Kearney's Paredes Warns Base AI Models Offer Zero Corporate Moat
Ai4
Because foundation models are commoditized utilities accessible to every competitor, businesses must anchor their competitive advantage in proprietary data and redesigned operational workflows rather than seat adoption metrics or autonomous agent overkill.
AI-assisted summary of Where Do Most Organizations and People Go Wrong With AI, What Are the Most Common Mistakes, and — original source. Check the original for full context.
Source published:
Detailed notes
## Background Published on YouTube by Ai4 on August 19, 2026, this presentation features Miguel Paredes, Partner and Fellow at Kearney and AI Fund. The session explores where most organizations and professionals go wrong when implementing artificial intelligence, highlighting common pitfalls and actionable strategies for building durable competitive advantage.
## TL;DR Miguel Paredes explains why treating AI as an isolated point solution or relying on commoditized foundation models as a proprietary moat leads enterprise strategies astray. Because base intelligence is accessible to everyone, sustainable differentiation shifts toward proprietary internal data, tacit domain knowledge, and fundamental workflow redesign rather than chasing superficial adoption metrics. He warns against agent overkill and illustrates how accelerating upstream throughput can overwhelm human decision-making bottlenecks without proper governance. To succeed, organizations must evaluate tasks across a spectrum of agentification and allow new AI capabilities to actively challenge and reshape overall corporate strategy.
## Atmosphere & style The presentation is pragmatic, intellectually grounded, and disarmingly self-deprecating, balancing rigorous management consulting frameworks with candid anecdotes about MIT, sommelier training, and failed stand-up comedy routines. Paredes speaks with reflective honesty about existential career identity crises and relatable modern frustrations, such as getting trapped in automated customer service loops. It feels like an engaging, hype-free masterclass that cuts through technical buzzwords to examine the messy, human realities of organizational transformation.
## Key takeaways - [06:02](https://www.youtube.com/watch?v=MZRAqkLEgmA&t=362s) Artificial intelligence is a General Purpose Technology (GPT) comparable to electricity or the steam engine, meaning it requires systemic workflow reconfiguration rather than treatment as an isolated point solution. - [08:32](https://www.youtube.com/watch?v=MZRAqkLEgmA&t=512s) Enterprise transformation follows a phased, deliberate metamorphosis rather than an overnight 'big bang,' as institutional inertia and change management cause organizational adoption to lag behind consumer awareness. - [09:42](https://www.youtube.com/watch?v=MZRAqkLEgmA&t=582s) Because frontier foundation models are accessible off the shelf to anyone via APIs, the base model functions as a commoditized utility and cannot serve as a defensible competitive moat. - [11:17](https://www.youtube.com/watch?v=MZRAqkLEgmA&t=677s) Durable competitive differentiation resides outside the model in proprietary enterprise data, tacit domain knowledge, and the ability to provide rich operational context. - [13:10](https://www.youtube.com/watch?v=MZRAqkLEgmA&t=790s) Measuring AI success through software adoption metrics such as seat count and login frequency is misleading, as high usage often masks inefficient or counterproductive practices. - [14:14](https://www.youtube.com/watch?v=MZRAqkLEgmA&t=854s) Deploying autonomous agents over broken or legacy processes merely accelerates operational failure, often creating expensive overkill where deterministic code or human workers remain superior. - [16:46](https://www.youtube.com/watch?v=MZRAqkLEgmA&t=1006s) Human personnel serve as a critical competitive advantage and brand safeguard, particularly in high-friction customer journeys where automated deflection loops erode client loyalty. - [19:32](https://www.youtube.com/watch?v=MZRAqkLEgmA&t=1172s) Accelerating task throughput with AI exposes downstream human bottlenecks, causing operational gridlock unless review workflows are restructured to accommodate human cognitive latency. - [21:05](https://www.youtube.com/watch?v=MZRAqkLEgmA&t=1265s) Failing to account for token economics leads to runaway inference costs, making architectural right-sizing and the use of open-weight or previous-generation models essential for high-volume tasks. - [22:07](https://www.youtube.com/watch?v=MZRAqkLEgmA&t=1327s) Model selection should be evaluated across task-specific multidimensional performance distributions rather than generic leaderboard benchmarks, backed by continuous monitoring against provider drift. - [23:28](https://www.youtube.com/watch?v=MZRAqkLEgmA&t=1408s) Organizations must adopt 'futures thinking' workshops to anticipate foundational shifts in consumer expectations and industry dynamics rather than merely tacking AI pilots onto existing models. - [27:25](https://www.youtube.com/watch?v=MZRAqkLEgmA&t=1645s) Workflows should be evaluated across a nuanced agentic spectrum—from human-only and AI-assisted to AI-augmented and fully autonomous—matching the level of autonomy to the specific job to be done. - [30:43](https://www.youtube.com/watch?v=MZRAqkLEgmA&t=1843s) Rather than merely aligning AI initiatives to legacy business strategies, enterprise leaders must use newly enabled AI capabilities to actively challenge and reinvent corporate strategy itself.
## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=MZRAqkLEgmA&t=4s) From MIT to Grocery AI: Miguel Paredes's Journey
Miguel Paredes introduces his background by recounting his path through applied artificial intelligence before it became mainstream. After completing his PhD at the MIT AI Lab (CSAIL) in 2014—joking that back then he was simply a "huge nerd" while today AI practitioners are considered "pretty cool," and self-deprecatingly calling himself the "dumbest person there"—he stepped into industry as the Chief Analytics and Data Officer for a massive Latin American conglomerate spanning ten distinct industries. This role gave him early hands-on experience deploying machine learning, deep learning, and natural language processing across traditional sectors such as banking, hospitality, cement and concrete, insurance, healthcare, fishing, and mining. Working with two global industry leaders—the world's largest tin mine and the world's largest fishmeal producer—highlighted how legacy, non-tech-native enterprises could unlock immense economic value simply by layering modern predictive models over existing, underutilized operational data.
Born and raised in Peru as half-American and half-Peruvian, Paredes eventually returned to the United States to serve as Vice President of AI and Data Science at grocery retail giant Albertsons Safeway. There, his teams analyzed massive volumes of consumer purchase data to decode household shopping behavior. One of his personal favorite projects combined his technical work with his passion for wine: having studied to become a sommelier, he led the development of algorithms that mapped dietary patterns from grocery receipts (tracking purchases of bread, butter, meat, and fish) to infer socioeconomic profiles and taste preferences at both individual and household levels, generating personalized wine recommendations.
- **MIT AI Lab (2014):** Early PhD research in applied machine learning prior to the modern generative AI boom. - **Conglomerate Leadership:** Managed analytics across 10 legacy sectors, including the world's largest tin mining operation and top global fishmeal producer. - **Retail Data Science at Albertsons Safeway:** Built household taste-preference inference models connecting everyday grocery baskets to curated wine recommendations.
---
[02:45](https://www.youtube.com/watch?v=MZRAqkLEgmA&t=165s) Enterprise AI, Board Governance, and the Human Identity Crisis
Paredes outlines his current multi-faceted roles across management consulting, academic research, and venture creation. As a partner at Kearney, he works directly with enterprise leaders navigating organizational transformation. He observes that there are currently no universal playbooks for enterprise AI; virtually every executive struggles with the same foundational questions: "Where do I start?", "Where should I focus?", and "Do I buy or build?" In parallel, Paredes conducts research at Harvard Business School alongside faculty members to examine AI adoption, workforce upskilling, and how corporate boards can shift from rigid, periodic meeting cadences into continuous, "living boards" powered by AI insights. He also collaborates with Andrew Ng at AI Fund to incubate and scale early-stage AI startups.
Beyond structural enterprise challenges, Paredes addresses the profound psychological and professional identity crisis triggered by generative AI. Professionals across every discipline—marketers, accountants, software engineers, and even veteran AI researchers—are grappling with existential questions about their future relevance and differential human value. Paredes admits that despite holding an MIT AI PhD and spending fifteen years in the discipline, he frequently questions his own unique value, jokingly noting his backup career ideas in wine tasting and stand-up comedy (admitting he completely bombed his first open-mic set despite using AI to help craft the routine).
- **Current Roles:** Partner at Kearney, Research Fellow at Harvard Business School, and collaborator at Andrew Ng's AI Fund. - **Enterprise Dilemma:** Universal lack of proven playbooks, forcing organizations to navigate custom build-versus-buy trade-offs. - **HBS Governance Research:** Redesigning corporate governance from episodic oversight into an AI-augmented "living board." - **The Identity Crisis:** A widespread re-evaluation of human differentiation across knowledge work, technical programming, and executive leadership.
---
[06:02](https://www.youtube.com/watch?v=MZRAqkLEgmA&t=362s) Mistake 0: Treating AI as a Point Solution Instead of a General Purpose Technology
The fundamental error organizations make—labeled "Mistake Zero"—is treating AI merely as an isolated technology, software feature, or point solution rather than recognizing it as a General Purpose Technology (GPT). In economics, a GPT represents a foundational, society-altering innovation comparable to electricity, the steam engine, the personal computer, Wi-Fi, or mobile telephony. These technologies do not just optimize existing workflows; they permeate every facet of societal infrastructure, redefine existing business models, and create entirely new economic ecosystems that future generations end up taking for granted.
```mermaid flowchart TD accTitle: General Purpose Technology Evolution subgraph GPT_Hist["Historical General Purpose Technologies"] E["Electricity"] S["Steam Engine"] PC["Personal Computing & Wi-Fi"] end
subgraph AI_Impact["Artificial Intelligence as a GPT"] AI["General Purpose AI"] W["Workflow & Process Reconfiguration"] Ind["New Industry & Business Model Creation"] end
GPT_Hist -->|"Historical Pattern of Pervasive Diffusion"| AI_Impact AI --> W AI --> Ind ```
To illustrate how embedded AI is already becoming in day-to-day productivity, Paredes recounts a recent lunch meeting with Brice Challamel from OpenAI. Challamel demonstrated how he coordinated their entire lunch plan using ChatGPT: taking Paredes's incoming text message, generating venue suggestions with photos, drafting a finalized response, and scheduling the calendar entry without writing or editing a single line of text manually. Just as the discovery of electricity reshaped global manufacturing and urban living over decades, AI is poised for radical, systemic impact—though its diffusion will be non-uniform across industries, geographies, and demographics.
- **Economic Definition of a GPT:** A pervasive technology that permanently transforms how society works, relates, and creates economic value (e.g., electricity, steam power, computing). - **Practical Workflow Integration:** Frontier professionals already delegate end-to-end operational tasks—such as contextual schedule orchestration and communication drafting—directly to LLMs. - **Uneven Trajectory:** The diffusion of AI capabilities will not happen uniformly, at the same speed, or with equal intensity across all sectors.
---
[08:32](https://www.youtube.com/watch?v=MZRAqkLEgmA&t=512s) Mistake 1: Expecting an Overnight "Big Bang" Transformation
The second major misconception is assuming that enterprise AI transformation will occur as a sudden, sweeping "big bang." Drawing on digital transformation frameworks taught at the IMD business school in Lausanne, Switzerland, Paredes references the classic caterpillar-to-butterfly metaphor to illustrate that deep organizational metamorphosis is typically deliberate, gradual, and complex.
While certain software-native domains experience rapid, explosive adoption, historical analysis of prior general-purpose technologies demonstrates that widespread industrial deployment often takes years or decades. Even in an interconnected digital era where information disseminates instantly across the internet, institutional inertia, regulatory requirements, talent shortages, and change management mean that enterprise-scale adoption remains fragmented, with large cohorts of organizations and knowledge workers still barely scratching the surface of AI integration.
- **Gradual Metamorphosis:** True structural transformation follows a phased evolution rather than instantaneous organizational overhaul. - **Historical Diffusion Cycles:** Past GPTs required decades to re-architect industrial workflows, train labor pools, and adjust regulatory environments. - **Dissemination vs. Adoption:** Instant global information flow accelerates technological awareness, but practical enterprise-wide deployment lags significantly behind consumer experimentation.
---
[09:42](https://www.youtube.com/watch?v=MZRAqkLEgmA&t=582s) Mistake 2: Believing the Model Itself Is Your Moat
A widespread strategic error among business leaders is assuming that the underlying foundation model constitutes a defensible competitive advantage. AI fundamentally lowers the barrier to entry for developing intelligent applications, democratizing access to state-of-the-art reasoning capabilities across the market. Paredes breaks this dynamic down into the technical economics of model training versus inference:
```mermaid flowchart LR accTitle: Training versus Inference Economics subgraph Training["Model Training (Commoditized Base)"] D["Massive Datasets"] --> T["Massive GPU Compute"] T --> M["Frontier Base Model\n(e.g., MD Degree)"] end
subgraph Inference["Enterprise Inference (True Differentiation)"] M --> INF["Operational Inference"] P["Proprietary Data & Context"] --> INF INF --> V["Business Value & Moat"] end ```
Training a frontier foundation model requires massive capital expenditure, tens of thousands of specialized GPUs, and vast volumes of data—analogous to sending a student through medical school until they graduate as a general MD. In contrast, the vast majority of enterprise workloads operate strictly at the inference stage: querying the trained model to perform domain-specific tasks, analyze data, and generate decisions, much like an MD diagnosing individual patients. Because top-tier frontier models from OpenAI, Anthropic, Google, and open-source ecosystems are instantly accessible via an API to anyone with an internet connection and a credit card, the model itself functions as a shared utility. Consequently, true long-term competitive advantage cannot come from the base model, forcing organizations to look elsewhere for differentiation.
- **Training vs. Inference:** Foundation training is capital- and compute-intensive, but standard business execution resides entirely within operational inference. - **Universal Model Access:** Proprietary frontier models and open-weights alternatives are equally accessible to competitors off the shelf. - **The Core Strategic Challenge:** If foundation models are fully commoditized utilities, enterprises must identify defensible moats through proprietary workflow integration, proprietary data assets, and execution capabilities.
[11:17](https://www.youtube.com/watch?v=MZRAqkLEgmA&t=677s) Sources of Differentiation Beyond Commodity Intelligence
Base artificial intelligence models are rapidly democratizing; because almost everyone has access to the exact same foundational intelligence, the model itself ceases to provide a durable edge. Differentiation shifts entirely to what surrounds the model. First and foremost is proprietary data. If an enterprise feeds an off-the-shelf model unique, proprietary data that competitors lack, that intelligence can surface insights it could never reveal to anyone else. By default, instances of a model know only what they learned from common training corpuses—analogous to graduates who all memorized the exact same curriculum in medical school. Only proprietary data enables them to offer differential value.
Equally vital is domain knowledge, which functions as an uncodified, unstructured form of data residing inside employees' heads and operational habits. Competitive advantage also stems from the practical skill of knowing how to direct these systems rather than obsessing over the underlying technical mechanics:
> "People who know how to use these tools, not necessarily understand what AI does, what is a GPT, what is a deep learning model. That's really not that important. It's really understanding how to use these tools to get to a purpose."
To illustrate this, Paredes offers an analogy: imagine sitting in a room with Albert Einstein. The group that knows how to prompt him effectively, feed him the right contextual background, and direct his inquiries will extract vastly superior value compared to another room that lacks context and has no idea how to interact with him. While AI still offers an advantage today simply because most organizations use it poorly, general model access will inevitably become commoditized.
- **Proprietary Data Moats:** General models share identical training weights; value emerges only when paired with exclusive internal records. - **Tacit Domain Knowledge:** Critical operational context is rarely digitized or cleanly structured, remaining locked in personnel workflows. - **Contextual Direction:** Prompt engineering and context delivery dictate performance far more than an employee's grasp of transformer mathematics.
---
[13:10](https://www.youtube.com/watch?v=MZRAqkLEgmA&t=790s) The Vanity of Adoption Metrics vs. Actual Transformation
When generative AI first exploded, enterprises reacted with intense risk aversion. Fearing catastrophic intellectual property leaks—such as widely publicized reports of proprietary Samsung semiconductor designs being pasted into external tools—IT security clamped down with strict proxy rules and firewall blocks. Over time, leadership gradually loosened restrictions and began rolling out enterprise tooling: free Microsoft Copilot Web, paid Copilot licenses for select cohorts, OpenAI Enterprise, and Claude Enterprise accounts.
However, management made a fundamental mistake by measuring success through crude adoption numbers. Tracking how many employees have a seat or how frequently they log in provides a deeply misleading signal of progress. High usage numbers can easily mask counterproductive behavior:
> "The question what you should try to measure—and I recognize this is not easy—is AI transformation, transformation of work or tasks. Because I could be using AI very incorrectly, inefficiently, or on things that I shouldn't, and my adoption metrics are going to go up."
- **The Corporate Adoption Arc:** Initial zero-trust lockdown (proxies/firewalls) transitioned directly into mass enterprise seat distribution. - **The Seat-Count Fallacy:** Elevated login and query frequencies often indicate inefficient wandering rather than meaningful business impact. - **True North Metric:** Organizations must look past software utilization rates and evaluate genuine task and workflow transformation.
---
[14:14](https://www.youtube.com/watch?v=MZRAqkLEgmA&t=854s) Slapping AI on Broken Processes and the Agent Overkill Trap
A classic technology implementation failure is taking an inefficient, broken business process and simply overlaying software on top of it; doing so merely causes the process to break at higher velocity. While applying AI to a suboptimal sub-step can occasionally be justified during initial experimentation to benchmark cost-effectiveness against human analog baselines, doing so at scale is disastrous. Leaders must step back and interrogate the fundamental "job to be done" rather than automating historical baggage.
This mistake is directly tied to the industry's rush to deploy autonomous agents. The prevailing consulting advice urges organizations to identify an administrative function—such as claims or receipts processing—and immediately replace it with an agentic workflow. Outside consultants frequently offer well-intentioned but disastrous advice because they do not understand the institutional nuances or unwritten rules that justify why certain steps exist in the first place.
```mermaid flowchart TD accTitle: Evaluation Flow for Business Process Automation subgraph Analysis["1. Process Re-examination"] A["Identify Task or Claim"] --> B{"Interrogate Job to Be Done"} B -->|"Redundant Step"| C["Eliminate Step Entirely"] B -->|"Codified Nuance Present"| D["Preserve Institutional Rule"] end
subgraph Execution["2. Solution Selection"] D --> E{"Assess Cost & Reliability"} E -->|"High Complexity & Latency Risk"| F["Retain Human Execution"] E -->|"Low Cost & Structured Path"| G["Simple Deterministic Script"] E -->|"Overkill / Expensive Route"| H["Autonomous AI Agent"] end ```
Agents are frequently far too expensive, unpredictable, and over-engineered for the problem at hand. In numerous operational environments, human workers remain significantly more reliable, more cost-effective, and simpler to deploy than an unpredictable agent swarm.
- **Automating Flaws:** Layering generative technology over broken processes merely accelerates operational failure. - **Hidden Institutional Knowledge:** Arbitrary-looking organizational checkpoints often exist to safeguard tacit compliance or legal constraints that external advisors overlook. - **Agent Overkill:** Autonomous agents are frequently chosen over cheaper deterministic code or dependable human workflows despite worse unit economics.
---
[16:46](https://www.youtube.com/watch?v=MZRAqkLEgmA&t=1006s) Humans as Competitive Advantage and the Xfinity Cancellation Nightmare
Far from becoming obsolete, human personnel will emerge as a definitive competitive advantage. Even in workflows where an agent is technically capable of handling interactions, deploying a human representative often produces substantially more enterprise value and preserves client loyalty. To highlight the perils of relying exclusively on automated deflection, Paredes recounts a maddening personal experience from the previous day:
> "Yesterday I was trying to cancel my Xfinity internet connection. Have you guys ever tried to cancel cable or internet? I mean, I hope AGI comes and can solve that, because like it's impossible. And I was on a loop with the AI, and I just started playing with it just to see what happens—trying to do some prompt injection—and it was really frustrating."
While automating support journeys reduces immediate payroll expenses, it often inflicts severe, invisible damage on the brand. Certain customer segments will actively pay a premium for human interaction, while in other scenarios, inserting a empathetic human into high-friction touchpoints avoids massive churn and brand erosion.
- **Human Moats:** Human beings represent a premium, value-creating layer in customer-facing workflows rather than merely a cost center to eliminate. - **Deflection Loops:** Rigid conversational bots frequently trap users in inescapable dead ends, driving users to test prompt injections out of pure exasperation. - **Brand Protection:** The immediate labor savings from automated support bots are frequently eclipsed by the long-term cost of alienating high-value accounts.
---
[18:05](https://www.youtube.com/watch?v=MZRAqkLEgmA&t=1085s) The Myth of the Flat Organization and the Evolving Enterprise OS
A popular thesis among industry executives argues that enterprise hierarchies will inevitably collapse into ultra-flat structures where small cadres of managers oversee vast swaths of autonomous agents. Paredes notes his close collaboration with Professor Suraj Srinivasan at Harvard Business School, who recently authored a *Harvard Business Review* paper exploring the emergence of human-agent managers. However, Paredes urges caution regarding claims that corporate management structures will radically flatten:
> "A lot of people claim that that's what's logical. I don't know. I really don't know... I think we need to be humble in knowing that there's—yes, a lot's going to change. We don't know what's going to change... What we have today is an organizational operating system that works in a certain way that is very much of the past."
While tactical operational transformations are already visible in software engineering and customer service, the fundamental enterprise operating model has not yet transformed. Current implementations merely bolt modern machine intelligence onto legacy corporate structures. Discovering what a native post-AI organizational operating system actually looks like will require years of experimentation.
- **Academic Context:** Harvard Business School's Suraj Srinivasan has documented the rise of "human-agent managers" directing digital workforces. - **Historical Parallels:** Past technology waves frequently promised the immediate death of corporate hierarchy, yet managerial layers adapted and survived. - **Legacy Operating System:** Current enterprise structures were engineered for a previous century; merging AI into these frameworks remains an unsolved design problem.
---
[19:32](https://www.youtube.com/watch?v=MZRAqkLEgmA&t=1172s) The Human Bottleneck and the Soccer Latency Metaphor
Accelerating discrete processing tasks with artificial intelligence frequently exposes severe friction elsewhere in the pipeline: the human bottleneck. When automated agents generate deliverables or process requests instantly, downstream workflows inevitably back up at points requiring human validation, oversight, or regulatory approval. Surging upstream throughput without redesigning the broader review architecture creates organizational gridlock.
To illustrate this, Paredes draws on an athletic ambition from his youth:
> "I put a little graph on the right of soccer. I used to want to be a professional soccer player. Unfortunately, I was born in Peru, where our soccer's not that good. And actually, I wasn't that good. And if you think about the ball, if the ball were to go faster between the different players, the players would have to increase also their ability to perceive and make decisions and be able to respond to it in an equally fast way. And that's not usually the case."
Because human decision-makers operate with inherent biological and cognitive latency, simply firing work at them faster results in dropped balls and paralysis. True adoption requires redesigning entire operational paths to balance high-speed AI output against human processing limits.
```mermaid flowchart LR accTitle: High-Velocity Throughput and the Human Bottleneck subgraph Ingestion["Automated Stream"] A["Agentic Generation"] -->|"Instantaneous Output"| B["Accelerated Pipeline"] end
subgraph Bottleneck["Decision Chokepoint"] B -->|"Volume Overload"| C{"Human Approval Gate"} C -.->|"Cognitive Latency"| D["Operational Backlog"] end
subgraph Resolution["Redesigned Workflow"] D --> E["Restructured Governance & Review"] end ```
- **Cognitive Latency:** Machines process inputs near-instantaneously, whereas human supervisors have biological limits on comprehension, perception, and sign-off speed. - **Throughput Havoc:** Accelerating one phase of a workflow without expanding supervisory capacity merely moves and amplifies operational bottlenecks. - **System Redesign:** Successful deployments focus on structural workflow orchestration rather than optimizing standalone execution speeds.
---
[21:05](https://www.youtube.com/watch?v=MZRAqkLEgmA&t=1265s) Ignoring Token Economics and Model Architecture Sizing
A frequent operational blind spot is failing to account for token economics—the underlying compute and inference costs incurred every time text is fragmented into numerical tokens for ingestion and generation. The business models surrounding these systems are shifting rapidly as foundational model providers transition away from subsidized rates intended to secure early market share and enterprise contract lock-in.
Engineering teams routinely make the expensive mistake of routing basic, high-volume enterprise tasks to the newest, most capable frontier models. In reality, sound architecture design dictates model right-sizing:
> "Really this is going to be changing a lot the architecture of how we design AI solutions. Because in many cases, you don't need the latest model. The one-year-old model is good enough. Or an open-weight model is good enough."
Organizations operating massive transactional processes face runaway inference bills if they fail to tier their workloads. Deploying open-weight architectures or older, optimized foundation models for standard routines preserves capital while delivering indistinguishable functional output.
- **Mechanics of Cost:** Ingestion and generation break prompts into discrete tokens whose cumulative inference cost scales rapidly with volume. - **Vanishing Subsidies:** Hyperscalers initially absorbed massive model deployment costs to drive adoption, but per-seat and token pricing structures are steadily normalizing. - **Architectural Tiering:** High-volume operational processes rarely require flagship frontier models; open-weight models or previous-generation systems deliver comparable results at a fraction of the expense.
[22:07](https://www.youtube.com/watch?v=MZRAqkLEgmA&t=1327s) Hidden Costs, Human Adoption, and the Fallacy of Single-Metric Benchmarks
When organizations scale up their use of AI inference and scoring, many look to migrate from paid cloud APIs toward hosting their own on-premise, open-weight models to curb variable per-token costs. Transitioning from variable operational expenditure to upfront capital expenditure can appear more cost-effective on paper, but leaders frequently overlook the substantial hidden costs—most notably the human cost. Because these systems are so new, enterprises lack baseline data on whether employees will actively embrace the tool or quietly boycott the workflow altogether.
```mermaid flowchart TD accTitle: Evaluation and Task Allocation Lifecycle subgraph Eval["Model Evaluation & Monitoring"] A["Define Job to Be Done"] --> B["Profile Multi-Metric Distribution"] B --> C["Continuous Monitoring for Provider Drift"] end subgraph Allocation["Spectrum of Agentification"] D["Human-Only"] --> E["AI-Assisted"] E --> F["AI-Augmented"] F --> G["Fully Autonomous Agentic"] end Eval --> Allocation ```
Another widespread pitfall is treating model performance as a single composite score and asking, "Is this the best model?" The answer depends entirely on the specific dimensions required for the job to be done: factual accuracy, hallucination rates, latency, or domain reasoning. Organizations must evaluate the full distribution of performance across their specific tasks rather than relying on generic leaderboard benchmarks. Furthermore, continuous monitoring is non-negotiable; when major foundation model providers deploy backend updates, downstream enterprise prompts and pipelines frequently break without warning.
- **CapEx vs. OpEx Trade-offs:** Shifting to open-weight on-premise models replaces per-call token fees with fixed hardware and maintenance expenses, but human adoption friction and integration overhead can offset savings. - **Continuous Evaluation:** Upstream provider model updates alter behavioral baselines, necessitating continuous regression testing across multi-metric evaluation suites.
---
[23:28](https://www.youtube.com/watch?v=MZRAqkLEgmA&t=1408s) The Mindset Shift: Electricity, Futures Thinking, and Search Paradigms
The final major mistake organizations make is failing to grasp the sheer magnitude of the required mindset shift. Comparing the current moment to the dawn of electrification, people in the pre-electric era were amazed just seeing an isolated light bulb at an exhibition, yet nobody could envision the downstream derivatives—refrigerators, toasters, telecommunications, or Wi-Fi. In the same way, AI presents an opportunity not merely to optimize existing tasks, but to invent entirely new categories of value creation. To break out of conventional mental models, executive teams should dedicate a half-day or full-day workshop to "futures thinking," examining how their entire industry and customer expectations will transform rather than asking where to tack on AI pilots.
This shift in consumer desires became unmistakable with the release of ChatGPT on November 30, 2022. For decades, users relied on search engines like Google, sifting through pages of ranked links, sponsored ads, and competing sources to assemble their own answers. ChatGPT revealed a latent, serendipitous consumer preference: people often preferred a single, direct, authoritative response, regardless of underlying biases or initial hallucinations. This dynamic has fundamentally reshaped venture building; rather than raising capital first to spend years engineering a product, teams now rapidly assemble functional AI prototypes to test market behavior directly before seeking equity funding.
- **Futures Thinking Workshops:** Executives must look past immediate pain points to map out overarching industry shifts, drawing inspiration from structured methodologies like those developed by IDEO and Google. - **Inversion of Venture Creation:** Rapid generative AI prototyping allows founders to validate product-market fit and user demand before diluting equity.
---
[27:25](https://www.youtube.com/watch?v=MZRAqkLEgmA&t=1645s) Reallocating Task Bundles and the Agentic Spectrum
Every organization is fundamentally a bundle of tasks designed to generate value for customers, shareholders, and employees. Today, virtually every enterprise exhibits a suboptimal allocation of tasks between humans and artificial intelligence. While current AI excels at a discrete set of capabilities, it is expanding rapidly; as the saying goes, "AI is the worst it will ever be today—it will only get better." Maximizing enterprise leverage requires taking an inventory of operational tasks and determining the appropriate degree of delegation.
Rather than viewing automation as a binary choice between human labor and autonomous agents, organizations should evaluate workflows along an "agentic" spectrum—a term popularized by Andrew Ng to reflect degrees of autonomy:
| Level of Autonomy | Operational Structure | Interaction Dynamic | |:--- |:--- |:--- | | **Human-Only** | Full manual execution | AI is not utilized in the workflow. | | **AI-Assisted** | Human-led with targeted speedups | AI accelerates discreet steps (e.g., summarization, drafting). | | **AI-Augmented** | System-led with human-in-the-loop | AI manages the baseline pipeline and prompts humans for edge cases. | | **Autonomous / Fully Agentic** | End-to-end delegation | Agent independently executes actions and API calls within a bounded scope. |
While completely autonomous agents remain limited to narrow domains, the practical value lies in constructing hybrid systems that match the right level of agentification to the specific job to be done.
---
[30:43](https://www.youtube.com/watch?v=MZRAqkLEgmA&t=1843s) Five Strategic Moves to Get Ahead
In previous technology waves—including big data, data science, and classical machine learning—the standard best practice was to align technology roadmaps strictly in support of existing corporate and functional strategies. While generative and agentic AI must continue enabling enterprise goals, they must also actively challenge the overarching strategy. Because what was operationally impossible yesterday is now feasible, corporate strategy itself must adapt to the new capabilities of AI rather than remaining anchored in legacy assumptions.
To build a sustainable competitive advantage, organizations should execute on five decisive moves:
1. **Start with Jobs to Be Done and Redesign Workflows:** Map out fundamental objectives from first principles before layering on software. Recognize that processes designed optimally for a non-AI era are now inherently suboptimal, and avoid focusing exclusively on broken workflows when well-functioning processes can achieve transformative leaps in efficiency. 2. **Choose the Right Level of Agentification:** Deliberately select whether a task requires assisted, augmented, or fully agentic execution rather than forcing binary automation. 3. **Turn Tacit Knowledge into Data and Context:** Capture the unwritten domain expertise, historical operational nuances, and intuition living inside employees' heads. This grounding is essential across software domains as well as physical, embodied, and industrial AI in manufacturing. 4. **Build an Enterprise AI Operating Model:** Develop an integrated, organization-wide operating architecture ("Enterprise 2.0 / AI-native") rather than treating AI as disconnected point solutions. 5. **Challenge Strategy with AI Possibilities:** Continuously revisit what products, business models, and service levels can now be created that were previously unthinkable.
## Notable quotes > [02:45](https://www.youtube.com/watch?v=MZRAqkLEgmA&t=165s) Where do I start?", "Where should I focus?", and "Do I buy or build? — Miguel Paredes
> [11:17](https://www.youtube.com/watch?v=MZRAqkLEgmA&t=677s) People who know how to use these tools, not necessarily understand what AI does, what is a GPT, what is a deep learning model. That's really not that important. It's really understanding how to use these tools to get to a purpose. — Miguel Paredes
> [13:10](https://www.youtube.com/watch?v=MZRAqkLEgmA&t=790s) The question what you should try to measure—and I recognize this is not easy—is AI transformation, transformation of work or tasks. Because I could be using AI very incorrectly, inefficiently, or on things that I shouldn't, and my adoption metrics are going to go up. — Miguel Paredes
> [16:46](https://www.youtube.com/watch?v=MZRAqkLEgmA&t=1006s) Yesterday I was trying to cancel my Xfinity internet connection. Have you guys ever tried to cancel cable or internet? I mean, I hope AGI comes and can solve that, because like it's impossible. And I was on a loop with the AI, and I just started playing with it just to see what happens—trying to do some prompt injection—and it was really frustrating. — Miguel Paredes
> [18:05](https://www.youtube.com/watch?v=MZRAqkLEgmA&t=1085s) A lot of people claim that that's what's logical. I don't know. I really don't know... I think we need to be humble in knowing that there's—yes, a lot's going to change. We don't know what's going to change... What we have today is an organizational operating system that works in a certain way that is very much of the past. — Miguel Paredes
> [19:32](https://www.youtube.com/watch?v=MZRAqkLEgmA&t=1172s) I put a little graph on the right of soccer. I used to want to be a professional soccer player. Unfortunately, I was born in Peru, where our soccer's not that good. And actually, I wasn't that good. And if you think about the ball, if the ball were to go faster between the different players, the players would have to increase also their ability to perceive and make decisions and be able to respond to it in an equally fast way. And that's not usually the case. — Miguel Paredes
> [21:05](https://www.youtube.com/watch?v=MZRAqkLEgmA&t=1265s) Really this is going to be changing a lot the architecture of how we design AI solutions. Because in many cases, you don't need the latest model. The one-year-old model is good enough. Or an open-weight model is good enough. — Miguel Paredes
> [22:07](https://www.youtube.com/watch?v=MZRAqkLEgmA&t=1327s) Is this the best model? — Miguel Paredes
> [27:25](https://www.youtube.com/watch?v=MZRAqkLEgmA&t=1645s) AI is the worst it will ever be today—it will only get better. — Miguel Paredes
## Mentioned Ai4, Miguel Paredes, MIT AI Lab (CSAIL), Albertsons Safeway, Kearney, Harvard Business School, Andrew Ng, AI Fund, Brice Challamel, OpenAI, IMD business school, Anthropic, Google, Microsoft Copilot Web, OpenAI Enterprise, Claude Enterprise, Samsung, Xfinity, Suraj Srinivasan, Harvard Business Review, ChatGPT, IDEO
## Video description Miguel Paredes - Partner / Fellow, Kearney / AI Fund
We explore where most organizations and people go wrong with AI, what are the most common mistakes, and how to get ahead.
[Source](https://www.youtube.com/watch?v=MZRAqkLEgmA)