家网络安全合作客户为例，该企业长期依赖“拦截骚扰电话（Robocalls）”这一单一卖点进行推广，投放规模遭遇严重瓶颈。生成式 AI 介入后，系统诊断出该主题的疲态，并推荐了全新的“从互联网上抹去你的个人痕迹（Vanish from the internet）”主题。这一全新角度切中了完全不同的潜在人群，最终助力客户在获客成本降低 10% 的同时，广告投放规模成功放大了 50%。 --- ### [16:28](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=988s) 营销战略演进与测试流程重塑展望未来，AI 对营销运营的深度改造将集中在两大核心板块： - **跨渠道战役适配（Campaign Strategy）**：当核心创意概念确立后（如“从互联网抹去痕迹”），生成式 AI 能够自动将其转化为契合不同渠道特性的原生素材。例如 TikTok 与 Facebook 所需的节奏与形式截然不同，AI 可以在保持核心主题一致的前提下完成自适应重组与元素拼装。 - **测试策略与日历智能化（Testing Strategy）**：企业内部消耗资源最多的往往是“注定失败的劣质测试”。未来由 AI 动态维护的自动化测试日历，将综合评估风险与业务收益，直接在源头剔除低价值测试方案，避免资源浪费。 Armen 在本段末尾总结道，AI 的卓越能力始终建立在“非 AI 知识”的根基之上；企业必须首先建立坚实的人类理解、直觉与商业逻辑认知，懂得如何识别真正的有效信号，才能最大化释放 AI 的自动化杠杆效能。 ### [18:40](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=1120s) 构建 AI 决策体系的阶梯与人机协同愿景要与 AI Agent 展开富有成效的交互，关键在于绝不能“跳步”。企业必须先扎实完成底层基础建设，从基础数据治理走向自动化流程，再融入机器学习（ML）能力，最终才能稳步迈进具有实质业务意义的“AI 驱动决策”（AI decision-making）。在拥抱持续迭代的 AI 方案时，技术架构应当具备足够的灵活性与前瞻性，团队应该“期待下一次基础模型的版本发布，而不是对更新感到恐惧”（"looking forward to the next model release, not dread it"）。只有能够无缝融合模型演进并持续放大业务杠杆的方案，才值得企业重金投入。 ```mermaid flowchart LR accTitle: 从基础数据到 AI 决策的演进路径 subgraph S1["基础建设期"] A["底层数据与基础设施"] --> B["流程自动化"] end subgraph S2["智能进阶期"] B --> C["机器学习 (ML) 预测"] C --> D["AI Agent 智能决策"] end ``` AI 的最终使命并非取代人类，而是“成倍放大人的影响力”（"multiply the human effect"）。通过利用 AI 过滤并剥离业务流中的冗余噪音，营销团队能够极大地拉近与客户之间的距离，重新回归有温度的“人与人之间的连接”，在显著优化终端客户体验的同时，切实减轻内部团队的工作负荷与疲劳感。 - **阶段推进原则**：不可跳过底层数据基础设施与自动化直接上马大模型；只有经历“基础数据 $\rightarrow$ 自动化 $\rightarrow$ 机器学习 $\rightarrow$ AI 决策”的完整链条，Agent 才能产生真实价值。 - **架构敏捷度**：打造适应性强的弹性系统，确保每次大模型升级都能直接转化为业务杠杆的提升。 - **人本赋能**：将 Agent 定位为噪音过滤器与效率放大器，赋能团队专注于深度的客户洞察与真实的人际连接。 ### [19:49](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=1189s) 现场问答：算法黑盒投放与精准触达的现实考量在演讲后的互动环节中，现场观众与讲者探讨了交付触达环节的确定性与预防性控制。对于传统的直邮（direct mailer）渠道，营销人员可以做到极其精准的一对一确定性触达；然而在以 Meta Ads 为代表的现代数字广告平台中，投放直接交由平台底层的黑盒推荐算法进行实时决策。在算法掌控分发逻辑的背景下，AI 虽无法做到绝对的一对一精准控制，但能有效辅助并引导算法走向更具针对性的优化方向。 - **渠道特性的差异**：直邮等传统渠道具备高确定性与精准性，而 Meta 等平台广告则重度依赖平台算法进行人群分发。 - **AI 在分发侧的定位**：AI 无法完全越过平台算法实现纯粹的一对一强行分发，但能够在前端优化输入策略，帮助平台算法更精准地捕捉目标受众。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=4s) 一人实现 10 亿美元估值的超级企业 — Sam Altman > [00:04](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=4s) 未来 5 年内 AI 催生的百万富翁数量将超过互联网诞生前 20 年的总和 — Jensen Huang > [04:21](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=261s) 使曲线向上平移并向左移动 — Armen Rostamian > [07:15](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=435s) 实现创意主题与广告素材的多样化 — Armen Rostamian > [10:12](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=612s) 从互联网上抹去你的个人痕迹（Vanish from the internet） — Armen Rostamian > [16:28](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=988s) AI 的卓越能力始终建立在“非 AI 知识”的根基之上 — Armen Rostamian > [18:40](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=1120s) looking forward to the next model release, not dread it（期待下一次基础模型的版本发布，而不是对更新感到恐惧） — Armen Rostamian > [18:40](https://www.youtube.com/watch?v=dJs2Dk7ARwk&t=1120s) 成倍放大人的影响力（multiply the human effect） — Armen Rostamian ## Mentioned SmartData, Armen Rostamian, Sam Altman, Jensen Huang, NVIDIA, Netflix, Meta Ads, TikTok, Media Mix Model (MMM), AI Agent, Generative AI ## Video description Armen Rostamian, President, Marketing Analytics, SmartData Abstracts What if you could act on insights before the signal even hits your dashboard? AI agents can accelerate marketing insights—shifting us to predictive growth engines powered by collective intuition. I’ll share real-world examples of how forward-thinking teams are deploying agents for media investment, creative testing, and lifecycle strategy—grounded in foresight and driven by automation. Join me to discover how the next era of marketing is being built on decisions made ahead of the curve. [Source](https://www.youtube.com/watch?v=dJs2Dk7ARwk)

Source quotes

[0:04] Sam Altman: 一人实现 10 亿美元估值的超级企业 [0:04] Jensen Huang: 未来 5 年内 AI 催生的百万富翁数量将超过互联网诞生前 20 年的总和 [4:21] Armen Rostamian: 使曲线向上平移并向左移动 [7:15] Armen Rostamian: 实现创意主题与广告素材的多样化 [10:12] Armen Rostamian: 从互联网上抹去你的个人痕迹（Vanish from the internet） [16:28] Armen Rostamian: AI 的卓越能力始终建立在“非 AI 知识”的根基之上 [18:40] Armen Rostamian: looking forward to the next model release, not dread it（期待下一次基础模型的版本发布，而不是对更新感到恐惧） [18:40] Armen Rostamian: 成倍放大人的影响力（multiply the human effect）

Source quotes

[0:04] Armen Rostamian: meeting them wherever they happen to be on their data journey [2:07] Armen Rostamian: die on the vine [4:21] Armen Rostamian: shifting the entire curve both upward and to the left [10:12] Armen Rostamian: hamster wheel [12:16] Armen Rostamian: robocalls are annoying, and we can stop them [12:16] Armen Rostamian: vanish from the internet [16:28] Armen Rostamian: built on top of non-AI knowledge [18:40] Armen Rostamian: look forward to the next model release, not dread it [18:40] Armen Rostamian: multiply the human effect

Original transcript

[0:04] I am Armen Rostamian. I lead the smart [0:08] data analytics and AI organization. We [0:10] are a business consulting firm and we [0:13] really focus on um enabling better [0:16] marketing decisions uh and meeting our [0:18] clients where they are. uh but [0:20] ultimately I think in this conversation [0:21] I really wanted to focus on AI uh use [0:24] cases and AI agents and how they could [0:27] better align organizations and [0:28] ultimately lead to better outcomes for [0:30] uh marketing uh companies and marketing [0:34] organizations within uh larger [0:36] organizations. So I wanted to start out [0:39] with [0:41] uh these two quotes. I know at this [0:44] conference we have the godmother and [0:45] godfather of of AI but I I think these [0:48] two guys are pretty influential [0:50] themselves. Uh and ultimately uh [0:53] starting with Sam's quote he's saying [0:54] that u he foresees a oneperson [0:57] organization becoming a billion-doll [1:00] organization in the near future. Um, and [1:05] Jensen Hang says there'll be more [1:07] millionaires in the next 5 years as a [1:09] result of AI than the first 20 years of [1:12] of uh the internet. And so ultimately [1:15] these two things speak to the same thing [1:17] in my estimation which is the [1:19] productivity of an individual and an [1:22] organization can be exponentially [1:24] increased by using AI. And so um and and [1:28] it's not always about cost cutting and [1:30] it's not always about being smart. It's [1:32] it's also really leveraging AI to [1:35] produce better alignment in your [1:38] organization. [1:39] And so, you know, when we think of the [1:41] numbers here, right now, Nvidia and [1:44] Netflix are probably the most productive [1:46] companies per employee and they produce [1:48] about 2.5 million in revenue per [1:51] employee. And so when you look at these [1:53] numbers uh you realize that you know we [1:56] are in the midst of a pretty substantial [1:58] revolution in in how productive a person [2:01] can be and how uh how much you can [2:04] multiply your own uh decision making. [2:07] And so ultimately when we talk about [2:11] um what more AI means in marketing [2:13] organizations [2:14] it really comes down to not having [2:17] excuses anymore. It's really comes down [2:20] to how novel and how powerful is your [2:23] idea and it reduces the no the amount of [2:27] noise and so the distance between the [2:30] vision and the execution is reduced. [2:33] There's less noise, less corporate [2:35] politics, you know, less departments [2:36] touching things, less people chiming in, [2:39] if you will. U and so ultimately all of [2:43] that efficiency and alignment should [2:45] yield better results. And then the [2:48] decisions become about strategy not [2:50] process. So whereas before you know a a [2:54] good enough idea would die on the vine [2:56] because of the organizational [2:58] complexities now by reducing the process [3:01] and really focusing on the strategy you [3:04] can uh yield better productivity per [3:07] person and ultimately um better success [3:11] and then there's fewer technical [3:13] barriers. So the person that is uh [3:17] closest to the customer can start making [3:20] decisions. So you're not looking for [3:22] unicorns anymore that had to be [3:25] empathetic towards the customer, had to [3:28] understand statistical regressions, had [3:31] to be a coder, had to do all of these [3:34] technical things all at once. Now, uh, [3:36] with the use of, uh, agentic solutions, [3:39] you could have your marketer, the person [3:43] that's thinking of themselves in the [3:45] customer's shoes, really making these [3:47] decisions in a more exact and direct [3:49] way. Uh, so I think that's going to be a [3:52] really powerful opportunity for any [3:54] marketer to become [3:56] very customer centric. And so again, [3:59] it's not about necessarily reducing cost [4:01] or becoming more efficient as an [4:03] organization per se. It's really freeing [4:06] up the mind space and the opportunity [4:08] for marketers to live uh through their [4:11] customers uh vision and ultimately serve [4:14] that customer better. So the quality [4:16] should improve as much as um the [4:19] efficiency does. [4:21] So if we go on like what I wanted to [4:24] start was uh here which is the classic [4:26] analytics growth curve. So this is you [4:30] know something that we all uh strive for [4:33] both in in business and just in as as [4:36] people where you know you're really [4:38] trying to one understand what's [4:39] happening uh the descriptive part and [4:42] then you grow into diagnosing things. So [4:45] you know describing is counting up [4:46] orders right doing all of those things [4:48] and then ultimately is the diagnostic [4:50] part which is like what's influencing [4:52] that number what's influencing these [4:54] outcomes um which then becomes uh once [4:58] you know that really well then you can [4:59] start predicting the future and then [5:02] ultimately you start prescribing [5:03] different things. So taking like a [5:05] simple example, you know, in descripting [5:08] in describing something accurately, [5:09] you're counting up orders and you're [5:11] saying, you know, our sales yesterday [5:12] was X dollars. And then once you grow [5:16] into the diagnostic, now you're saying [5:17] that the X dollars in sales were a [5:20] function of our visitors to our website [5:22] and our conversion rate. So, you know, [5:24] really starting to break down that into [5:26] a diagnostic. And so once you have those [5:29] components, now you can start predicting [5:30] and say like, well, tomorrow's [5:32] conversion rate is going to be X and our [5:34] our visitor count is going to be Y, so [5:36] we predict we'll have X amount of sales. [5:39] And once you start being able to do [5:41] that, then ultimately you start [5:42] prescribing things and saying, well, now [5:44] let's invest our resources towards [5:46] improving our salesunnel. Uh, ultimately [5:48] improving our conversion rate from 2% to [5:51] three with the visitors we expect. This [5:53] is the return we'll get on this project. [5:56] So you know that's the classic curve and [5:59] now AI could really help you along this [6:01] curve. You know first thing it could do [6:02] for any organization is is move you [6:04] through this uh steps and ultimately [6:07] there's you can't skip any of these [6:09] steps you know so you have to move along [6:11] these but then ulti [6:14] [Music] [6:16] helping shift up the curve and also [6:19] shift the curve left. So you'll have a [6:23] competitive edge if you're enabled by AI [6:27] and it'll be simpler for you to execute [6:29] like for each one of these steps you [6:31] should be able to ultimately deliver a [6:33] better outcome and a more efficient [6:35] outcome. But the what I started with was [6:39] that AI will turn analytics into [6:41] operations and you know I think the the [6:43] the fifth frontier here is really the [6:46] generative. And so what used to be let [6:49] me describe and prescribe a solution [6:52] that let's improve the sales funnel now [6:55] it becomes the opportunity of generative [6:57] AI to be an extension of that to be a [7:00] continuation of this level of [7:02] understanding uh and ultimately yield uh [7:05] a a outcome with less resources that's [7:08] more productive than uh what was in the [7:11] classic case. So I think uh I I'll walk [7:15] through a particular example here and [7:17] and you could picture it in in different [7:19] settings but for the sake of time what I [7:21] wanted to do was really focus on uh the [7:23] most qualitative piece of uh of [7:26] marketing which is oftentimes the [7:28] creative uh performance. How well does a [7:32] creative ad that's been developed by [7:35] humans uh engage on these platforms and [7:39] with the eventual viewer? So um there's [7:42] always these algorithms between you and [7:44] the customer that you have to consider [7:46] but ultimately um how can we be more [7:48] methodical and quantitative in that [7:50] approach for something that used to be [7:53] qualitative. You know we've all [7:54] interacted with creative teams uh or [7:58] even uh influencers that you interact [8:00] with and uh you know it's always an [8:02] interesting cast of characters u but [8:05] ultimately they they have that creative [8:07] side of them that organizations need. [8:10] Now, how do you get the most out of [8:11] that? So, running through this uh [8:14] example quickly, [8:17] you know, we are talking about a [8:18] specific organization that's spending a [8:20] million dollars uh and has a six-month [8:22] return. And so, you need to be able to [8:24] count things accurately to be able to [8:26] quantify uh that component of it. And [8:30] so, the next step is as you're [8:31] diagnosing, you see that your [8:33] click-through rate, your conversion [8:34] rates have dropped on meta ads. Now, how [8:37] do you know that? You need to know that [8:38] by enabling some ML uh leveraged [8:42] attribution models to really hone in [8:45] that the meta ads are where the issue [8:47] is. But ultimately, let's say, you know, [8:49] along this curve, you you've grown [8:51] enough to to to do that. And then you [8:54] start using your media mix model, which [8:56] is essentially a statistical prediction [8:58] model. And that enables you to [9:01] ultimately um realize that you know the [9:04] the that your spend is excessive. So you [9:08] you've hit a point where you're on a [9:10] degradation curve. You're losing [9:12] efficiency. And so it's ultimately going [9:14] to tell you to reduce spend. So you [9:16] don't want to reduce spend. You don't [9:18] want your CAC to spike. Uh so what do [9:20] you do? You uh the ultimate prescription [9:22] of this is to diversify your themes. Uh [9:26] diversify your ads. Um, and so in a [9:30] typical organization, what happens then [9:31] is you go out and you ask a bunch of [9:33] creatives, hey, help me find new [9:36] audiences [9:38] uh, and and you try that. But [9:40] ultimately, what I think with generative [9:41] AI, you could be a lot more prescriptive [9:45] uh, with uh, the generation of these ads [9:47] and and it could be really powerful. And [9:49] so what I foresee really successful [9:52] companies doing is this isn't going to [9:54] be the exclusive solution. you're still [9:56] going to need that human element of it, [9:58] but ultimately your prompts into [10:01] generative AI are also going to be your [10:04] prompts into your creative departments, [10:06] your creative uh uh assets. And so that [10:09] in and of itself is powerful. [10:12] So here's wi faced with that problem of [10:16] not enough creative uh diversity the the [10:21] boxes on the left is the creative [10:23] process without AI and you know and the [10:27] blurring out of the what the vision was. [10:30] Uh so it starts with the analytics team. [10:32] They they deliver the insight and then [10:36] the brand team takes that is typically [10:39] rather defensive about that starts [10:41] defending their ideas says uh we have [10:43] enough thing and but ultimately let's [10:45] say they they consider that and and um [10:48] look to uh to recommend a test. They [10:52] come up with some ideas and then that [10:54] goes into a brief. They write that up. [10:57] And so you can already see the amount of [10:59] human uh interpretation that's coming [11:02] through here. Uh that brief gets handed [11:05] off to a creative department to uh [11:07] usergenerated networks. They read [11:10] through it. They interpret it a certain [11:11] way. Uh ultimately they come up with uh [11:14] you know some assets. Those come back. [11:17] They're reviewed and then the best ones [11:20] quote unquote best ones are selected. [11:23] They go into a media buying team. that [11:26] media buyers buy against a certain [11:28] audience if it's lined up or not uh is [11:31] is their role. But you know that's where [11:33] it gets fuzzy as well. Ultimately it [11:35] goes to a finance department to approve [11:37] the test budget and all that. And so all [11:40] of that is going to take months, right? [11:43] And we got five uh departments. You've [11:47] needed humans uh to create these uh [11:50] assets. So, that's an investment, uh, [11:53] you know, $50,000 investment. You've [11:56] navigated through a potential of six [11:58] communication layers with six failure [12:01] points. Um, and ultimately, you know, [12:04] you get back and and it's just this [12:06] hamster wheel of going through this uh [12:08] iteration and and trying to get [12:10] incrementally better, but ultimately [12:12] it's it's leaving a lot to chance. So [12:16] what the creative asset uh agents could [12:20] do is is two things. One um the growth [12:24] team can manage this right they have the [12:26] agents they can interact with them not [12:28] all these departments are needed. So now [12:30] you've got a focused core group of [12:32] people that are making the most relevant [12:34] decisions. And what the first layer of [12:36] the uh solution can be is to categorize [12:39] and describe your creatives in a very [12:41] quantitative way rather than a [12:44] qualitative way. So rather than saying [12:47] hey I I think this influencer fit our [12:50] demo better, it it's actually going to [12:53] say in the first 3 seconds our hook was [12:56] X. It was uh a female influencer in her [13:00] home making a pitch about um uh cyber [13:04] security. Um and ultimately that's what [13:08] and you can break down all of these [13:09] components into this big table meta [13:12] table of of describing your creative ad [13:15] in a very methodical way rather than [13:17] leaving it to a creative agent uh agency [13:20] or a creative partner to uh you know [13:23] look through the ad as a human and [13:25] categorize it. you know the the amount [13:27] of subjectivity there is just extra uh [13:29] astronomical. So by first describing it [13:33] very exactly now you know the variables [13:36] that are successful the variables that [13:39] are not contributing and so now you can [13:42] hold the variables that work constant [13:44] and change the variables that don't. So [13:46] the first 3 seconds of the ad might be [13:48] great but the last 5 seconds of it may [13:51] not include a call to action that's [13:53] clear enough. So when when a human is [13:56] looking at that ad, who knows how they [13:57] interpret it. They might think the first [13:59] 3 seconds wasn't as as powerful. So [14:02] ultimately I think this is where uh you [14:05] know the first layer of uh agentic help [14:08] can come is the standardization of a [14:11] qualitative thing which is the creative [14:14] asset. Then the second part, you know, [14:17] once again, you've followed that chain [14:18] and now you're past the prescriptive, [14:20] now you're going into generative. Uh the [14:23] generative part uh and here's an example [14:25] of a cyber security company that we [14:27] worked with. uh they were really having [14:31] a a very successful uh business and uh [14:35] we did this analysis on their creatives [14:37] and and what it yielded was that they [14:40] were really good at a certain theme and [14:43] most of their ad spend was at this theme [14:45] which was hey robocalls are annoying you [14:49] we can help stop them. Robocall stop [14:52] robocalls. So like when you looked at [14:54] their ad spend, the majority was around [14:57] this theme and it was and so it was [15:00] working but it wasn't helping scale and [15:02] the more money they threw at it, the [15:04] further they slip back on this uh [15:06] degradation curve and ultimately were [15:08] stuck at that scale. So what the [15:11] generative AI did was ultimately review [15:13] that point out the fact that that's true [15:16] and then ultimately recommended a net [15:19] new theme which was additive to their [15:22] business and the theme was this vanish [15:25] from the internet theme. So uh distinct [15:29] from the robocall angle but more towards [15:33] uh the the elimination of yourself from [15:36] the internet was resonated with a [15:37] different group of folks. So what they [15:39] were able to do was spend against that [15:42] audience with this new creative up their [15:45] uh spend by 50% while keeping their CAC [15:48] the same and also lowering their CAC. So [15:51] they lowered their CAC by 10% their co [15:54] acquisition cost while scaling by 50% by [15:58] having this net new theme. So, really [16:00] powerful opportunity and one that, you [16:03] know, a human probably could have come [16:04] up with this. You know, it's not like [16:06] this revolutionary amazing thing, but I [16:09] think what it was able to do was be [16:10] methodical, point something out, get it [16:12] tested, don't have to go through this [16:15] hamster wheel of organizational debt, [16:18] and ultimately yield a result that was [16:20] productive for them quickly with a lot [16:23] less investment and a lot less points of [16:26] failure. [16:28] So looking ahead, I think this really [16:30] should impact the operations of of [16:32] marketing organizations particularly [16:34] around campaign strategy and testing [16:35] strategy. Whereas you know you could [16:38] create a campaign now uh it needs to be [16:42] channel specific. There's all these [16:43] layers of of uh different uh things that [16:47] you have to consider for each channel. [16:48] Well ultimately generative AI can help [16:50] you solve for that. So you come up with [16:53] a theme with the concept vanish from the [16:55] internet. Now it could adapt that theme [16:57] into channel level creatives that are [17:00] productive for you. U so a Tik Tok ad [17:03] and a Facebook ad need to be different [17:04] enough but needs to have that [17:06] consistency. [17:07] Um and then soon I think you could start [17:09] creating like novel pieces of creative [17:12] that get stitched together by generative [17:13] AI. And so if you think of like the [17:16] creatives going into a generative AI [17:18] engine instead of the end user then you [17:21] could start uh getting that right too. [17:25] testing strategy. I think you know uh [17:28] eliminating bad tests is probably uh the [17:32] area where the most resources get soaked [17:35] up in an organization. A a test that has [17:37] no chance of success but still resources [17:40] are dedicated to it and get wasted [17:41] there. So ultimately I think you know uh [17:44] what it can oops I was supposed to sorry [17:47] about that I was looking at it on my [17:49] screen but ultimately like it's going to [17:51] eliminate bad tests um and that soak up [17:54] a disproportionate amount of resources [17:56] and then ultimately I think in the near [17:58] future there'll be a testing calendar [17:59] that takes into consideration risk and [18:01] opportunity and everything else that [18:04] ultimately uh gets delivered to [18:06] organizations [18:08] uh and it's upkept by the AI [18:12] So that's it. So in summary, I think you [18:14] know u in summary I'd say that you know [18:18] the AI functionality and capability is [18:21] built on nonAI knowledge. Um and so you [18:26] know you have to invest first in [18:28] building that knowledge and it's a human [18:31] understanding and intuition first. you [18:33] know, you have to know what's going to [18:36] work and and you have to be able to [18:38] recognize those things first and then [18:40] you could uh interact with an agent in a [18:42] in a productive way. But I would say [18:44] that the foundation is always uh starts [18:47] there. And like I said, you can't skip [18:49] any steps. You have to be able to do the [18:51] foundational work. You have to be able [18:52] to do the automation, the ML, and then [18:54] you can get to AI decision- making in a [18:56] meaningful way. [18:59] Uh and then the the last couple concepts [19:01] is one um you know with each iteration [19:03] of AI solutions we should be looking [19:06] forward to the next model release not [19:09] dread it. So coming up with solutions [19:11] that are flexible enough and are taking [19:13] that into consideration to increase your [19:16] leverage is going to be ultimately the [19:18] AI solutions that you need to invest in. [19:21] And then lastly is, you know, to [19:23] leverage AI to multiply the human [19:25] effect, get you closer to the customer, [19:28] get you closer to that human-toum [19:30] connection while dropping off the noise [19:33] in the process in a way that is going to [19:35] yield uh a better experience for your [19:38] customer and and for your teams. So [19:41] that's it. Uh, thank you for your time [19:44] and I'll be sure. [19:49] >> Uh, [19:50] >> that would be awesome. [19:51] >> Well, I think the delivery part is the [19:54] preventative part because I don't like [19:56] maybe if it's a mailer, I could be very [19:58] precise with that. But with a meta ad, [20:01] it's going into an algorithm that's [20:02] making those some of those decisions for [20:04] you. So, yeah, you could help that [20:06] along. Yeah, you could help that along [20:07] for sure and make it more targeted. Uh [20:10] but it won't be a onetoone for