Ex-DeepMind Researcher Deploys Pomo AI to Automate Brand Ad Teams
Ai4
Small teams can outmaneuver enterprise marketing departments by using autonomous agents with decision memory to bypass corporate approval chains and agency retainers.
## Background
Published by Ai4 on YouTube on August 17, 2026, this presentation features Praneet Dutta, founder of usepomo.ai and former Google DeepMind researcher. The session outlines how autonomous AI agents enable small, two-person marketing teams to outperform massive enterprise brands by bypassing bureaucratic approval chains and agency retainers to drive revenue directly across major ad platforms.
## TL;DR
Praneet Dutta explains how lean teams can leverage autonomous AI agents to unify fragmented data—including inventory, POS records, and external market signals—into a centralized intelligence layer. Drawing from his frontier research at Google DeepMind and execution at Pomo, he illustrates how to avoid reactive discounting and introduces 'decision lineage' to preserve institutional marketing memory. He presents a three-tier risk delegation framework that pairs continuous background research and guardrailed creative launches with human-led strategic judgment. Finally, he demonstrates how a domain-first workflow enables small brands to execute multi-channel campaigns, dynamic PDP creation, and competitor tracking with unprecedented velocity.
## Atmosphere & style
The presentation carries a crisp, pragmatic, and insider-driven tone that bridges deep frontier AI engineering with the scrappy realities of early-stage growth. Dutta speaks with authoritative clarity, punctuating technical concepts with dry humor—such as mocking superficial AI summaries as the analytical equivalent of stating 'water is wet.' The delivery feels empowering and methodical, creating a sense of strategic urgency while providing a grounded, highly actionable blueprint for lean operators.
## Key takeaways
- [00:04](https://www.youtube.com/watch?v=8aCXoSXbJqc&t=4s) Former Google DeepMind researcher Praneet Dutta built usepomo.ai with a 6-person team to demonstrate how autonomous AI agents give small teams massive operational leverage.
- [03:14](https://www.youtube.com/watch?v=8aCXoSXbJqc&t=194s) Modern marketing requires monitoring fragmented channels, from POS and retail inventory to emerging Answer Engine Optimization (AEO) on AI search engines like ChatGPT and Claude.
- [03:14](https://www.youtube.com/watch?v=8aCXoSXbJqc&t=194s) Unifying siloed internal and external data signals into a centralized intelligence layer eliminates costly context-switching for resource-constrained growth teams.
- [05:01](https://www.youtube.com/watch?v=8aCXoSXbJqc&t=301s) High-performing marketers avoid knee-jerk competitive reactions, using agentic synthesis to evaluate competitor moves against internal margins, inventory levels, and ICP overlap.
- [06:49](https://www.youtube.com/watch?v=8aCXoSXbJqc&t=409s) Lean teams function as agile autonomous pods that can outpace 2,000-person enterprises burdened by slow approval chains by offloading research and data collation to AI.
- [09:13](https://www.youtube.com/watch?v=8aCXoSXbJqc&t=553s) Effective AI marketing requires moving beyond superficial generalities to uncover non-consensus, high-leverage insights that link internal bottlenecks to competitor weaknesses.
- [10:14](https://www.youtube.com/watch?v=8aCXoSXbJqc&t=614s) Sustainable growth architecture combines three core pillars: internal operational truth (Company Brain), external market dynamics (Market State), and historical choice tracking (Decision Lineage).
- [10:14](https://www.youtube.com/watch?v=8aCXoSXbJqc&t=614s) Capturing Decision Lineage turns past marketing experiments and human approvals into institutional memory, allowing agents to continuously calibrate future recommendations.
- [13:40](https://www.youtube.com/watch?v=8aCXoSXbJqc&t=820s) Delegating work to AI requires a three-tier risk model: continuous autonomous synthesis, human-reviewed tactical execution, and human-led strategic direction.
- [13:40](https://www.youtube.com/watch?v=8aCXoSXbJqc&t=820s) Competitive marketing has evolved from periodic, batch-based campaigns to continuous, always-on micro-experimentation and real-time iteration.
- [15:54](https://www.youtube.com/watch?v=8aCXoSXbJqc&t=954s) Domain-first onboarding enables autonomous agents to instantly extract ICP targeting, competitive intelligence, and white space opportunities using just a company website URL.
## Detailed walkthrough
### [00:04](https://www.youtube.com/watch?v=8aCXoSXbJqc&t=4s) From Frontier Research at DeepMind to Empowering Lean Teams
Praneet Dutta opens his talk by introducing his personal journey from Big Tech to entrepreneurship. Straight out of university, he spent over eight years building enterprise software and over a decade in applied artificial intelligence, starting back when convolutional neural networks were just emerging and large language models did not yet exist. Remarking on how an entire applied AI conference ten years ago "wouldn't fill up like half a room of this size," he reflects on the explosion of interest across the industry. Dutta spent significant time as a researcher at Google DeepMind working directly on frontier models—leading capabilities on projects including Imagen 3, Imagen 4, and Gemini Nano—as well as specialized applications such as AI for soccer analytics, renewables forecasting, and reinforcement learning for HVAC systems.
Seeing the frontier shift rapidly from passive conversational chatbots to autonomous agents that act on behalf of users, Dutta and his co-founder (whom he worked with in his early days at Google) left Big Tech to launch Pomo (usepomo.ai). Their core conviction is that AI acts as the ultimate democratizer, granting small teams the agency and velocity previously restricted to massive enterprise organizations. Proving their own thesis on lean execution, they built the entire Pomo platform with a core team of just six people—including only four to five engineers and Dutta handling growth and marketing.
- **Speaker Background:** Praneet Dutta, Founder of usepomo.ai; former Google DeepMind frontier model researcher.
- **Key Frontier Projects:** Imagen 3, Imagen 4, Gemini Nano, AI for soccer tactical analysis, renewable energy forecasting, and RL-driven HVAC thermal control.
- **Pomo Team Footprint:** Entire platform engineered and launched with only 4 to 5 engineers and 6 total team members.
---
[03:14](https://www.youtube.com/watch?v=8aCXoSXbJqc&t=194s) The Fragmentation of Modern Growth and Decision-Making
For lean teams and SMBs, growth hinges on a small group of operators making high-stakes decisions with limited resources. Marketers must constantly evaluate past performance: how budget was allocated over the trailing 60 days, which SKUs are driving high margins versus sitting stagnant in retail channels, and where customer demand is shifting. Furthermore, the discovery landscape is undergoing a structural transformation. Prospective buyers increasingly rely on AI search agents like ChatGPT or Claude for product recommendations, turning Answer Engine Optimization (AEO) and dynamic indexing into essential battlegrounds. Simultaneously, organic conversations on TikTok, Instagram, and LinkedIn dictate what products gain traction and inform future product development.
In the legacy MarTech ecosystem, handling these disparate dynamics required stitching together fragmented point solutions—separate tools for paid ad campaigns, social listening, retail inventory tracking, and AEO. Dutta emphasizes that attention and time are a small team's most scarce assets. Pomo was designed to unify these siloed market signals into a centralized intelligence layer that helps operators make faster, data-backed strategic moves rather than context-switching across dozens of dashboards.
```mermaid
flowchart TD
accTitle: Unified Growth Intelligence Architecture
subgraph SOUR["Data & Market Ingestion"]
A["Company Brain<br/>POS, CRM, Inventory & Ad Audits"]
B["Market State<br/>Paid Ads, Social Listening & AEO"]
end
subgraph AGNT["Autonomous Synthesis"]
C["Analytical Agents<br/>Correlation, Causation & White-Space Discovery"]
end
subgraph EXEC["Execution & Feedback"]
E["Multi-Channel Action<br/>Win-Backs, Creatives & AI SEO"]
F[("Decision Lineage Memory<br/>Continuous Learning Loop")]
end
A --> C
B --> C
C --> D
D --> E
E --> F
F --> C
```
---
[05:01](https://www.youtube.com/watch?v=8aCXoSXbJqc&t=301s) Strategic Synthesis vs. Knee-Jerk Reactions: The Dewey Decimal Example
To illustrate how agentic intelligence alters decision-making, Dutta walks through a hypothetical direct-to-consumer skincare brand called "Dewey Decimal." Suppose Dewey Decimal's primary competitor abruptly launches a 20% discount promotion on an evening ad run. When the marketing team logs in the next morning, the instinctive, knee-jerk reaction is often to match or beat the offer with a 25% discount without reviewing underlying numbers.
Dutta argues that prudent, high-performing marketers refuse to jump directly into ad spend. Instead, they examine foundational data first:
- Does the competitor's discounted product actually intersect with Dewey Decimal's Ideal Customer Profile (ICP)?
- Are they capturing an unaddressed white space, or simply offloading dead inventory?
- What are Dewey Decimal's current inventory levels, product margins, and existing search demand?
- How is the brand stocked across retail distribution relative to e-commerce?
Rather than automatically generating an ad or lowering prices, Pomo's agents analyze the competitive offer against trailing seven-day search volumes, paid and organic social performance, margin buffers, and retail stock. The system synthesizes these disparate signals to provide a nuanced recommendation: whether to aggressively launch a counter-campaign, maintain a holding pattern to monitor the competitor's stock depletion, or ignore the promotion entirely.
---
[06:49](https://www.youtube.com/watch?v=8aCXoSXbJqc&t=409s) The Autonomy Advantage: Enterprise Pods vs. Agile Startups
Drawing on his years inside Big Tech, Dutta offers a "spicy take" on enterprise efficiency versus startup agility. While large enterprises possess massive advantages in brand equity, capital reserves, proprietary distribution, and specialized personnel, their highest-velocity units are almost always small, cross-functional pods granted strict autonomy, clear geographical or product focus, and direct ownership over end results. However, enterprises often drown these pods in approval chains and bureaucracy.
In contrast, a startup or SMB *is* that single autonomous pod. Lean teams do not have the luxury of dedicated sub-departments for data aggregation, POS analysis, and competitive scraping. By leveraging autonomous agents to perform the tedious information gathering and initial synthesis, small operators can dedicate their mental bandwidth purely to strategic judgment. Humans decide whether a synthesized insight aligns with company goals, leaving the background execution and data collation to AI.
---
[09:13](https://www.youtube.com/watch?v=8aCXoSXbJqc&t=553s) The Judgment Imperative: Avoiding Superficial AI Insights
When delegating research to general-purpose agents, Dutta cautions against accepting superficial summaries. Prompting a generic agent to scrape a sector often yields platitudes—such as reporting that "customers increasingly drink tea during daily routines," which Dutta mocks as the analytical equivalent of stating that "water is wet."
True marketing leverage requires uncovering non-consensus, controversial, or hidden insights that link disparate operational data points. A high-leverage agentic system cross-references competitor traction against internal conversion slowdowns, identifying precise market white spaces where customer acquisition has stalled. Operators can then use that unified intelligence to take decisive, revenue-generating downstream actions rather than relying on shallow industry generalities.
---
[10:14](https://www.youtube.com/watch?v=8aCXoSXbJqc&t=614s) The Three Pillars of Modern Growth: The Power of Decision Lineage
Dutta outlines three foundational pillars defining the future of business operations:
1. **Company Brain:** Internal proprietary assets, including Point of Sale (POS) records, CRM pipelines, ad account audit logs, margin data, and warehouse inventory.
2. **Market State:** Real-time external signals across search indexing, competitor ad libraries, social conversations, and earned media.
3. **Decision Lineage:** The captured historical record of past operational choices and their downstream business outcomes.
| Growth Pillar | Composition & Source Signals | Primary Function in Autonomous Workflow |
|:--- |:--- |:--- |
| **Company Brain** | Point of Sale (POS), CRM data, historical ad audits, margin sheets, warehouse inventory levels | Provides the ground truth of internal business health, supply constraints, and unit economics. |
| **Market State** | Competitor ad creative tracking, social listening (TikTok/IG/LinkedIn), AI search rankings (AEO), earned media | Identifies external market movements, competitor pricing shifts, and customer sentiment changes. |
| **Decision Lineage** | Stored history of human operator approvals, campaign pivots, discount tests, and resulting revenue outcomes | Serves as the agentic memory layer, allowing models to learn from past wins/mistakes and refine recommendations. |
Dutta notes that decision history is historically discarded: an employee runs a discount, observes the result, leaves the company, and the institutional learning disappears. By embedding decision history directly into an agentic memory architecture, the platform learns which campaigns worked, which hypotheses failed, and how the human expert refined previous proposals. Over time, the system produces increasingly calibrated, higher-ROI strategies.
---
[13:40](https://www.youtube.com/watch?v=8aCXoSXbJqc&t=820s) Guardrails, Risk Tiers, and Continuous Marketing Execution
Addressing common failures in agentic automation—such as agents deleting critical directories or sending unauthorized emails—Dutta articulates a clear three-tier delegation matrix based on consequence and risk:
- **Low-Consequence Information Synthesis (Autonomous 24/7):** Competitive monitoring, market scraping, POS auditing, and multi-source data aggregation. Agents run continuously in the background while the team sleeps or codes.
- **Medium-Risk Tactical Execution (Guardrailed Delegation):** Creating new mutant ad creatives based on winning variants, reallocating ad budgets between Meta, Google, and TikTok, modifying `robots.txt` for AI search visibility, or drafting win-back email sequences. The agent prepares the end-to-end assets; the human operator reviews and clicks "yes" to push them live.
- **High-Risk Strategic Decisions (Human-Led):** Fundamental brand positioning, long-term pricing restructuring, and new product development remain firmly anchored in human executive judgment.
Dutta concludes this part by stressing that modern marketing has moved away from rigid, batch-based workflows that run once every few weeks. Instead, the brands winning market share are shifting to continuous, "always-on" micro-experimentation—constantly deploying variations, evaluating customer response loops, and iterating in real time.
[15:54](https://www.youtube.com/watch?v=8aCXoSXbJqc&t=954s) The Domain-First Agent Workflow and the Future of Autonomous SMB Growth
Looking ahead at the trajectory of marketing, the adoption of autonomous agents is poised to become standard operating procedure across the industry over the next few years. Onboarding onto a platform like usepomo.ai is designed to be completely frictionless: all an agent needs to get started is a single company domain. While connecting point-of-sale (POS) systems and integrating paid advertising accounts enriches the system's operational context, a simple website URL yields a surprising depth of actionable intelligence right out of the box. From just a domain scan, the AI extracts comprehensive insights about the broader market landscape, pinpoints the ideal customer profile (ICP), analyzes competitor positioning, and maps out existing customer sentiment.
```mermaid
flowchart LR
accTitle: usepomo.ai Domain to Execution Pipeline
subgraph Ingestion ["Data Ingestion"]
A["Company Domain"] --> B["Website Extraction"]
C["Ad Accounts & POS"] -.-> B
end
subgraph Intelligence ["Market Discovery"]
B --> D["ICP & Market Mapping"]
B --> E["Competitor & White Space Intel"]
end
subgraph Action ["Autonomous Execution"]
D --> F["Creative Launch & PDPs"]
E --> G["Ad Audits & Social Outreach"]
end
```
In early deployments across both D2C and B2B brands, teams running these workflows achieved massive leverage in just a matter of days. Rather than spending weeks stalled in agency retainer loops or internal approval chains, small teams were immediately able to perform ad account audits, uncover market white space, autonomously spin up high-converting product detail pages (PDPs), generate and launch fresh ad creatives, and directly engage prospective buyers talking about their brand across social media channels. As foundational AI models advance in multi-step, long-term reasoning tasks, this capability will increasingly empower small and mid-sized businesses to scale their growth engines and "do a lot more with less."
- **Input Requirements:** Fully operational from just a website domain, with optional integrations for ad accounts and point-of-sale systems to enrich performance feedback loops.
- **Immediate Outputs:** Competitor intelligence, ICP targeting, automated ad account audits, white-space opportunity mapping, dynamic PDP generation, and real-time social engagement.
- **Audience Q&A:** The presentation concluded with two and a half minutes reserved for questions from the audience regarding implementation and agent capabilities.
## Notable quotes
> [00:04](https://www.youtube.com/watch?v=8aCXoSXbJqc&t=4s) wouldn't fill up like half a room of this size — Praneet Dutta
> [09:13](https://www.youtube.com/watch?v=8aCXoSXbJqc&t=553s) water is wet — Praneet Dutta
> [06:49](https://www.youtube.com/watch?v=8aCXoSXbJqc&t=409s) spicy take — Praneet Dutta
> [15:54](https://www.youtube.com/watch?v=8aCXoSXbJqc&t=954s) do a lot more with less — Praneet Dutta
## Video description
Praneet Dutta - Founder, usepomo.ai
Google, Meta, and TikTok have spent a decade building AI that decides what your customers see, and until now that intelligence worked for the platforms, not for you. I spent years building AI at Google DeepMind, and now we're putting that class of agents directly in the hands of brands: an AI team that tracks competitors' ads, plans the month, produces the creative, and launches campaigns across Google, Meta, and TikTok, reporting back in revenue instead of impressions. Meanwhile, the 2,000-person brand is still routing that same work through approval chains and agency retainers. This talk shares unfiltered lessons from real small brands running this way today: where agents already outperform a department, where they fail, and the decisions you should never hand over. You'll leave knowing what to delegate on day one and why being small just became the advantage.
Praneet Dutta, founder of usepomo.ai and former Google DeepMind researcher, argues that autonomous AI agents allow small teams to outperform large enterprises by unifying fragmented data streams, automating tactical execution, and preserving institutional decision memory.
Replacing Knee-Jerk Reactions With Unified Intelligence (3:14)
Dutta points out that growth marketing has become severely fragmented across disparate systems, including retail point-of-sale (POS) metrics, warehouse inventory, Answer Engine Optimization (AEO), and organic conversations on platforms like TikTok and LinkedIn. For small operators with limited time, this fragmentation often prompts hasty, reactive decisions. For instance, when a competitor launches an evening discount, the typical reflex is to immediately match the price cut without auditing underlying numbers.
Using a hypothetical direct-to-consumer skincare brand called "Dewey Decimal," Dutta illustrates how an integrated agentic architecture circumvents knee-jerk spending. Instead of generating reactionary ads, the system correlates the competitor's campaign against trailing seven-day search volume, internal margin buffers, and retail versus e-commerce stock levels. This allows operators to determine whether the competitor is simply clearing dead inventory or targeting an unaddressed white space before choosing to counter, wait, or ignore the move entirely.
Small Pod Autonomy and Non-Consensus Insights (6:49)
Drawing on his experience at Google DeepMind—where he worked on frontier models including Imagen 3, Imagen 4, and Gemini Nano—Dutta contrasts corporate bureaucracy with startup velocity. Large enterprises find their greatest efficiency within small, cross-functional pods, yet often paralyze them with approval layers. Startups naturally operate as autonomous pods, but they lack dedicated personnel to manually aggregate POS records, scrape ad libraries, and monitor social channels.
However, Dutta warns that automating these processes with generic AI prompts yields useless platitudes, such as reporting that consumers drink tea during routines. True leverage requires uncovering non-consensus findings that link external shifts with internal operational realities, such as detecting specific market white spaces where customer acquisition has stalled alongside internal conversion slowdowns.
Three Growth Pillars and Decision Lineage Memory (10:14)
Dutta structures modern growth operations into three core components: the internal "Company Brain" (proprietary POS, CRM, margin, and inventory data), the external "Market State" (competitor ads, social sentiment, and AI search indexing), and "Decision Lineage." While most companies treat internal data and market tracking as standard point solutions, decision lineage is routinely overlooked.
In standard corporate environments, institutional memory vanishes when an employee tests a promotion or campaign and later leaves the company. By recording the historical lineage of human approvals, strategic adjustments, and resulting revenue outcomes within an agentic memory layer, the system continually learns which hypotheses succeeded or failed, delivering increasingly calibrated strategic recommendations over time.
Guardrailed Automation and Domain-Driven Deployment (13:40)
To prevent runaway failures—such as unauthorized emails or deleted databases—Dutta advocates a three-tier delegation model. Low-consequence tasks, like continuous market scraping and POS auditing, run autonomously around the clock. Medium-risk tactical tasks, including drafting mutant ad creatives, generating product detail pages (PDPs), and adjusting cross-platform ad budgets, are prepared end-to-end by agents but require explicit human sign-off. High-consequence choices, like core brand positioning and pricing overhauls, remain strictly human-led.
Dutta reports that this agentic workflow can initiate from a single company domain scan to extract ideal customer profiles (ICPs) and competitor positioning. In early B2B and D2C deployments, this setup shifted marketing from periodic, batch-based campaigns into continuous, real-time micro-experimentation, allowing a six-person team to manage complex multi-channel operations.
Dutta forecasts that autonomous agent workflows will become standard operating procedure for growth teams within a few years, though the ultimate ceiling for lean brands will depend on ongoing improvements in multi-step AI reasoning and disciplined integration of deep proprietary data beyond simple domain scans.
精炼阅读摘要
usepomo.ai 创始人兼前 Google DeepMind 研究员 Praneet Dutta 提出，自主 AI 智能体正在打破传统营销的规模壁垒；通过将内部运营数据、实时市场动态与决策谱系相融合，极简团队能够以超越大型企业的速度和精度完成战略分析与执行。
碎片化增长困局与数据驱动的战略合成 (3:14)
对于中小型企业与精益团队而言，增长决策往往受制于有限的精力和碎片化的技术工具。现代营销人员不仅需要复盘过去60天的预算分配与各 SKU 的利润率，还要应对生成式 AI 搜索推荐（如 ChatGPT 与 Claude 带来的回答引擎优化 AEO）以及社交媒体原生内容对用户心智的重构。过去，团队必须在付费广告、社交聆听、库存监控等割裂的独立软件之间频繁切换，消耗了大量宝贵的战略精力。
Dutta 举例指出，当直接面向消费者（D2C）的护肤品牌 Dewey Decimal 遭遇竞争对手突然推出的八折促销时，传统反应往往是未经计算便盲目跟进更大幅度的降价。而高效的战略合成则要求系统先行交叉分析多维数据：对手的折扣是否触达品牌的目标客户群体（ICP）、是在抢占空白市场还是清理滞销库存、自身产品的利润空间与渠道库存能否支撑防御。通过算法统筹分析过去7天的搜索热度与库存缓冲，团队方能在跟进反击、静观其变或完全忽略之间做出审慎抉择。
现代业务增长的三大支柱与决策谱系 (10:14)
Dutta 构建了一套支撑自主营销的底层框架，其核心由三大支柱组成：一是“企业大脑”（Company Brain），涵盖销售终端（POS）记录、CRM 销售漏斗、历史广告审计日志、各品类毛利与仓储库存等内部经营事实；二是“市场状态”（Market State），即竞品广告投放库、社交媒体舆情、AI 搜索引擎索引及公关声量等外部实时信号；三是“决策谱系”（Decision Lineage），即过去业务决策及其对应的后续商业成果记录。
在传统企业运营中，员工实施促销或调整策略后一旦离职，相关实践经验与制度记忆往往随之流失。Dutta 强调，将决策历史内嵌为智能体的长期记忆层，能够让 AI 系统明确知晓过去的成败原因，以及人类专家此前如何修正方案。借助这种持续迭代的因果反馈闭环，智能体在后续推荐中所给出的策略校准度与投资回报率将显著提升，从而摆脱浅层 AI 生成类似“顾客在日常生活中喜欢喝茶”这种显而易见且缺乏商业价值的表面洞察。
基于风险层级的自动化护栏与持续微实验 (13:40)
为规避自主智能体在执行层面可能造成的误操作风险，Dutta 提出了三级任务委派矩阵。第一层为低后果的信息合成任务，包括竞品监测、市场抓取和 POS 数据审计，由智能体在后台全天候无缝运行；第二层为中度风险的战术执行，涵盖根据优质素材衍生的变体广告、在 Meta、Google 和 TikTok 之间重新调配预算、修改 robots.txt 以优化 AI 搜索收录，或起草客户召回邮件。该类任务由智能体完成端到端资产构建，但必须由人类操作员最终审核并确认发布；第三层为高风险战略决策，核心品牌定位、长期定价重塑与新品研发仍严格保留在人类经营者的判断范畴。
Dutta 认为，现代品牌竞争已经彻底脱离了数周运行一次的周期性批处理模式。借助分级授权机制，企业得以转向“全天候”的持续微实验形态，在安全受控的边界内高频次投放广告变体、追踪用户转化反馈并实时迭代。
敏捷团队结构优势与低门槛落地路径 (6:49)
基于在科技巨头的工作经历，Dutta 对比了大型企业与敏捷初创团队的执行效率。尽管巨头坐拥资本储备、品牌资产和分销渠道，其最高效的业务单元依然是权责清晰、拥有自主权的小型跨职能团队，然而繁复的企业层级与冗长审批流往往阻碍了此类团队的推进速度。初创企业天然具备敏捷特性，当其借助 AI 代替专职团队完成繁重的底层数据汇总与初步分析后，核心成员得以将全部心智带宽聚焦于最终的商业战略裁决。
从应用落地来看，新型智能体系统的部署门槛正被大幅压低。Dutta 介绍，团队只需输入企业域名，系统即可自动解析行业全景、界定目标客群画像、分析竞品站位并提炼用户情绪；若进一步接入 POS 与广告账户，则能形成更精准的闭环。早期测试显示，D2C 与 B2B 团队在数天内即可独立运行广告账户审计、挖掘空白市场，并自主生成高转化率的产品详情页（PDP），展示出精益团队在多步骤复杂推理技术支持下的规模化运营潜力。
Dutta 总结认为，随着底层基础模型在长期多步推理能力上的演进，自主智能体将在未来数年内成为中小企业增长的标准配置。然而，尽管数据提取与执行自动化已显著降低运营门槛，核心的品牌战略、定价架构与高风险决策仍无法脱离人类专家的专业判断与治理边界。
Detailed notes — zh
## Background
该内容由频道 Ai4 于 2026 年 8 月 17 日发布在 YouTube 平台上。视频主讲人是 usepomo.ai 的创始人、前 Google DeepMind 研究员 Praneet Dutta，主要探讨小型团队如何借助新一代 AI Agent 统一营销决策与全自动化执行，从而打破大企业冗长审批与高昂代运营的枷锁，实现以小博大的市场竞争优势。
## TL;DR
Praneet Dutta 结合在 DeepMind 的前沿研发履历与创办 Pomo 的实战经验，指出消费决策链路正从传统渠道转向 AI 搜索与动态社媒，导致营销工具与数据严重碎片化。Pomo 通过构建由企业专有大脑、市场全局状态与决策血统组成的多智能体闭环系统，自动聚合 POS、广告与竞品动态，并提供深度非共识洞察。团队提出了低风险全自主、中风险人类一键确认、高风险人类掌舵的分级协作边界，让 2 人规模的极简敏捷团队也能依托强大的 Agentic Memory 实现超越千人品牌的增长迭代速度。
## Atmosphere & style
演讲兼具硅谷前沿技术专家的深厚积淀与初创企业构建者的实战锋芒，语调坦诚、犀利且充满激情。Praneet 毫无保留地分享了从大厂离职创业的心路历程，在剖析大企业组织冗余与平庸洞察时直击痛点、妙语连珠。整场分享节奏明快、逻辑严密，既有扎实的系统架构剖析，又充满赋能中小团队的感染力，带给观众极强的启发性与行动冲动。
## Key takeaways
- [00:04](https://www.youtube.com/watch?v=8aCXoSXbJqc&t=4s) 前沿 AI 正在成为强大的平民化工具，让几人规模的敏捷小团队拥有匹敌甚至超越大品牌营销部门的执行速度与能力。
- [02:01](https://www.youtube.com/watch?v=8aCXoSXbJqc&t=121s) 大模型正从被动的聊天机器人向能够自主执行复杂行动的多智能体（Agent）演进，成为驱动中小企业增长的核心杠杆。
- [02:01](https://www.youtube.com/watch?v=8aCXoSXbJqc&t=121s) 现代消费者决策链路向 ChatGPT/Claude 等 AI 搜索（AEO/GEO）及动态社媒转移，倒逼企业告别分散割裂的点状营销工具，转向统一智能体决策。
- [05:01](https://www.youtube.com/watch?v=8aCXoSXbJqc&t=301s) 面对竞品降价等市场冲击，优秀的营销决策不应是盲目跟风，而需交由 Agent 交叉分析客群画像、毛利率、库存与履约等全域数据后给出行动建议。
- [06:53](https://www.youtube.com/watch?v=8aCXoSXbJqc&t=413s) 大企业最高效的运作单元是具备自主权的独立小 Pod，而初创团队本身就是一个 Pod，其核心竞争力在于将繁琐数据处理交给 Agent，人类专注终极判断力。
- [09:13](https://www.youtube.com/watch?v=8aCXoSXbJqc&t=553s) 通用 Agent 输出的浅层常识毫无指导价值，高质量智能体必须能穿透海量噪音，挖掘出反直觉、非共识但高度可执行的深度商业洞察。
- [09:47](https://www.youtube.com/watch?v=8aCXoSXbJqc&t=587s) 未来企业智能系统由三大支柱构成：整合内部 POS/CRM 的企业专有大脑、聚合全网广告与 AI 搜索可见度的市场全局状态，以及记录决策脉络的决策血统系统。
- [11:32](https://www.youtube.com/watch?v=8aCXoSXbJqc&t=692s) 多智能体闭环工作流涵盖后台 24/7 侦察、因果分析建议、人类专家审批确认、多渠道自动化执行以及业务数据回流迭代的 Agent 记忆库。
- [12:24](https://www.youtube.com/watch?v=8aCXoSXbJqc&t=744s) 在完整闭环下，获得人类授权的 Agent 可全自动接管老客挽回邮件、广告素材搭建、SEO/AEO 优化及精准促销等全渠道落地执行。
- [13:40](https://www.youtube.com/watch?v=8aCXoSXbJqc&t=820s) 企业需根据风险层级划分智能体权责：低风险任务全自主 24 小时运转，中风险任务在设防规则下一键审批执行，高风险长期战略则由人类绝对掌舵。
- [13:40](https://www.youtube.com/watch?v=8aCXoSXbJqc&t=820s) 现代营销模式已从数天/数周一次的批处理运作，全面升级为“永远在线、持续微调、高频反馈”的常态化连续运转机制。
- [15:22](https://www.youtube.com/watch?v=8aCXoSXbJqc&t=922s) 中小品牌仅需提供一个官网 URL 并打通基础数据，即可借助具备长程推理能力的 Agent 快速完成账户审计、利基市场挖掘和自动化增长跃升。
## Detailed walkthrough
### [00:04](https://www.youtube.com/watch?v=8aCXoSXbJqc&t=4s) 告别大厂投身创业：AI 时代属于精简敏捷的小团队
主讲人 Praneet Dutta 开场向观众致意，并附上了自己的 LinkedIn 二维码以便后续交流。他坦言今天的分享对自己而言非常私密且意义非凡。在大学毕业后，他直接进入大科技公司工作了 8 年多，积累了丰富的企业级构建经验。然而就在最近，他毅然决定辞职创立自己的公司 usepomo.ai。支撑这一决定的核心信念在于：当前正处于一个不可思议的构建时代，在前沿实验室中看到的 AI 正在成为极其强大的“平民化工具”（democratizer），它赋予了具备行动力（agency）的团队前所未有的超高执行速度。
谈及过往背景，Praneet 表示自己从事应用 AI 领域已有十余年之久。彼时大语言模型（LLM）尚未诞生，卷积神经网络（CNN）也才刚刚起步。十年前一整场 AI 大会的参会者甚至坐不满半个会议室，而如今整个领域彻底爆发、受关注度呈指数级增长，这种转变令人惊叹。此前他在 Google DeepMind 担任前沿模型的研究员，主导了特定能力方向的研究，参与过 Imagen 3、Imagen 4 以及 Gemini Nano 等核心项目；此外，他还跨界参与过 AI 足球、可再生能源预测以及利用强化学习控制暖通空调（HVAC）系统等探索。
- **个人背景**：从大学毕业后在 Google 等大厂深耕 8 年以上，拥有十余年应用 AI 研发经验。
- **DeepMind 研发履历**：深度参与 Imagen 3、Imagen 4、Gemini Nano 等前沿模型，并研究过物理实体层面的 AI 应用（如 HVAC 温控、新能源预测）。
- **极简团队验证**：Pomo 整个平台仅由 6 个人开发完成，除开 Praneet 本人兼顾市场等杂务外，核心研发实际上仅靠 4 至 5 名工程师便搭建起了全套系统。
---
### [02:01](https://www.youtube.com/watch?v=8aCXoSXbJqc&t=121s) 营销科技的割裂困境与统一智能体的诞生契机
在前沿大模型研发过程中，Praneet 与团队观察到一个关键转变：AI 正在超越单纯的被动聊天机器人（chatbot），演化为能够替人类执行复杂行动的 Agent（智能体）。虽然硅谷的企业巨头最先跟进，但跳出硅谷圈子，绝大多数中小企业（SMB）在如何利用 AI 赋能增长方面依然处于极早期阶段，并存在诸多误解。这促使 Praneet 与早年在 Google 共同奋斗的联合创始人离职创办了 Pomo，旨在帮助精简团队实现飞跃。
对于中小品牌而言，少数几个人每天必须做出极其关键的商业抉择：过去 60 天广告预算如何高效分配、哪些高利润产品面临断货、哪些产品在零售渠道滞销等。与此同时，消费者的决策链路发生了巨变：新一代客户正转向通过 ChatGPT、Claude 等 AI 搜索（AEO/GEO）寻找产品建议，社交平台（TikTok、Instagram、LinkedIn）也充斥着动态的用户购买问询。在以往的 MarTech 领域，团队需要为广告投放、社媒舆情倾听、AI 搜索引擎优化等各个环节分别采购点状工具（Point-wise solutions），导致注意力严重碎片化。Pomo 的愿景正是利用日益强大的 Agent 统一全渠道信号，进行智能化决策整合。
```mermaid
flowchart TD
accTitle: 传统碎片化MarTech架构与统一Agent平台对比
subgraph S1["传统分散工具模式 (Point Solutions)"]
A["广告归因工具"]
B["社媒舆情监测"]
C["AI搜索(AEO)优化"]
D["库存与POS数据"]
end
subgraph S2["分散人工决策流"]
E["人工拼凑碎片数据\n高沟通成本与认知负荷"]
end
subgraph S3["Pomo 统一智能体平台"]
F["统一数据源采集\n(POS/社媒/广告/竞品)"]
G["Agent 深度交叉分析与提炼"]
H["战略建议\n(执行 / 观望 / 忽略)"]
I["人类专家进行高维判断"]
J["自动化执行与反馈闭环"]
end
A --> E
B --> E
C --> E
D --> E
F --> G --> H --> I --> J
```
- **决策维度剧增**：涵盖销售端数据（高毛利单品、库存深度）、AI 搜索曝光度（是否被 ChatGPT/Claude 准确索引与推荐）、社媒趋势洞察等。
- **工具割裂痛点**：传统模式下企业不得不维护多套独立系统，不仅耗费巨额订阅费与时间，还造成数据孤岛与反应迟缓。
---
### [05:01](https://www.youtube.com/watch?v=8aCXoSXbJqc&t=301s) 虚拟案例 Dewey Decimal：竞品降价下的非条件反射式决策
为了具象化说明问题，Praneet 举了一个虚构的护肤品牌“Dessie Dewey Decimal”（致敬图书分类法与数学概念）。假设 Dewey Decimal 是一个人员精干的护肤品牌，某天晚上竞争对手突然发动价格战，推出全场 20% 折扣（8折促销）。
第二天一早登录后台，许多运营人员的第一反应往往是条件反射式的跟进降价，甚至头脑发热地直接打出 25% 折扣，却完全不看数据支撑。但优秀的营销人员在砸钱投流或改价前，第一步永远是严审核心数据指标：
1. **目标客群契合度**：竞品促销的具体产品是否真正触及了我方的理想客户画像（ICP）？
2. **市场空白与竞争关系**：对方是在抢占某种细分蓝海，还是与我方核心爆品正面冲突？
3. **自身经营现状**：我方对应产品的毛利率是多少？当前库存是否充足？市场需求趋势如何？
4. **全渠道履约**：线下零售货架与线上独立站的铺货及履约状况是否允许承接这波流量？
Pomo 的核心逻辑正是让 AI Agent 在底层自动聚合 POS 销售终端、竞品动态和多渠道广告等海量数据，完成繁琐的数据清洗与交叉计算，最终给出一目了然的决策建议：究竟是应当“立即采取行动（Act）”、保持“观望态势（Hold & Watch）”，还是“完全忽略（Ignore）”。
---
### [06:53](https://www.youtube.com/watch?v=8aCXoSXbJqc&t=413s) 大企业 Pods 与创业团队：将精力聚焦于“人类判断力”
回顾在大厂的经历，Praneet 分享了一个略带犀利的观点（"spicy take"）：大企业虽然坐拥品牌资产、海量资本和专业分工优势，但真正最高效的运作单元往往是那些具备高度自主权、明确目标考核与最终决策权的“小型敏捷战斗组”（Pods）。例如一个专门负责某特定国家/地区单一产品线增长的小 Pod，其迭代速度远超臃肿的层级汇报链。
这种组织机制与 AI 时代的精简团队有着异曲同工之妙。在初创企业或 SMB 中，整个公司本质上就是一个单独的 Pod。小团队没有冗余的跨职能部门去层层拉扯，因此核心竞争力必须落在**人类的终极判断力（Human Judgment）**上。
- **底层与后端繁重工作**：多平台信息抓取、POS 报表整合、基础数据提炼全部移交给具备超高效率的 AI Agent。
- **顶层战略把控**：人类管理者负责审视 Agent 最终交付的综合报告，评估战略合理性，决定资源走向，再指令 Agent 执行末端投放与配置。
---
### [09:13](https://www.youtube.com/watch?v=8aCXoSXbJqc&t=553s) 告别平庸废话：Agent 需要挖掘非共识的深度洞察
在谈及“人类判断力”与 Agent 的配合时，Praneet 强调了洞察质量的重要性。如果只是给基础的 Claude 或通用 Agent 发送极其宽泛的抓取指令，它产出的结论往往流于表面、毫无价值。
比如让 Agent 分析饮品市场，它可能会煞有介事地汇报：“越来越多消费者喜欢在日常生活中喝茶”。Praneet 调侃道：“这简直就像在说‘水是湿的’一样毫无意义。”全行业人尽皆知的浅层信息对业务没有任何指导作用。真正高价值的系统，必须能够帮助团队穿透海量噪音，挖掘出那些具有**非共识性（non-consensus）**、反直觉甚至带有争议性但高度可执行的深度商业洞察。
### [09:47](https://www.youtube.com/watch?v=8aCXoSXbJqc&t=587s) 统一数据信号与企业大脑的三大支柱
如果让营销团队花大量时间深挖数据，往往耗时费力且难以迅速得出结论。而通过将所有分散的数据源串联起来，系统能够主动发出预警并提出建议，例如敏锐捕捉到“竞争对手正在占领某块市场空白，而你自身产品的转化率正在放缓”。这种将多方信号统一、直接提炼出确定性结论并赋能下游营销目标的能力，正是未来商业智能的核心所在。
预测未来企业运作形态的演变，主要取决于三大核心要素：
- **企业专有大脑（Company Brain）**：整合 POS 销售终端、CRM 客户关系管理系统、广告历史审计等企业内部专有数据资产。
- **市场全局状态（Market State）**：聚合高度碎片化的外部数据，覆盖全渠道广告投放、社交媒体互动、AI 搜索引擎可见度（AI Search Visibility）以及口碑媒体（Earned Media）等。
- **决策历史与决策血统（Decision History / System of Decision Lineage）**：过去企业的决策逻辑往往深锁在部门孤岛中（例如某次打折促销的决策背景与后续表现往往随着项目结束而被遗忘丢弃）。在 Agent 与人类协同的新模式下，Agent 向专家学习“该做什么与不该做什么”，记录完整的决策脉络，从而在后续闭环中持续优化推荐质量。
### [11:32](https://www.youtube.com/watch?v=8aCXoSXbJqc&t=692s) Agent 协作闭环与决策记忆系统
企业未来的工作流将演变为一个由多智能体协同运转的闭环系统。前端 Agent 负责 24 小时全网侦察与数据采集；分析型 Agent 负责交叉比对内外部数据，建立因果与相关性分析，输出具备可操作性的具体判断；营销专家则担任审核与把关角色（选择同意、驳回或微调方案）；一旦人类确认放行，执行 Agent 就会迅速展开行动，无论是上线广告活动、调整定价还是推出促销策略。
```mermaid
flowchart TD
accTitle: Agentic Closed Loop Workflow
subgraph ScoutAndAnalyze["数据感知与分析层"]
A["后台侦察 Agent<br/>(24/7 市场与竞品监控)"] --> B["分析 Agent<br/>(数据因果关联与决策建议)"]
end
subgraph HumanApproval["人机协同层"]
B --> C{"人类专家把关<br/>(确认 / 修改 / 驳回)"}
end
subgraph ExecutionAndFeedback["执行与记忆层"]
C -- "确认授权" --> D["执行 Agent<br/>(投放广告/修改定价/内容部署)"]
D --> E["业务结果追踪<br/>(B2B/D2C 闭环成效)"]
E --> F[("企业级 Agentic 记忆库<br/>(决策血统与历史沉淀)")]
F -. "自我迭代与优化".-> B
end
```
在执行完毕后，所有真实的业务表现数据都会沉淀回企业内部的“Agent 记忆库（Agentic Memory）”中。下一次生成策略时，系统会基于历史成功与失败的经验持续自我迭代，使推荐越来越聪明。
### [12:24](https://www.youtube.com/watch?v=8aCXoSXbJqc&t=744s) 护肤品牌实战：从全域归因到多渠道全自动执行
以精华液（Serum）护肤品牌为例，当竞争对手突然上线了一项极具冲击力的新促销政策时，Agent 并不会盲目冲动地“立刻新建一条广告”。相反，它会全面扫描并综合计算多个维度的实时数据：
- 竞品具体优惠力度与条款
- 过去 7 天内全网关于“精华液”的搜索趋势热度
- 自身各渠道表现（付费广告投放数据与社媒自然流量）
- 品牌在 AI 搜索引擎中的可见度与推荐排名
- 企业当前的实时库存水平与商品利润率空间
在统揽全局后，Agent 会输出多种策略选项供运营者选择。一旦获得人类确认，系统便会全权接管下游落地执行：自动起草老客挽回邮件（Email Win-Back Campaign）、生成针对性广告素材、修改官网的 `robots.txt` 以优化 AI 搜索引擎的抓取索引、在站点发布针对用户痛点的客户案例，或直接上线定向促销活动。那些保持高频实验并从中持续学习的企业，其增长势能将远远甩开依赖单次偶然尝试的传统团队。
### [13:40](https://www.youtube.com/watch?v=8aCXoSXbJqc&t=820s) Agent 的权责边界：低风险放行、中风险设防、高风险由人类掌舵
面对许多人对自主 Agent 曾出现误删文件、误发邮件等失控问题的担忧，企业必须根据任务的风险等级划定明确的协作边界：
| 风险层级 | 典型应用场景 | 运作模式与人类参与度 |
|:--- |:--- |:--- |
| **低风险 / 低后果任务** | 知识提炼（Knowledge Synthesis）、竞品 24/7 监控、全网市场情报收集分析 | **全自主运行**：无需人类值守，可在后台或夜间 24 小时常态化运转 |
| **中等风险 / 客户触达任务** | 基于新变体素材（Mutant Creatives）搭建广告组、在 TikTok / Google / Meta 之间灵活迁移预算 | **设防执行（Defined Guardrails）**：Agent 提出具体调整方案，人类一键点击确认后由系统执行并持续监控 |
| **高风险 / 长期战略任务** | 品牌核心定位（Brand Positioning）、全新产品研发与立项决策 | **人类主导决策**：高度依赖业务负责人的直觉与商业判断，Agent 仅在底层提供决策辅助分析 |
正如现今优秀消费品牌与 B2B 企业的实践所示，现代营销已经彻底告别了过去“每隔数周或数天运行一次”的批处理（Batch-based）模式，全面转向“永远在线、持续微调、快速反馈”的常态化连续运转机制。
### [15:22](https://www.youtube.com/watch?v=8aCXoSXbJqc&t=922s) 从极简输入到全能增长：用极简起点赋能中小品牌
将强大的 AI 营销能力落地并不复杂。在系统平台上，用户甚至只需要提供一个企业官方网址（Domain URL），Agent 便能深入挖掘出极其丰富的洞察：包括目标客户画像（ICP）、市场定位盲区、竞品动向以及社交网络上的真实用户声量。随着进一步打通 POS 系统与广告投放账户，数据飞轮还将变得更加强劲。
在与多家 D2C 和 B2B 品牌的实际合作中，即便只接入短短几天，这些小团队便展现出了惊人的效能跃升：
- 高效完成深度广告账户审计并快速上线创意素材
- 准确锁定未被满足的利基市场空白（White Space）
- 自动化优化生成高转化率的产品详情页（PDP）
- 主动搜寻社交平台上讨论品牌的用户并开展精准互动
随着 Agent 在长程推理（Long-term Reasoning）任务上的持续突破，中小企业（SMB）完全能够打破传统团队规模与高昂代运营费用的枷锁，以极轻量的人员配置实现超越巨头品牌的增长效率。
## Notable quotes
> [00:04](https://www.youtube.com/watch?v=8aCXoSXbJqc&t=4s) 在前沿实验室中看到的 AI 正在成为极其强大的“平民化工具”（democratizer），它赋予了具备行动力（agency）的团队前所未有的超高执行速度。 — Praneet Dutta
> [00:04](https://www.youtube.com/watch?v=8aCXoSXbJqc&t=4s) Pomo 整个平台仅由 6 个人开发完成，除开 Praneet 本人兼顾市场等杂务外，核心研发实际上仅靠 4 至 5 名工程师便搭建起了全套系统。 — Praneet Dutta
> [02:01](https://www.youtube.com/watch?v=8aCXoSXbJqc&t=121s) AI 正在超越单纯的被动聊天机器人（chatbot），演化为能够替人类执行复杂行动的 Agent（智能体）。 — Praneet Dutta
> [06:53](https://www.youtube.com/watch?v=8aCXoSXbJqc&t=413s) 大企业虽然坐拥品牌资产、海量资本和专业分工优势，但真正最高效的运作单元往往是那些具备高度自主权、明确目标考核与最终决策权的“小型敏捷战斗组”（Pods）。 — Praneet Dutta
> [09:13](https://www.youtube.com/watch?v=8aCXoSXbJqc&t=553s) 真正高价值的系统，必须能够帮助团队穿透海量噪音，挖掘出那些具有非共识性（non-consensus）、反直觉甚至带有争议性但高度可执行的深度商业洞察。 — Praneet Dutta
> [09:47](https://www.youtube.com/watch?v=8aCXoSXbJqc&t=587s) 在 Agent 与人类协同的新模式下，Agent 向专家学习“该做什么与不该做什么”，记录完整的决策脉络，从而在后续闭环中持续优化推荐质量。 — Praneet Dutta
> [12:24](https://www.youtube.com/watch?v=8aCXoSXbJqc&t=744s) 那些保持高频实验并从中持续学习的企业，其增长势能将远远甩开依赖单次偶然尝试的传统团队。 — Praneet Dutta
> [13:40](https://www.youtube.com/watch?v=8aCXoSXbJqc&t=820s) 现代营销已经彻底告别了过去“每隔数周或数天运行一次”的批处理（Batch-based）模式，全面转向“永远在线、持续微调、快速反馈”的常态化连续运转机制。 — Praneet Dutta
> [15:22](https://www.youtube.com/watch?v=8aCXoSXbJqc&t=922s) 随着 Agent 在长程推理（Long-term Reasoning）任务上的持续突破，中小企业（SMB）完全能够打破传统团队规模与高昂代运营费用的枷锁，以极轻量的人员配置实现超越巨头品牌的增长效率。 — Praneet Dutta
## Mentioned
Praneet Dutta, usepomo.ai, Google, Google DeepMind, Meta, TikTok, Imagen 3, Imagen 4, Gemini Nano, ChatGPT, Claude, Instagram, LinkedIn, Dessie Dewey Decimal
## Video description
Praneet Dutta - Founder, usepomo.ai
Google, Meta, and TikTok have spent a decade building AI that decides what your customers see, and until now that intelligence worked for the platforms, not for you. I spent years building AI at Google DeepMind, and now we're putting that class of agents directly in the hands of brands: an AI team that tracks competitors' ads, plans the month, produces the creative, and launches campaigns across Google, Meta, and TikTok, reporting back in revenue instead of impressions. Meanwhile, the 2,000-person brand is still routing that same work through approval chains and agency retainers. This talk shares unfiltered lessons from real small brands running this way today: where agents already outperform a department, where they fail, and the decisions you should never hand over. You'll leave knowing what to delegate on day one and why being small just became the advantage.
[Source](https://www.youtube.com/watch?v=8aCXoSXbJqc)
[0:04] Praneet Dutta: wouldn't fill up like half a room of this size
[9:13] Praneet Dutta: water is wet
[6:49] Praneet Dutta: spicy take
[15:54] Praneet Dutta: do a lot more with less
Original transcript
[0:04] Thank you all for joining in for my talk. My
name's Praneet. I put in my LinkedIn QR code
[0:08] if anyone want to connect after the talk. Today
I'm going to be sharing something that's very
[0:14] personal to me. I was in big tech for more than
8 years. I joined straight out of university.
[0:20] Spent a lot of time building in a great
enterprise, right? And very recently,
[0:25] I'll get into it, I quit to build my own company,
right? And our conviction behind this is that now
[0:30] is an incredible building, right? AI, the
way we see it working in the frontier lab,
[0:36] is a democratizer. It enables teams with agency
to move a lot faster, right? So, you know,
[0:42] we quit to actually help SMBs, like most of you.
Actually, how many SMBs are out here in the crowd?
[0:49] Okay, the rest of you enterprise? No?
[0:53] Okay, I guess we'll see in the questions. But,
Yeah, you know, I spent over a decade building in
[1:00] applied AI, right? Back before LLMs were a thing
and convolutional neural networks were still,
[1:05] you know, just getting set up. So, it's been
an interesting experience to see how, you know,
[1:10] the field that, you know, was probably
wouldn't fill up like half a room of this size,
[1:14] like for an entire conference like 10 years ago,
has now exploded and there's so much interest,
[1:18] right? It's amazing to see that. Specifically, I
was in Google's DeepMind group. I was a researcher
[1:23] in the frontier models. I led aspects in certain
capabilities. So, ImageNet 3, ImageNet 4,
[1:30] and then Gemini Nano Bananas. Worked on a
bunch of other stuff, you know, AI for soccer,
[1:36] AI for renewables forecasting, reinforcement
learning for controlling HVAC systems.
[1:40] So, if you ever interested in talking about
how, you know, the broader applications behind
[1:43] beyond chatbots and agents and LLMs in
atoms, not bits, I'm also happy to chat.
[1:51] So, as I mentioned, you know, it it was an amazing
time to be building. I I working on the frontier
[1:56] of generative AI models, right? But an insight
that we saw was as we heard from the previous
[2:01] presenters, the models were getting increasingly
more intelligent, right? They were able to move
[2:06] from beyond a chatbot towards something that acts
for you, right? And we saw a huge opportunity
[2:11] there. Obviously, enterprise, you know, are some
of the first to jump and you we heard about token
[2:15] mixing. Um all a lot of it's been happening,
right? But amongst like other smaller companies
[2:21] that's outside our bubble here in Silicon Valley,
there are a lot of misconceptions and you know, I
[2:25] would say they're pretty early in the days of how
they leverage AI to grow their teams. Obviously,
[2:29] there are few companies that have been outstanding
like I've seen lean teams do amazing things,
[2:33] right? Um and that's what kind of encouraged me
and my co-founder who we worked together in our
[2:38] early days at Google to leave and start Pomo
primarily for like lean growth teams, but you
[2:43] know, it's been it's been incredible for us to see
like all the way from like SMBs to enterprise use
[2:47] our platform, right? And as a lot of testament
to how we doubled down our own thesis of how,
[2:53] you know, small teams can do incredible things,
we built an entire platform with just six people,
[2:58] right? Like I mean and that includes folks like
me that spends most of my time on marketing and
[3:02] everything else, right? So, probably like four
to five engineers built out an entire platform
[3:06] on Pomo which I'm about to show you in the next,
you know, 15 minutes that we have left, right? Um
[3:14] So, before I actually talk about, you know, what
the actual advantage is, right? If you think of a
[3:18] startup for those of you like, you know, in the
SMB space might also it's resonate with them.
[3:23] You have few people that have to make a
handful of decisions, but these are very,
[3:27] very important decisions, right? How did I best
allocate budget over the last 60 days, you know,
[3:33] like what did I do? How did it perform, right?
Um how did my sales work out, right? Like what's
[3:39] selling out? Like which what are high margin
products? What are those that are probably not
[3:43] being stocked in retail stores? What are those
that have, you know, high visibility and people
[3:47] are talking about them, right? On visibility, you
know, as you know, consumers these days are moving
[3:53] a lot to AI search. Your next customer will be
actually asking the chat GPT or cloud agent in
[3:57] terms of what's happening. This is dynamic,
you know, the rules have changed slightly.
[4:01] How do you get indexed by the right audience for
the right things? What's your products, you know,
[4:06] call to action, right? And are you being indexed
over there, right? And then social as a channel,
[4:12] as we all know, has also been growing. Like buyers
have been, you know, asking recommendations on the
[4:16] TikToks, Instagram, LinkedIns of the world. How
do you get in front of those conversations? How
[4:21] do you make sense of like what people are talking
about for what they're what you can provide now
[4:26] versus future product development, right? So, as
you can see, there's like tons of sources, right?
[4:32] Um in the MarTech space that we are in, before
agents became mainstream, you would probably need
[4:37] a different point-wise solutions for each of
these. One for campaigns, one social listening,
[4:41] one for AEO, right? Um but what we saw was like
obviously fragmentation. Your time and attention
[4:48] is valuable, right? Agents are getting a lot
better, and we envision a single platform that
[4:53] can help you unify all of these different signals
and make decisions a lot more intelligently.
[5:01] So, um for the purpose of today's presentation, I
made up a fictional brand, Dessie Dewey Decimal.
[5:06] For those of you, I mean, in in maths, you will
kind of get the reference. Um so, we'll work with
[5:11] Dewey Decimal's case over here. So, let's say
they are a skin care brand, right? Um you know,
[5:17] it's a lean team, you know, they're trying to like
get the skin care products out there. And let's
[5:21] say on one fine evening, their competitor launches
a promotion saying that they are doing a 20% off,
[5:26] right? Now, when they log in back in the morning,
a knee-jerk reaction can be that, you know what,
[5:32] 20% off, let's match that, you know, let's do 25%
off, right? Uh without looking at any of the data.
[5:39] But what we've been hearing and what I
think is the prudent thing to do from
[5:42] smart marketers is that before you actually
go and like spend more ad or marketing budget,
[5:48] you first look at the numbers, right? Some
decisions can be does this product that they
[5:52] are launching promotion on is it actually relevant
to our ICP, right? Is there a certain white space
[5:57] in the market they're capturing? How are our
products doing? You know, are they high margin?
[6:01] Do we have inventory? Is there demand, right?
Um do we directly compete as I mentioned? Um
[6:07] are we stocked in retail stores if you're selling,
you know, retail or on e-commerce, right? So, what
[6:12] we are hearing from marketers is that they want
to first understand not only like what happened,
[6:16] but strategically what's the best way for them to
react to this, right? Before they go and just make
[6:21] a decision and spend like a bunch of, you know,
dollars on ads or some other campaign, right?
[6:26] So, that's kind of like what we are hoping to be
doing here at Pomo, right? We want to like do all
[6:31] the data crunching, have have agents like look at
the individual data sources, see what's working,
[6:36] what's not, and then give you recommendation in
terms of like something we think you should act
[6:40] on, something you think you should hold and watch
how it's going, or probably just ignore, right?
[6:49] So, this kind of comes back to taking a bit of a
segue, right? Like when you think about I I was in
[6:53] Big Tech, right? Um I'm not saying that startups
are going to like, you know, just kill it and do
[6:58] amazing, which, you know, it will do, but I I
think enterprises are also uniquely positioned
[7:03] in this case, right? Uh they do have brand value,
they do have distribution, they have, you know,
[7:08] specialists, they do have, you know, enough
capital to invest in different bets, right?
[7:13] But the thing about enterprise and some of
you might resonate with some of you is that
[7:19] the most efficient teams, at least in my
case, right? It's my my my spicy take,
[7:23] were like smaller pods in larger companies
that have clear autonomy, have clear direction,
[7:28] and have clear um sort of like um um say over
the end results, right? Those teams seem to
[7:35] have moved the fastest, right? It would probably
just be siloed like let's say you have like one
[7:39] enterprise pod that works on one product in one
geography, and their goal is for like, you know,
[7:44] success in that one specific region over a period
of time, right? In an enterprise you probably have
[7:49] like 1,000 teams like those, right? But each of
these can be kind of like brought into a like a
[7:52] unique pod that's obviously cross-functional,
but, you know, they work faster, right?
[7:58] The same principle kind of follows here
though, right? Like it's like a small team
[8:02] with AI agents the opportunity is there where
their judgment is where most of the time should
[8:06] be spent, right? All of the information gathering
in terms of looking at your point of sales, all
[8:11] the other signals, those are things that LLMs and
AI agents are uniquely positioned to summarize and
[8:16] do very quickly, right? And then they spend their
time looking at the final synthesis and saying
[8:20] that does this make sense or not? And then finally
agents can also act on the other side of things,
[8:25] right? So, you know, enterprises and we've been
seeing, you know, companies we've been talking to
[8:29] do really, really well in this regard and kind
of reorganizing teams around this, right? Um
[8:36] but startups, you know, that one pod is your
entire company, right? Your time is valuable,
[8:41] you don't have cross-functional teams, right? Um
for SMBs you need to make sure how much of this
[8:47] work can you outsource to something that's
reasonably good and fast and gets you there
[8:51] while you focus on what matters, which is a human
judgment, right? And I think that's the principle
[8:57] how things are going to be going. Like all of us
we all have like specialty kind of domains that we
[9:01] work in. We'll see agents taking up more of the
back end work, the mind-boggling work, while as
[9:05] we focus on the judgment that actually matters
in terms of guiding them and working with them.
[9:13] So, when you talk about judgment, we can look
at a specific example of what this means, right?
[9:18] So, in this case, um we, as I mentioned,
our time is very important here, right?
[9:23] If we ask agents to just like, you know,
you ask a cloud agent to scrape the market,
[9:27] give you an insight, it might come up with
something super shallow. Like in this case,
[9:31] customers increasingly drink tea
during daily routines, you know?
[9:34] That's like saying water is wet, right? I mean,
everyone in the in the domain knows this, right?
[9:38] But, what you actually want is a unique, perhaps
you know, non-consensus or a controversial insight
[9:43] that tells you something new which you can
act on, right? Probably you would need like a,
[9:47] you know, like a lot of more time to kind of
go deep through data and from this conclusion.
[9:51] But, what if it tells you something that's okay,
you know, the competitor's winning this white
[9:55] space while your conversion's flattening.
Uh this is what we see is that's happening
[9:58] when we link together all the data sources and
this is what we recommend you do. So, it kind
[10:03] of unifies the signals, give you a conclusion
and then you can act on that, right? This is
[10:07] a decision that you can actually use downstream
and, you know, use towards your marketing goals.
[10:14] So, in terms of how I think businesses are going
to change, again, this is our prediction, right?
[10:18] Um we see com- we see like three unique factors,
right? The first two might be more obvious. You
[10:22] have a company kind of like information or brain,
right? You have your point of sales, your CRM,
[10:28] your advertising audits, you know, your
company proprietary kind of like brain, right?
[10:33] Then you have the market state,
right? These days, you know,
[10:35] people are increasingly focusing the marketing
efforts across different channels, right? So,
[10:40] you can look at the fragmented data across ads,
social media, AI search visibility, and, you know,
[10:46] uh earned media, and a bunch of
other channels as well, right?
[10:50] The third thing, decision history, that's
where things get very interesting, right?
[10:55] As I mentioned, you know, like previously,
a lot of company proprietary information,
[10:58] like in terms of how decisions are made, is
lost in silos, right? Someone might decide
[11:03] to discount a certain product and then
launch it and you see how it performs.
[11:07] That information is mostly thrown away, right?
But, in this new world where we see agents and
[11:13] human working together and agents learning from
experts in terms of what to do or not to do,
[11:17] this decision, what we call the system of
decision lineage, becomes more and more important.
[11:22] Because once an agent works with you to go
through an entire workflow, it's able to see
[11:26] what works, what doesn't work, learn, and keep
optimizing and making smarter recommendations.
[11:32] So, this is sort of a loop that we envision
that'll keep happening, right? We'll have, you
[11:35] know, agents doing all the back-end work for you,
scouting, seeing what's out there in the market,
[11:39] right? Based on that, you can have a separate set
of agents that take these kind of sort of like
[11:43] data sources, and then try to build causations
or correlations, and give like, you know,
[11:48] tasks that are things you can act on, like
judgments, right? Then it works with you,
[11:52] the expert, in terms of like what, you know, do
you agree, disagree, do you want to edit or not?
[11:57] Once you give a thumbs-up, then prepares the work,
right? Whether it's a campaign, whether it's like
[12:01] a pricing change, a promotion, right? And then you
execute, right? B2B is now great become a great
[12:07] place where you can actually have a closed-loop
system. You can see how things are doing, right?
[12:11] All of this information will now be stacked in
your company's internal agentic memory, right?
[12:15] And every time you have a new recommendation,
[12:17] it learns from what worked and not worked
in the past, and keeps iterating from there.
[12:24] Um so, if we go back to the example of, um you
know, um the serum brand we spoke about, right?
[12:29] If let's say your competitor launches a new offer
that, you know, affects, you know, your market,
[12:34] it scrapes through all this information
in terms of looking at the offer,
[12:37] you know, serum searches last 7 days, your
channel performance, both paid as well as social,
[12:42] how are you performing in terms
of like, uh your AI search,
[12:46] um what's your margin and inventory like, and
then based on that, it then comes at conclusion.
[12:51] It doesn't immediately decide that, oh, it's time
for me to go ahead and create an ad, right? So,
[12:55] it comes up with like different options like this,
which you can iterate on, and it goes from there.
[13:00] And then downstream, once you say
yes, you know, at least in Pomo,
[13:03] and we see this happening more generally, it
goes ahead and handles the execution, right?
[13:07] Whether it's like drafting an email win-back
campaign, whether it's like generating creatives,
[13:11] whether it's making changes to your, you know,
website's robots.txt, so you're better indexed
[13:15] on AI search, whether it's placing a new
case study that answers a question, you know,
[13:19] whether it's like launching a promotion, right?
All of this is stuff that, you know, it goes
[13:23] ahead and does it for you. So, what we kind of see
that is going to be happening over here, right?
[13:29] Companies that keep experimenting and keep trying
out new experiments and learning from those will
[13:34] do better than those that do a single shot and
then, you know, try to get an influence from that.
[13:40] Now, this slide is pretty important. This is
where I think a lot of you will probably have
[13:43] follow-up questions on, right? You know, we How
many of you have used OpenClaw in the past? Right.
[13:50] So, some of you would have seen things go wrong,
right? You know, you know, writing like emails
[13:54] that shouldn't have been sent or deleting files
and folders, right? Um this is where our take is.
[14:00] Agents are great at knowledge synthesis, you
know, getting together different, you know,
[14:04] um data sources and being able
to get inferences from that.
[14:07] So, as a sort of like analyst or a competitive
monitoring or an information gathering tool for
[14:12] almost all tasks, you know, they are quite good.
They can be running while you sleep, you know,
[14:16] while we are coding or doing other things. I run
my agents overnight while it does stuff for me.
[14:21] So, for things like that that are low consequence
and stuff more on background information
[14:24] gathering, we run our agents 24/7, right? Now, as
you start moving more towards the output side of
[14:32] stuff, like customer-facing sort of aspects,
that's where things get a bit more trickier.
[14:36] Um this obviously depends more on a
case-by-case basis, but what we see is
[14:40] for things that are low stakes, for example,
like creating advertising campaign with a
[14:44] new creative mutant that wasn't performed in
well in the past, right? Or changing budget or
[14:50] from one platform to the other, from let's
say TikTok to Google to Meta Ads, right?
[14:54] These are things where you want agents to execute,
but with defined guardrails. So, it might give
[14:57] you a recommendation, you click yes, and it goes
ahead and does it for you and monitors. And then,
[15:02] for the last part, you know, which are things
which are more high-risk and more long-term,
[15:05] you know, like positioning or creating a new
product, this is something where we feel like, you
[15:10] know, our marketers or, you know, your business
owner's judgment matters the most, right? And
[15:15] we do see agents helping out more on a judgment
stack over here, but this will take a bit longer.
[15:22] And, you know, to kind of close the loop on
this, as I mentioned, we see companies that are
[15:26] able to execute faster, test out many different
micro campaigns, promotions, offers, you know,
[15:32] those that are able to learn, see how
consumers or their customers react,
[15:36] and then adapt on the basis of that. That's
where we see some of the best companies will win,
[15:40] right? I mean, you you've seen articles from,
you know, the best consumer and B2B brands that
[15:44] marketing's moving more from a process that's more
batch-based, it runs once every few days or weeks,
[15:49] um to something that runs continuously, it's
always on, right? So, you know, my my bold
[15:54] prediction is that this is what we're going to see
most companies do in the next few years, right?
[16:01] Cool. Um I know we're a bit short on time, so I'll
maybe just skip to what's more relevant, right? Um
[16:06] if I just go back here. So, you know, in in a
platform like ours, it's pretty simple. All we
[16:11] need is your company domain, right? You can start
putting in more information like your point of
[16:15] sales, integrate your ad accounts, and that's when
information get richer. But even from just your
[16:20] website, you'll be amazed to know how much you
can learn about not only your market, your ICP,
[16:24] who you're targeting, what your competitors
are doing, what people are talking about you.
[16:28] It's very, very information-rich that even
we were surprised about, right? So, you know,
[16:33] early days, but we worked with like a few brands,
both D2C and B2B, and we saw like even in the
[16:37] order of a few days, they were able to like do
so much more, launch creatives, better ad audits,
[16:42] understand white space, you know, generate better
PDPs, right? Reach out to customers talking about
[16:47] them on, you know, social media platforms, and get
more done, right? And this is just the beginning.
[16:53] So, you know, we're very excited as agents get
more and more um advanced in long-term reasoning
[16:59] task, how much more can they help brands um
optimize their marketing and growth efforts,
[17:04] and you know, scale SMBs to do a lot more with
less. So, I have exactly 2 and 1/2 minutes left,
[17:11] so I'll probably leave it now for any
questions um that the audience has for me.
