Ex-DeepMind Lead Says Foundation Models Commoditize as Agent Systems Win
Ai4
Foundation models are rapidly commoditizing into baseline utilities, shifting enterprise value to contextual agent harnesses that fuse internal business data with external market signals.
## Background
Published on YouTube by the Ai4 channel on August 15, 2026, this conference presentation features Praneet Dutta, founder of Pomo and former Google DeepMind GenAI research tech lead. The session outlines the shift from narrow, isolated marketing copilots to autonomous growth systems capable of analyzing market dynamics, orchestrating strategy, and iteratively learning through closed-loop human-in-the-loop controls.
## TL;DR
Drawing on his foundation model research at Google DeepMind, Praneet Dutta argues that raw model capabilities are commoditizing, shifting defensible value to agentic harnesses that can reason across complex, multi-source enterprise data. Rather than relegating generative AI to superficial promotional copy or reactive price cuts, businesses must architect a 'living commercial model' that pairs internal operational state like unit margins and ICPs with external signals like competitor maneuvers and generative search visibility. Dutta outlines a calibrated governance framework where 24/7 background research and low-risk tactical shifts operate autonomously, while high-stakes brand and pricing decisions remain guarded by human experts. Finally, he introduces 'decision lineage' as an operational feedback loop that records and learns from every intervention, transforming marketing into a measurable, compounding science.
## Atmosphere & style
The presentation carries the pragmatic, demystifying tone of a seasoned researcher stepping out of the AI lab and into the messy realities of commercial execution. Dutta speaks with understated authority and candid realism, puncturing industry illusions around zero-shot autonomy and soaring API costs without cynicism. The pacing is focused and analytical, making the session feel less like a polished vendor pitch and more like a structured engineering briefing on how to bridge frontier models with defensible enterprise software.
## Key takeaways
- [00:04](https://www.youtube.com/watch?v=6jYv47iAQis&t=4s) Foundation models are rapidly commoditizing as open-source alternatives quickly catch up to proprietary models, meaning raw generative capabilities will become table stakes.
- [00:04](https://www.youtube.com/watch?v=6jYv47iAQis&t=4s) Enduring commercial value lies in software harnesses that enable models to reconcile disparate, contradictory data feeds and synthesize actionable business conclusions.
- [02:02](https://www.youtube.com/watch?v=6jYv47iAQis&t=122s) Modern marketing operators suffer from cognitive overload rather than a lack of data, struggling to isolate cause and effect across competing channels, CRM conversions, and AI search visibility.
- [02:02](https://www.youtube.com/watch?v=6jYv47iAQis&t=122s) Current generative AI tools are largely underutilized as low-leverage copy generators rather than strategic filters that isolate critical commercial correlations.
- [02:02](https://www.youtube.com/watch?v=6jYv47iAQis&t=122s) Generic, off-the-shelf agents fail at commercial strategy because they lack internal context like unit margins, target ICPs, and specific product advantages.
- [04:48](https://www.youtube.com/watch?v=6jYv47iAQis&t=288s) Enterprises will build 'living commercial models' that combine internal operations (CRM, ad accounts, margins) with external signals (AI search ranking, competitor moves) into persistent brand memory.
- [04:48](https://www.youtube.com/watch?v=6jYv47iAQis&t=288s) Zero-shot full autonomy across marketing operations is impractical due to soaring API costs, reasoning failures, and substantial brand safety risks.
- [04:48](https://www.youtube.com/watch?v=6jYv47iAQis&t=288s) Effective agent systems require domain-specific guardrails and deliberate human-in-the-loop validation to safely self-improve over time.
- [07:04](https://www.youtube.com/watch?v=6jYv47iAQis&t=424s) Context-grounded agents evaluate competitor maneuvers strategically—recommending aggressive defense, waiting, or outright disregard based on margins and ICP alignment rather than knee-jerk price cuts.
- [07:04](https://www.youtube.com/watch?v=6jYv47iAQis&t=424s) While current AI focuses narrowly on promotional copy, the larger economic opportunity spans the broader 'Five Ps' of marketing, including pricing elasticity, placement, and product strategy.
- [09:02](https://www.youtube.com/watch?v=6jYv47iAQis&t=542s) Autonomy should be calibrated by judgment risk: background market research can operate 24/7 autonomously, while tactical tweaks use one-click approvals and structural positioning requires expert oversight.
- [09:02](https://www.youtube.com/watch?v=6jYv47iAQis&t=542s) 'Decision lineage' creates a closed feedback loop by logging contexts, recommendations, human modifications, and real-world outcomes to continuously refine brand memory.
## Detailed walkthrough
### [00:04](https://www.youtube.com/watch?v=6jYv47iAQis&t=4s) From Foundation Models at DeepMind to Commoditization: A Spicy Take
Praneet Dutta opens his session by noting the shift in scenery for both AI4 and himself: while previous years were held at the MGM Grand, the conference has now moved to the Venetian, and he is stepping on stage as an early-stage founder rather than a researcher representing a major tech laboratory. Reflecting on his background as a research tech lead at Google DeepMind, he mentions his work across foundation model post-training teams—touching models such as Imagen 3, Imagen 4, Gemini Nano, and Bananas. While acknowledging the remarkable strides these teams continue to make in the raw capabilities and efficiency of next-generation models, he offers what he terms his "spicy take": foundation models themselves are rapidly becoming commoditized and will no longer serve as an enduring "special sauce."
With open-source architectures trailing closed-source leaders by mere months, Dutta argues that raw reasoning and generative capacity will soon be table stakes. While the broader industry remains fixated on moving from conversational chatbots to tool-calling actions, he and his co-founders at Pomo identify a far larger opportunity. The frontier lies not in simply generating text or images, but in surrounding these models with robust harnesses capable of reasoning across dozens of disparate data feeds, reconciling contradictory signals, and synthesizing high-confidence conclusions that businesses can actually execute.
- Dutta previously spoke at AI4 as a big tech foundation model researcher before founding Pomo (usepomo.ai).
- Experience spans post-training on Google DeepMind projects, including Imagen 3, Imagen 4, and Gemini Nano.
- Believes model capabilities will become a baseline utility, shifting value to orchestration harnesses that handle multi-source reasoning.
---
[02:02](https://www.youtube.com/watch?v=6jYv47iAQis&t=122s) Navigating the Signal Deluge in Modern Marketing Ops
Turning to the day-to-day realities of marketing operations, Dutta highlights a pain point shared by early-stage startups and enterprise leadership alike: driving quantifiable return on marketing spend while being overwhelmed by information overload. Modern growth operators are inundated by hundreds of shifting signals every single day. Teams must track rival campaigns, monitor cross-channel social commentary, assess point-of-sale and CRM conversions, and now reckon with an entirely new inbound vector—brand visibility within generative search engines like ChatGPT, Claude, and Perplexity.
```mermaid
flowchart TD
accTitle: Multi-Source Signal Integration and Strategic Synthesis
subgraph S["Signal Ingestion"]
S1["Competitor Moves & Promos"]
S2["Cross-Channel Social Sentiment"]
S3["Generative AI Visibility: ChatGPT, Perplexity"]
S4["CRM & Point-of-Sale Conversion"]
end
subgraph E["Reasoning & Synthesis Engine"]
R1["Multi-Source Correlation Engine"]
R2["Deep Company Context: Margins, ICP, SKUs"]
end
subgraph A["Downstream Action"]
A1["Strategic Executive Decisions"]
A2["Targeted Campaigns across 5 Ps"]
end
The fundamental bottleneck is rarely a lack of data; it is the cognitive load of translation. Marketers struggle to cleanly isolate cause and effect—such as determining whether an uptick in customer inquiries was triggered by a promotional tweak, an algorithm update, or organic commentary. While generative AI is frequently relegated to low-leverage execution tasks like churning out ad variants and marketing copy, its true economic value emerges when it helps operators filter the noise, isolate key business correlations, and pinpoint the handful of strategic decisions worthy of executive attention.
Dutta illustrates this gap through the shortcomings of naive off-the-shelf agents. An operator might task an unanchored ChatGPT or Claude agent with monitoring brand sentiment. While it may return high-level observations—such as noting that category interest is trending upward on TikTok or Instagram—it lacks the situational awareness required for sound decision-making. Without visibility into unit margins, target Ideal Customer Profiles (ICPs), or specific formulation advantages (such as the distinct ingredients of a consumer packaged goods beverage), generic media monitoring yields surface-level observations rather than defensible commercial strategy.
---
[04:48](https://www.youtube.com/watch?v=6jYv47iAQis&t=288s) Architecting a Living Commercial Model and Agent Harnesses
To bridge this operational divide, Dutta posits that modern enterprises will inevitably construct what he calls a "living commercial model" of their business stack. Because autonomous agents excel at state tracking and continuous memory retention, they can maintain a synchronized operational picture that fuses internal operations with external market dynamics.
Internally, this commercial model ingests enterprise state: active advertising accounts, transaction ledgers, inventory velocity, and CRM records. Externally, it monitors customer sentiment, rival messaging, and generative AI search indexing, which can shift dramatically from one week to the next. By anchoring these disparate inputs within a persistent brand memory, companies can systematically audit historical outcomes, determine what interventions worked, and build cumulative institutional intelligence.
```mermaid
flowchart LR
accTitle: Living Commercial Model and Decision Architecture
subgraph Context["Dual-State Inputs"]
direction TB
Int["Internal State<br/>(CRM, Ads, Margins, POS)"]
Ext["External Signals<br/>(AI Search, Rivals, Sentiment)"]
end
subgraph Core["Pomo Architecture"]
direction TB
LCM["Living Commercial Model<br/>& Brand Memory"]
Harness["Agent Harness & Guardrails"]
end
subgraph Exec["Execution Loop"]
direction TB
HITL["Expert-in-the-Loop Review"]
Action["Market Action & Feedback"]
end
From an engineering perspective, Dutta warns against the naive assumption that frontier models can simply "zero shot" end-to-end commercial operations. While closed-model providers often present zero-shot autonomy as an out-of-the-box reality, teams attempting this approach quickly find themselves faced with staggering API costs and unreliable reasoning loops. Even as reasoning models absorb more cognitive tasks, an unaddressed "white space" of brand safety and strategic error remains. Building a viable agent system demands rigorous software harnesses and domain-specific guardrails, pairing automated execution with deliberate expert-in-the-loop validation so the system safely self-improves over time.
---
[07:04](https://www.youtube.com/watch?v=6jYv47iAQis&t=424s) The Five Ps: Why Strategic Nuance Beats Knee-Jerk Reactions
To ground the distinction between superficial automation and deep contextual reasoning, Dutta walks through a concrete competitive scenario: an enterprise resource planning (ERP) competitor suddenly announces a 20% promotional discount. A reactive, uncalibrated marketing instinct might jump to match or undercut them immediately: "Let's just make it 25%." Dutta cautions that ungrounded price wars destroy unit economics without solving underlying strategic hurdles.
An agent grounded in a company's full commercial context approaches the problem through distinct operational paths based on target customer profiles and profit margins:
- **Aggressive Direct Defense:** If the rival's SKU targets the exact same ICP and product tier, and the company maintains healthy margins while chasing market share, the system validates launching an aggressive, limited-time promotion reinforced with product-value case studies.
- **Wait and Watch:** If the rival's promotion targets an adjacent white space or non-core feature set, the prudent operational choice is monitoring sentiment in community threads and customer forums to assess whether prospective buyers actually care before burning margin.
- **Strategic Disregard:** If the rival's discount applies to an irrelevant tier or uncompetitive product line, the system recommends ignoring the maneuver entirely, preserving marketing budget and brand equity.
Polling the audience to see how many attendees come from a classic marketing background, Dutta points out that while the marketing discipline rests upon the "Five Ps" (Product, Price, Place, Promotion, and People), current generative AI tools overwhelmingly target only promotional copy and ad creation. The broader opportunity is deploying AI agents to inform the foundational levers of commercial strategy: analyzing pricing elasticity, evaluating distribution placement, and guiding product development.
---
[09:02](https://www.youtube.com/watch?v=6jYv47iAQis&t=542s) Calibrating Autonomy and the Future of Decision Lineage
Addressing the central question of governance—"how much autonomy is too much?"—Dutta outlines a framework that separates autonomous research from real-world execution based on judgment risk. Information retrieval, market scanning, and competitive research should operate continuously in the background without human supervision. Operators should comfortably task agents to run exhaustive market analyses overnight, waking up to distilled executive summaries.
```mermaid
flowchart TD
accTitle: Autonomy Risk Spectrum
subgraph LowRisk["Autonomous / Low Risk"]
R["24/7 Market Research & Signal Monitoring"]
O["One-Click Approval Actions:<br/>• Budget shifts across ad channels<br/>• Creative copy adjustments<br/>• Keyword defense maneuvers"]
end
subgraph HighRisk["Expert-in-the-Loop / High Risk"]
P["Fundamental Value Prop Repositioning"]
C["Core Pricing Strategy Adjustments"]
G["Enterprise Go-To-Market Shifts"]
end
R --> O
O -->|Escalation Required| HighRisk
```
Operational interventions, by contrast, must be gated according to their risk profile:
- **Low-Stakes Actions (One-Click Approvals):** Adjusting tactical budget allocations between Instagram and LinkedIn, rotating ad copy that counterbalances competitor keyword bidding, or deploying promotional variants require minimal friction and can be executed via simple one-click approvals.
- **High-Stakes Decisions (Expert-in-the-Loop):** Structural interventions—such as fundamentally altering the core positioning or messaging of an enterprise product—carry high brand and commercial liability. Leaving these unconstrained risks waking up to community backlash on platforms like X and LinkedIn. Until multi-agent coordination matures over the next 5 to 10 years, high-judgment decisions must be co-piloted by human experts.
Dutta concludes by framing the future of autonomous growth around "decision lineage." Borrowing the closed-loop feedback principles of reinforcement learning, decision lineage tracks the entire trajectory of an agentic workflow: reading internal context, synthesizing market signals, recommending an intervention, logging executive modifications, observing the real-world outcome, and updating core brand memory. This closed feedback loop allows B2B growth engines to continuously learn from historical performance, transforming marketing from disconnected creative bursts into a self-improving operational science. He invites attendees to explore Pomo through a conference-specific trial before opening the floor to audience questions.
## Notable quotes
> [00:04](https://www.youtube.com/watch?v=6jYv47iAQis&t=4s) foundation models themselves are rapidly becoming commoditized and will no longer serve as an enduring 'special sauce.' — Praneet Dutta
> [00:04](https://www.youtube.com/watch?v=6jYv47iAQis&t=4s) The frontier lies not in simply generating text or images, but in surrounding these models with robust harnesses capable of reasoning across dozens of disparate data feeds, reconciling contradictory signals, and synthesizing high-confidence conclusions that businesses can actually execute. — Praneet Dutta
> [02:02](https://www.youtube.com/watch?v=6jYv47iAQis&t=122s) The fundamental bottleneck is rarely a lack of data; it is the cognitive load of translation. — Praneet Dutta
> [02:02](https://www.youtube.com/watch?v=6jYv47iAQis&t=122s) While generative AI is frequently relegated to low-leverage execution tasks like churning out ad variants and marketing copy, its true economic value emerges when it helps operators filter the noise, isolate key business correlations, and pinpoint the handful of strategic decisions worthy of executive attention. — Praneet Dutta
> [04:48](https://www.youtube.com/watch?v=6jYv47iAQis&t=288s) living commercial model — Praneet Dutta
> [04:48](https://www.youtube.com/watch?v=6jYv47iAQis&t=288s) While closed-model providers often present zero-shot autonomy as an out-of-the-box reality, teams attempting this approach quickly find themselves faced with staggering API costs and unreliable reasoning loops. — Praneet Dutta
> [07:04](https://www.youtube.com/watch?v=6jYv47iAQis&t=424s) Let's just make it 25%. — Praneet Dutta
> [09:02](https://www.youtube.com/watch?v=6jYv47iAQis&t=542s) how much autonomy is too much? — Praneet Dutta
## Video description
Praneet Dutta - Founder, usepomo.ai
Paige Liao - Software Engineer, MachFlow AI (POMO)
Marketing AI has mostly stopped at copilots: useful for isolated tasks, but weak at turning insight into action. In this talk, Praneet Dutta, founder of Pomo and former DeepMind GenAI research tech lead, will show why the next leap is autonomous growth systems: AI agents that can research markets, generate strategy, create campaigns, launch experiments, and learn from real-world feedback. Drawing from his work on foundation models at DeepMind and now building Pomo, he will walk through the architecture behind an agentic marketing system, including creative generation, decision intelligence, feedback loops, and human-in-the-loop controls. Attendees will learn what it actually takes to move from AI demos to production systems that drive measurable growth. This talk is for founders, operators, and AI practitioners who want a practical view of where autonomous marketing is going and how to build it safely.
前DeepMind技术主管：基础模型走向大宗化，Agent系统成取胜关键
基础模型正迅速大宗化为底层公用设施，企业价值正转向融合内部业务数据与外部市场信号的上下文Agent控制架构。
基于在Google DeepMind的基础模型研究经验，Praneet Dutta指出，原生模型能力正趋于大宗化，具备防御性的商业价值正转移至能够跨复杂、多源企业数据进行推理的Agent架构。企业不应将生成式AI仅用于撰写浅层宣传文案或被动降价，而必须构建一个“动态商业模型”，将单位利润率和目标客户画像（ICP）等内部运营状态，与竞对动态、生成式搜索曝光度等外部信号相结合。Dutta还概述了一个权责明确的治理框架：全天候后台调研与低风险战术调整可自主运行，而高风险的品牌和定价决策仍由人类专家把关。最后，他提出了作为运营反馈闭环的“决策谱系”（decision lineage），通过记录每一次干预并从中学习，将市场营销转变为一门可衡量且具备复利效应的科学。
前 Google DeepMind 研究技术负责人 Praneet Dutta 提出，随着开源模型迅速赶超闭源模型，基础模型本身正在快速商品化；未来的持久商业价值在于能协调多源复杂数据并提炼行动结论的软件架构（Harness）。 (0:04)
现代营销人员普遍面临信号过载难题，难以在竞对动作、社媒舆情与生成式 AI 搜索可见性中厘清因果，而通用开箱即用 Agent 因缺乏企业利润率和目标画像等上下文，无法形成有效商业策略。 (2:02)
企业将需要构建融合内部运营与外部市场动态的“实时商业模型”（Living Commercial Model），通过持久的品牌记忆系统性复盘历史效果，克服纯 Zero-shot 全自主带来的高昂 API 成本与品牌安全风险。 (4:48)
自主增长系统需按风险等级校准控制权：全天候市场调研与预算微调可实现自动化或一键审批，而涉及核心定价与产品定位等高风险决策则必须保留专家在环（Human-in-the-Loop）机制。 (9:02)
Reading summary
Praneet Dutta, former Google DeepMind research lead and founder of Pomo, argues that foundation models are rapidly commoditizing into table-stakes utilities, shifting commercial defensibility to orchestration harnesses. In his view, sustainable enterprise value requires integrating external market signals with an internal operational model of unit margins and customer profiles rather than relying on ungrounded copy generation.
Foundation Model Commoditization and the Orchestration Frontier (0:04)
Reflecting on his research experience across foundation model post-training teams at Google DeepMind—including Imagen 3, Imagen 4, and Gemini Nano—Dutta contends that raw model capabilities are commoditizing, with open-source alternatives trailing proprietary leaders by only months. Consequently, he argues that standalone generative capabilities and basic tool-calling will no longer serve as an enduring competitive advantage.
Instead, Dutta asserts that the technical frontier lies in developing surrounding software harnesses. These systems must ingest dozens of disparate and conflicting data feeds, reconcile contradictory signals, and synthesize high-confidence strategic conclusions that enterprise operators can safely execute.
Bridging the Signal Deluge with Living Commercial Models (2:02)
Marketing operators face persistent information overload across hundreds of shifting daily signals, including competitor campaigns, cross-channel social commentary, CRM and point-of-sale conversion data, and visibility within generative search engines such as ChatGPT and Perplexity. Dutta points out that generic, off-the-shelf agents fail in this environment because they lack awareness of internal unit margins, formulation specifics, and target Ideal Customer Profiles (ICPs), resulting in superficial observations rather than defensible strategy.
To bridge this gap, Dutta posits that enterprises must construct a 'living commercial model' that fuses internal operations (transaction ledgers, active ad accounts, inventory velocity) with external market monitoring. He warns against assuming frontier models can manage these operations zero-shot, noting that unconstrained zero-shot workflows create unsustainable API expenses, erratic reasoning loops, and brand safety liabilities.
Contextual Decision-Making Across the Five Ps (7:04)
Dutta critiques current generative AI applications for concentrating almost exclusively on promotional ad copy, neglecting the broader strategic foundation of the 'Five Ps': Product, Price, Place, Promotion, and People. He argues that AI agents should instead be tasked with analyzing core business levers such as pricing elasticity, distribution channels, and product development.
To illustrate contextual reasoning over knee-jerk reactions, Dutta examines an enterprise resource planning competitor introducing a 20% discount. While an uncalibrated reaction might trigger a margin-destroying price cut to 25%, an agent grounded in company economics evaluates strategic options: mounting an aggressive direct defense if margins allow and the target ICP aligns, taking a wait-and-watch approach to assess buyer response in adjacent segments, or ignoring the move entirely if it affects an irrelevant product tier.
Risk-Gated Autonomy and Closed-Loop Decision Lineage (9:02)
Dutta proposes governing growth agents according to judgment risk. Continuous data ingestion, market monitoring, and competitive research run fully autonomously in the background. Low-stakes operational adjustments—such as reallocating channel ad spend, adjusting creative copy, or executing keyword defense—are deployed through one-click human approvals. High-stakes structural shifts, such as altering core pricing or foundational product positioning, require expert-in-the-loop oversight to prevent brand backlash, with Dutta forecasting that reliable autonomous multi-agent coordination remains 5 to 10 years away.
To drive continuous system improvement, Dutta introduces 'decision lineage,' an architecture derived from reinforcement learning principles. By systematically tracking the initial market context, agent recommendations, human executive modifications, and downstream real-world outcomes, the system updates its core brand memory, turning growth operations into an auditable, self-improving science.
Dutta concludes that fully autonomous commercial execution remains premature, necessitating human-in-the-loop governance for strategic and brand-critical decisions. The open challenge for enterprises is transitioning growth workflows from intuitive, uncoordinated marketing tasks into closed-loop systems anchored by persistent commercial memory and auditable decision lineage.
[0:04] Praneet Dutta: foundation models themselves are rapidly becoming commoditized and will no longer serve as an enduring 'special sauce.'
[0:04] Praneet Dutta: The frontier lies not in simply generating text or images, but in surrounding these models with robust harnesses capable of reasoning across dozens of disparate data feeds, reconciling contradictory signals, and synthesizing high-confidence conclusions that businesses can actually execute.
[2:02] Praneet Dutta: The fundamental bottleneck is rarely a lack of data; it is the cognitive load of translation.
[2:02] Praneet Dutta: While generative AI is frequently relegated to low-leverage execution tasks like churning out ad variants and marketing copy, its true economic value emerges when it helps operators filter the noise, isolate key business correlations, and pinpoint the handful of strategic decisions worthy of executive attention.
[4:48] Praneet Dutta: living commercial model
[4:48] Praneet Dutta: While closed-model providers often present zero-shot autonomy as an out-of-the-box reality, teams attempting this approach quickly find themselves faced with staggering API costs and unreliable reasoning loops.
[7:04] Praneet Dutta: Let's just make it 25%.
[9:02] Praneet Dutta: how much autonomy is too much?
[9:02] Praneet Dutta: decision lineage
Original transcript
[0:04] Thank you everyone for joining in. I know
it's been a busy morning for all of you. So,
[0:07] appreciate you um staying with
me here for my talk, right?
[0:11] Um you know, I've enjoyed AI4 over the
years. Um I think this is a change for
[0:15] both AI4 and us. I think last year they were at
the MGM Grand and now they're at the Venetian.
[0:20] And the last two years where I spoke, um I was in
a big tech lab working on the foundation models.
[0:25] Uh but now I'm speaking as a founder, which is um
also a very new experience for me as well, right?
[0:32] So, it's very just a bit of a brief background
about me. I spent most of my time working on the
[0:36] foundation models, on post-training teams,
such as Imagen 3, 4, Gemini Nano, Bananas,
[0:41] and I think the teams have done amazing
work on the new generation of models.
[0:45] Um and that's what our founding team also has
experience in the space. There's a lot of talk
[0:51] in terms of like how do we improve the models?
And I think that arms race will keep improving,
[0:54] right? And you know, we'll see better and
better intelligence at like lower costs.
[0:59] Uh but, you know, my hot take is I don't think
the models are going to be like the special
[1:03] sauce anymore. We're seeing a certain amount
Oh, can you hear me? Okay, perfect. Thank you
[1:08] for correcting me. So, I'll go back to my spicy
take. Um I think the models will be commoditized,
[1:14] right? We're already seeing that. Close-sourced
is like just a few months behind open So,
[1:18] yeah, sorry. Open-source is just a few months
behind close-source, right? Um so, we'll start
[1:23] seeing like more emergent abilities in models
as we keep going forward in the future, right?
[1:28] And a lot of people keep talking about the amazing
generation capabilities. Models are great at,
[1:32] you know, like reasoning and like generating new
concepts. We are starting to see the shift from
[1:36] chatbots to actions, right? Um but the reason
the what I saw and some of my co-founders saw
[1:42] was that there's a much wider opportunity with
the way in which intelligence is growing, right?
[1:48] Um these models and the harnesses around them,
they are great at reasoning through multiple
[1:52] sources and then being able to infer a conclusion
it shares with you, right? Obviously, there's
[1:56] the elements of hallucination stuff, you know,
but these are getting better day by day, right?
[2:02] So, what we see is like moving beyond just the
models towards a layer where they can actually
[2:08] reason through multiple different sources and
then help share a conclusion, right? That's a
[2:12] much more broader opportunity, right? So, in our
case, you know, we are building our AI agents for
[2:19] marketing ops for you know, all the way from like
tiny startups till enterprises. You know, a lot of
[2:25] you have the same problem, like better ROI on your
marketing spend, right? Um for those of you who
[2:31] are in charge of your marketing or sales or your
leadership team, some of this will make sense,
[2:35] right? You're getting inundated in like tens of
signals or like hundreds of signals every day.
[2:41] What your competitors are doing, what are
people talking about you on different channels,
[2:46] how does your AI visibility? Now, we have another
channel for inbound, you know, what people are
[2:50] saying on ChatGPT, Claude's, perplexities of
the world. Um how is your demand shifting if
[2:55] you connect your point of sales or your you
know, your CRMs? How how is that translating?
[3:02] If you have access to all of these different
signals, that's great. But then, it's this
[3:06] additional layer of translation where you have
to reason through all of them, what you know,
[3:10] what's the correlation? Did you launch a new
promotion? Did cause more inbound, right?
[3:14] That's something we have seen like some of the
best marketers struggle with, right? Obviously,
[3:17] you can use like AI for like making better ads,
generating writing creatives, right? Um that's
[3:22] one aspect to it. But being able to actually make
sense of what's happening in the market and then
[3:29] um kind of like summarize what are the
most important decisions for you cuz our
[3:32] time as marketers, business owners, um folks
working on building amazing stuff is valuable.
[3:38] That's the real ROI that we see, right? And then
downstream, you can execute whether it's like,
[3:41] you know, the five P's in marketing, you
can execute in any of those five P's.
[3:47] Um so, we take a specific case, right? Let's say
you kind of like link up your ChatGPT or Claude
[3:51] agent and act ask it to act like a media monitor,
right? If you're like running a startup or a
[3:55] growth team and you ask it, "Okay, keep telling me
what competitors are doing and what's happening."
[4:00] You know, it might be able to give you insights
like, "Okay, interest in your category is going
[4:04] up. People are maybe posting more on Instagram,
TikTok, all channels you care about, right?" But,
[4:11] without the full context of your business
in terms of what's happening, right?
[4:16] Does it fully understand like who you're selling
to, what the ICP is, what are the margins you make
[4:20] on this product, right? Are people interested in
this certain offering? Like, you know, if you sell
[4:26] CPG drink, right? Do they care about
the ingredients that you have, right?
[4:29] Only once you have full understanding of this,
can you get something that moves from like
[4:34] a high-level media monitor towards a decision
that you as a business owner can act on, right?
[4:39] So, in general, the more specific it is, it might
be wrong, but at least it moves in one direction
[4:43] of reasoning. That's what's more useful for
business owners that we've been talking to.
[4:48] So, the way we kind of see this evolving,
right? We're seeing, you know, some of the
[4:51] best, you know, think of the best companies in
marketing, they already have done this, right?
[4:55] But, if that was not, you know, quite obvious
and a bit more subtle, is every company will
[5:00] start building a living kind of commercial
model of what their stack looks like, right?
[5:05] As I spoke about, agents are great at reasoning
and holding state and holding memory, right?
[5:09] So, what this means specifically in
the marketing domain is we have your
[5:13] like specific company state, your commerce, your
advertising accounts, you know, your CRM, right?
[5:19] And then we have everything happening externally,
you know, what what shoppers are talking about,
[5:23] what competitors are doing, right? How is your AI
visibility? What are new blog posts and articles
[5:28] that, you know, influencing that, right? And this
changes every day. You know, you might ask your
[5:34] Perplexity agent a question today and the answer
will be very different, you know, maybe in a week
[5:37] from now, right? So, it's very important for
companies, what we're seeing, to have a kind
[5:42] of like living commercial model in terms of what
does their um, brand memory look like, how that's
[5:48] updating, and what can it learn from what happened
in the past, right? And that's sort of the problem
[5:52] they're tackling here at Pomo, building up this
kind of like living commercial model, so companies
[5:57] can learn, can always like look up what worked,
what didn't work, then based on that take action.
[6:04] Um, from a technical perspective, speaking about
how this works would be a much longer um, talk.
[6:09] But as you you know, some of you that work
with like the some of the largest closed model
[6:12] providers, the usual answer would sometimes be
that yeah, you can maybe just zero shot it and the
[6:17] model's smart enough to do it. Sure, but you know,
your API bills will speak differently, right?
[6:23] Um, so and that's one factor along with the
fact that you know, would you I I think models
[6:28] are getting a lot smarter, right? But even
today with like some of the best harnesses,
[6:33] they're eating up the judgment stack in terms
of what they're able to do independently,
[6:37] but there's still this amount of white space where
you won't trust it. You we all heard about like
[6:41] what happened with the open clause of the world
back in you know, like six or seven months ago,
[6:44] right? So, um, there's amount of work that goes
into like setting up a harness, setting up the
[6:50] adequate guardrails, so while these models can
reason and get stuff done, you want to make sure
[6:54] that is a certain amount of expert in the loop, so
you can guide them in the right direction as they
[6:59] keep self-improving that regard. So, if we kind of
like take a specific example, right? Um, let's say
[7:05] you set up your media monitor and you see oh, your
closest competitor has like a 20% discount on the
[7:10] ERP uh, plan, right? Um, uh, a simple actual
be you know what, we need to respond to that,
[7:18] right? That's actual customer eyeballs.
Let's just you know, we make it 25%, right?
[7:23] That's obviously not how you know, you would
want things to work, right? If that happens. Um,
[7:28] once you have a system that can more, you know,
like deeply understand, you know, your company,
[7:32] your products, your SKUs, the margins you
make on them, this is more of a deeper
[7:37] discussion that you need to reason through
and then decide how to act on on right?
[7:41] So, let's say you know, your competitor is you
know has an SKU that's targeting the same ICP
[7:46] you're targeting, you know, same offering,
it completely competes with you. You're
[7:50] pretty profitable, but you're looking at chasing
growth. All that we understand for your business.
[7:54] Yeah, sure. You know, launch a limited promotion,
you know, make it compelling, advertise, you know,
[7:59] what your value prop is, case studies, it'll help
you, right? Um but what if it's not as relevant?
[8:05] There's a certain white space in the market
that you know they are not touching or slightly
[8:08] different, but they may or may not go there. Sure.
Then it kind of makes sense to kind of like wait
[8:13] and watch, see what people are talking about
on the comment threads, you know, what are the
[8:17] main kind of discussion points, and then if this
becomes something relevant, you might go for it.
[8:23] And then if it's completely relevant,
why even waste your kind of like margins
[8:26] advertising dollars on this, right? You can
just choose to ignore it, see how that happens,
[8:29] see what happens from this, and then choose to
kind of act on this in the future. So, um how many
[8:35] of you are actually from a marketing background
in the room? I should have started with that.
[8:40] Okay. I I see one. So, you know there's
five P's of marketing, right? Most most of
[8:44] generative AI targets is like one or two, which
is like better AI ads and promotions, right?
[8:48] But everything else, the product, the placement,
the price, these are like more fundamental
[8:52] discussions that we think that generative AI can
help teams massively grow and learn from, right?
[8:57] And that's kind of like a discussion point where
this comes into. Um something that might be, you
[9:03] know, in your minds as we talk about this is like,
"Okay, but how much autonomy is too much, right?"
[9:08] Um the way we see it is for everything on the
research side, stuff that you know you can
[9:14] ask your GPT agent at night, "Do this
research for me. I'm I'm sleeping. I'll
[9:17] check it in the morning, right?" All of that
is something that should be on by default 24/7,
[9:22] right? Agents should always be looking, um seeing
what's happening, and then summarizing, right?
[9:27] Um the action aspect of it, being able to actually
look in the market then have an inference, right?
[9:33] Being able to execute with guardrails,
that's where things get more exciting.
[9:37] The way we see it is for low stake task like in
the case of marketing which is whether it's like
[9:42] creating a new promotion, changing budget
from let's say Instagram to LinkedIn ads,
[9:47] um, you know, you see one copy creative is
working well, your customers kind of attacking
[9:52] your keywords. Adapting those and giving you
kind of um, counter actions. Those are stuff
[9:57] that can be one click approval, right? But
what if we ask like a deeper question? Um,
[10:03] like, okay, we think that your positioning should
be fundamentally changed for one of your products,
[10:07] right? That's something which you don't want your
agent to kind of go out in the market and then
[10:11] just like run something and then, you know, you
wake up to angry people on, you know, LinkedIn, X,
[10:16] and all of the channels saying that what's going
on here, right? So, we kind of like go as we go
[10:20] up the judgment stack, we want this to be more
of an expert and agent in the loop, right? Um,
[10:25] we do see things changing in the future. Maybe in
5 or 10 years from now every company has their own
[10:30] custom agent that works with the marketing agent,
the sales agent, accounting agent. But for now,
[10:34] I think given the amount of um, um, long horizon
success we've seen so far, we do fundamentally see
[10:42] that, you know, for a task higher in judgment
risk, uh, we do need experts in the loop.
[10:49] So, I'll kind of like end up by saying that, you
know, we see that companies and we're seeing some
[10:53] of the larger ones still adopting this already.
They are building something called a decision
[10:57] system of decision lineage. Agents that, you
know, understand your company's context, takes
[11:02] take actions on that. You see how they are
perceived and then based on that it's able
[11:07] to actually learn and then keep improving. That's
what's going to be the future. For those of you
[11:11] that, you know, have a background in reinforcement
learning, this closed loop would be familiar.
[11:15] Uh, unfortunately for most B2B companies, that's
starting to get more and more possible now. Right?
[11:22] Um, yeah, so, you know, I mean, we are
super excited for the future, you know,
[11:26] I think it's it's a it's it's a very, very
interesting way in which the intelligence,
[11:30] the cost of intelligence is going up,
right? We are excited to start promo. We
[11:34] have seen some great results with, you know,
early customers on the platform, right? So,
[11:39] would love to answer any questions you
have about our vision, what we are doing,
[11:43] and we're also expand Sorry, extending an AI photo
offer where you can get like a free trial and
[11:48] custom discount code. So, you can scan the QR
code if you want to check this out. Okay. Well,
[11:53] thank you so much and I'm here to answer any
questions that all of you have. Thank you.
