Divya van Mahajan

The New Problems AI Is Creating (And How People Are Solving Them)

ai-daily-brief-podcast
Listen to episode →

Overview

This episode of the AI Daily Brief — a daily podcast and video covering significant news and discussions in artificial intelligence — examines the new operational and organisational problems that agentic AI is creating inside businesses, and how forward-thinking companies are actively solving them. The host argues that the AI conversation has matured from existential debate (“is AI going to be a thing?”) to sophisticated, practical problem-solving (“how do we do this well?”). No guest speaker is named; the content is presented by the show’s regular host.

Source video: URL not provided in the transcript.

Prerequisites

  • Basic familiarity with large language models (LLMs) and generative AI tools (e.g., ChatGPT, GPT-4)
  • General understanding of enterprise software adoption and SaaS cost models
  • Awareness of the concept of agentic AI (AI systems that act autonomously to complete multi-step tasks)
  • Familiarity with terms such as tokens, token consumption, and inference costs
  • General knowledge of organizational change management principles

Main Points

1. The Conversation Around AI Has Fundamentally Shifted

  • A year ago, significant debate remained about whether AI would prove transformative; enterprises were still hoping the trend would pass.
  • That debate is now largely settled; the conversation has moved to agentic AI use cases and how to operationalize them.
  • The new questions being asked are more sophisticated: how to allocate AI resources, manage costs, and govern usage across an organization.
  • The shift from “if” to “how” represents meaningful progress in enterprise AI maturity.

2. Misconception 1 — AI Will Immediately Generate a Productivity Boom (EY)

  • Historical analogies show that major technologies (steam engine: ~100 years; electricity: ~50 years; computers: ~10 years) take significant time to produce measurable economy-wide productivity gains.
  • The first stage of any technology revolution is infrastructure build-out: data centres, semiconductors, power generation, talent development.
  • Inside organizations, AI productivity is jagged — dramatic gains in some areas, stubborn stagnation in others.
  • The transition to agentic working creates new work (oversight, integration, process redesign) that consumes time in the short term, partially offsetting gains.
  • Organizations are working through these challenges iteratively rather than passively lamenting underperformance.

3. Misconception 2 — AI Is Nearly Free (EY)

  • Unlike traditional SaaS, AI carries a meaningful marginal cost per use: every prompt consumes tokens, compute, and electricity.
  • As AI embeds across organizations, costs accumulate rapidly and shift from a capital expenditure to a recurring operating expense.
  • Many early adopters have exhausted annual AI budgets within months due to higher-than-expected employee usage.
  • Responses include token budgets, usage caps, and tiered governance — but also pathways for employees to apply for additional budget based on demonstrated value.
  • Frontier reasoning models deliver greater capability at higher token cost, intensifying the economics challenge.
  • The host’s framing: organizations should manage AI like any other capital allocation, matching the level of AI “intelligence” to the complexity of the problem.
  • Media coverage misrepresents this as organizational panic; in reality, enterprises anticipated this shift and are adapting deliberately.

4. Misconception 3 — AI Will Make Labor Redundant (EY)

  • Two groups have most strongly promoted this claim: leading AI lab executives and business leaders using AI as justification for layoffs.
  • The host is skeptical of both: lab executives are walking back claims; business leaders were using AI as a convenient, market-accepted excuse.
  • AI will reshape job profiles and have labor market impacts, but mass redundancy is not the observed outcome.
  • Cases of companies rehiring people previously laid off with AI cited as cause are beginning to undermine this narrative.

5. The AI Slop Problem and Institutional Responses

  • Widespread use of AI for writing has produced a flood of low-quality, padded, unverified content — “AI slop.”
  • Social platforms (e.g., LinkedIn) are introducing user-reporting mechanisms (“seems like AI slop” buttons); AI detectors are appearing on platforms like Substack.
  • Clay co-founder Varun Anand published a company-wide AI writing policy with four guiding principles:
    1. Stand behind every idea — authors are responsible for every sentence they share.
    2. Writing is thinking — circumventing the writing process degrades the author’s understanding of the subject.
    3. Respect readers’ time — a document generated from a short prompt and sent to readers who must spend more time consuming it than the author spent creating it is disrespectful.
    4. Longer is not better — if the output of a prompt is long and padded, consider sharing just the prompt instead.
  • Crucially, the policy does not ban AI; it is an injunction against laziness and a defense of process value.
  • The policy attracted ~8,377 engagements on LinkedIn, suggesting rapid uptake across organizations.

6. Redesigning Work Around AI — Lessons from OpenAI’s CFO

  • Sarah Fryer, CFO of OpenAI, published lessons from building an “AI-native finance function,” targeting a zero-day financial close and continuously updated automated forecasting.
  • Five practical lessons she identified:
    1. Give everyone access, then create a reason to use it — pair open access with structured experimentation around real business problems; bottom-up experimentation and top-down strategy are both required.
    2. Finance professionals are becoming builders — 40% of finance professionals’ specialized AI use involves work outside traditional finance; 22% involves engineering-related tasks; teams are building custom dashboards and live tools.
    3. Measure value per unit of intelligence — evaluate each workflow on four questions: Did AI complete work that mattered? What did it actually cost (including human review time)? Was the output good enough to use? Did it help move faster or make a better decision?
    4. Avoid over-simplified ROI metrics — token count alone is insufficient, but requiring proof of ROI for every use case will bias organisations toward easy, low-leverage wins rather than transformative redesign.
    5. Redesign work around decisions that matter — adoption requires clear accountability, room to experiment, and measurement of dependable AI output.

7. Strategic AI Governance — From Vendor Selection to Organizational Architecture

  • Section CEO Greg Shove advised CEOs to avoid both “token maxing” and “token minimizing”; accept a 1–2 year horizon before full productivity gains are visible.
  • Recommendation: identify a lighthouse team for deep, fast transformation rather than spreading enablement thinly across the entire organization.
  • Warning against the 12-month stall — the period after initial rollout excitement when skepticism sets in; the response is to get specific about which teams are blocked and why.
  • BCG Global Chair Rich Lesser noted that CEO questions have evolved from “which model should we use?” to “are we committing too much too soon to an evolving ecosystem?”
  • BCG and Microsoft’s Satya Nadella both emphasize the concept of an enterprise cortex — the organization’s proprietary IP, data, business rules, and process knowledge — as the essential harness that must be owned and controlled, capable of working with any model or combination of models.

8. The De-skilling Risk and the Underinvestment in Human Capability

  • BCG research identifies distributed de-skilling as the risk most leaders are not tracking: the quiet erosion of judgment, critical thinking, and problem-framing across a workforce while adoption metrics look strong.
  • Half of BCG-surveyed leaders said they are already observing this; over 60% expect it to be a real threat within 3–5 years.
  • Fewer than 1 in 5 employees feel confident using AI tools; ~2 in 3 said they would be more willing to support change if their effort was recognized.
  • Token usage is not a proxy for adoption confidence; confidence is built by reinforcing the right behaviors.
  • KPMG’s Adaptability Report found executives are 2x more likely to increase technology investment than to invest in employee training; fewer than 10% named workforce training as a primary objective despite 57% prioritizing performance and efficiency.
  • KPMG data point: among leaders who increased workforce investment, 37% reported revenue growth of 20%+ over three years, versus 25% for business leaders overall.
  • Conclusion: technology investment and talent investment are not a trade-off; organizations that advance both together see better outcomes.

9. The Tragedy of the Cognitive Commons — A Horizon Problem

  • Concept introduced via a paper discussed by Zara Zhang: the tragedy of the cognitive commons.
  • Core argument: checking AI output requires deep expertise; deep expertise is built through years of junior-level “grunt work”; AI is automating grunt work first; therefore, organizations are simultaneously building systems that require expert oversight while eliminating the pipeline that produces experts.
  • Each company acting rationally (eliminating junior roles) produces an irrational collective outcome: professions that eventually cannot catch AI errors.
  • The host frames this as a horizon problem worth examining rather than an inevitability — the key open question is whether there are alternative pathways for developing deep expertise that do not rely solely on traditional junior role pipelines.

Key Concepts

  • Agentic AI: AI systems capable of autonomously executing multi-step tasks with reduced human intervention, representing a shift beyond simple prompt-and-response interaction.
  • AI slop: Low-quality, AI-generated content that is verbose, unverified, and produced without meaningful human judgment or editorial responsibility.
  • Token budget / usage cap: Organizational governance mechanism that limits the number of AI tokens (units of compute consumed per interaction) available to individuals or teams in a given period.
  • Jagged productivity: The uneven distribution of AI-driven productivity gains across tasks and roles within an organization — dramatic in some areas, negligible in others.
  • Enterprise cortex: An organization’s proprietary layer of IP, data, business rules, process knowledge, and strategic context that should form the foundation of its AI architecture, independent of any specific model vendor.
  • Lighthouse team: A focused organizational unit selected for deep, intensive AI transformation to generate a high-visibility proof of concept, as distinct from broad, shallow enterprise-wide rollout.
  • Value per unit of intelligence: A proposed evaluation framework for AI spend that assesses whether the outcome of AI-assisted work justifies its full cost, including human review time and quality.
  • Tragedy of the cognitive commons: The systemic risk that widespread AI automation of junior-level work erodes the shared pool of deep professional expertise across an entire field, eventually degrading the collective ability to supervise and validate AI outputs.
  • Distributed de-skilling: The gradual, organization-wide erosion of critical thinking, judgment, and problem-framing capabilities as AI handles more cognitive tasks, often invisible in standard adoption metrics.
  • Zero-day close: A finance operations ambition in which financial reporting is produced in real time rather than at end-of-period, enabled by AI-driven continuous reconciliation and data integration.
  • Cognitive commons: The shared pool of professional expertise within a field or society from which all practitioners draw, and which is sustained by the accumulation of experience across successive generations of practitioners.

Summary

The central argument of this episode is that AI inside organizations has passed through an initial phase of uncertainty and hype into a more mature and demanding phase of practical problem-solving. The host uses a series of recent publications and posts — from EY, OpenAI’s CFO Sarah Fryer, BCG, KPMG, Clay’s Varun Anand, and Section’s Greg Shove — to illustrate that the new challenges of agentic AI are real, recognized, and being actively addressed by serious organizations. These challenges include the non-linear nature of productivity gains, the shift of AI from a fixed SaaS cost to a variable operating expense requiring capital-allocation discipline, the proliferation of low-quality AI-generated content, the need to redesign work and governance architecture around an organization’s own proprietary knowledge base, and the chronic underinvestment in human capability alongside technology spend. Looking further ahead, the host flags the “tragedy of the cognitive commons” — the structural risk that automating junior-level work destroys the pipeline for producing the expert judgment needed to oversee AI — as a problem worth anticipating before it becomes acute. The overall message is optimistic: the sophistication of the questions being asked has improved dramatically, organizations are sharing their learnings openly, and the path forward lies in continuing to name, examine, and collaboratively solve the new problems that AI’s opportunities inevitably create.