Why Most AI Products Are Just Chatbots Wearing Makeup

 

You've seen this a hundred times.

A polished demo. Natural language input. A confident pitch about "AI-powered transformation." But strip away the interface and branding, and you're left with the same interaction pattern: type a message, get a generated response. Chatbot. Different logo, same chat window.

The problem isn't that these products are useless. It's that most teams, and most PMs, can't tell the difference between a product that uses an LLM and a product that is an LLM wrapper with a login page. That confusion costs money.

The Chatbot Wrapper Epidemic

Here's why this keeps happening:

Speed beats substance. Building a chat interface takes days. Building a real product takes quarters. When leadership wants "an AI strategy" by end-of-quarter, a text box is the fastest thing to ship. Investors reward velocity. The market rewards demos. So teams ship chatbots.

Chat hides the hard decisions. A blank text box avoids the product thinking: Which problem? Which workflow? What counts as done? When the interface is "ask anything," nobody has to answer those questions. The customer does that work instead.

Foundation models make it easy. With a single API call, any team can make something that sounds intelligent. That lowers the barrier to shipping but also lowers differentiation. If your value proposition is just "we wrote a good prompt," it's not defensible.

Here's the uncomfortable part: most teams know this. They know they're shipping makeup. They ship it anyway because the market doesn't punish it fast enough. Investors reward the demo. Early users treat it like a feature. By the time competitors ship something real, your team is already staffed up and your roadmap is locked in. But that's also where the risk lives in the gap between "this shipped" and "this matters."

The Test: Remove the Chat UI. What's Left?

Here's the first question every PM should ask about an "AI product":

If you remove the conversational interface, does the core capability still work?

If the answer is yes; if the product is still valuable; you might have something real. A routing engine, a classification system, a workflow automation. Something that solves a problem differently.

If the answer is no; if it falls apart without the chat; then the chat was the product. Everything else was theater.

Most products in the market fail this test.

Five Levels of AI Product Maturity

Not every "AI product" is created equal. Think of them on a spectrum:

Level 1: AI Interface
The existing product unchanged. AI provides a new way to interact with it.
Example: "Ask our CRM anything." You type questions; AI answers. The underlying workflow stays manual. You still read, decide, copy, paste, execute.

Level 2: AI-Assisted Workflow
AI begins participating in the work, but humans remain the primary orchestrator.
Example: Support ticket arrives → AI summarizes → AI proposes diagnosis → human approves → system updates ticket. You accelerated individual steps, not the workflow.

Level 3: AI-Orchestrated Workflow
The system coordinates the work, not the user.
Example: Support ticket arrives → system classifies → retrieves history → determines resolution → updates CRM → escalates exceptions. You provide an objective; the system executes a process.

Level 4: AI-Native Product
AI is core to the product architecture. The product couldn't exist in this form without it.
Example: A system that continuously observes data, interprets conditions, generates hypotheses, makes decisions, executes actions, evaluates outcomes, and learns from feedback.

Level 5: Adaptive System
The product learns from outcomes and adapts behavior based on changing inputs, feedback, and policies.
Example: A financial system that not only processes transactions but improves its routing logic based on market conditions and outcome patterns.

A product doesn't need Level 5 to be valuable. A narrow Level 3 can create enormous impact if it reliably solves a painful, expensive, frequent problem. The point is: is there actual workflow change, or just a new interface?

Five Questions That Separate Substance From Makeup

1. Does it remove work or create another step?

A superficial product gives users one more place to ask questions. A strong product eliminates steps in the end-to-end job.

• Weak: A meeting-notes chatbot that transcribes and summarizes a call.
• Strong: The system automatically links the meeting to the account, identifies commitments, assigns follow-ups, updates the CRM, drafts customer communications, tracks completion, surfaces unresolved risk.

2. Does the system understand context?

Generic AI produces generic assistance. Real products work with the right context at the right moment.

• Weak: A chatbot that answers questions with public knowledge.
• Strong: A system that knows the employee's role, department, payroll jurisdiction, benefit eligibility, prior cases, and relevant policy version. It produces a decision that is actually useful.

3. Can it safely execute?

Text generation is not execution. I watched a support team adopt an AI solution that drafted perfect responses but never updated the customer record. Six months in, they were maintaining two systems: one for what the AI said, one for what actually happened. Products become materially more valuable when they connect to systems of record, use tools, and produce verified changes; not just suggestions.

• Weak: "Here's what you should do."
• Strong: It updates the CRM, schedules the technician, prepares the claim, flags fraud, initiates the approval workflow. Safely.

4. Is performance measured on outcomes?

Many AI products are judged by impressive examples, not systematic performance. That's a red flag. Ask the vendor to share their worst-case scenario, not their best. Ask what happens in the long tail of edge cases. If they can only show you the highlight reel, you're looking at a demo, not a system.

Real products have metrics: Resolution time down 40%. First-pass approval rate up 65%. Error rate below 2%. These are task-specific measures.

5. Does each interaction make the product better?

A generic chatbot starts every interaction with roughly the same capabilities. A true product accumulates advantage.

• Weak: You ask the same question next month, get the same quality answer.
• Strong: User feedback, edits, approvals, and outcome labels flow into operational learning. The system improves from your usage of it.

What to Build (Or Buy)

If you're shipping an AI product, don't start with the chat box. Start here:

What painful workflow would disappear if this problem was solved?

Name it. Measure it today. Then ask: does AI solve this differently? If yes, does AI fundamentally redesign how the work gets done?

If the answer is yes, build around that capability. Design the interface for the outcome, not for the model. Maybe it's a dashboard. Maybe it's automated workflows that don't require user input at all. Maybe it's a hybrid: chat for exceptions, structured forms for routine work.

Then measure whether it moved the needle. Not "our users can now chat with the system." But "we reduced processing time from 5 days to 8 hours" or "we eliminated 60% of manual categorization" or "customers resolved issues without escalation 3x more often."

The Uncomfortable Truth

Most AI products shipping today are feature releases, not products. They're incremental UI improvements wrapped in the promise of AI. They tap into real demand and real FOMO. They attract funding and headlines.

But they're not solving fundamental problems differently.

The products that will matter in two years are the ones that can answer honestly: "Without AI, could this problem be solved at all?"

If the answer is yes; if it could be solved, just slower or more expensively; you're competing on efficiency, not innovation.

If the answer is no; if this problem didn't exist before AI, or couldn't be solved before; you're building something real.

Most of what's being marketed as AI products right now? Chat layer on top of a thing. That thing would work fine without the chat.

Smart PMs are asking harder questions before they ship. They're asking: "What would we do differently if we couldn't use a chat interface?"

That question will separate the real products from the dressed-up ones.

Holes in Anthropic’s Predictions about USA Economy by 2030

 

 Anthropic’s report is a useful scenario-accounting framework, not a forecast or a full general-equilibrium model. Its main gap is that several outcomes presented as downstream implications -  GDP, labor share, unemployment, and wage divergence - are highly sensitive to simplifying assumptions that are exogenous, coarse, or only weakly empirically identified over a 2026–2030 horizon.

Part 1 can be found here:

LinkedIn: https://www.linkedin.com/pulse/anthropic-predicts-2030s-economy-tushar-jain-qrupc

SubStack: https://tjain13.substack.com/p/anthropic-predicts-2030s-economy?r=w9onv&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true

Blog: https://agileanswer.blogspot.com/2026/09/anthropic-predicts-2030s-economy.html

 

Bottom-line assessment

The paper’s strongest contribution is transparency: it converts five AI-related inputs - task exposure, diffusion, per-task productivity gain, automation versus augmentation, and “reinstatement” of new labor tasks - into comparable economic scenarios. But that also makes its output only as credible as the assumed paths for those inputs and the model’s structural choices.

A compact way to frame the critique:

 

Most consequential modeling gaps

1. The two-sector labor-market abstraction is too coarse

The model divides the economy into only two worker groups:

• “Cognitive” occupations: management, professional, sales, and office work which are AI-exposed.
• “All other” occupations: described broadly as manual and interpersonal work which are assumed not to be directly exposed.

That is analytically convenient but increasingly hard to defend as a near-term economic representation:

• AI exposure varies substantially within professions. A lawyer, radiologist, accountant, teacher, software engineer, marketing analyst, and executive all combine tasks with very different exposure, complementarity, liability, and human-trust requirements.
• Physical and frontline work is not a single insulated residual. Computer vision, planning systems, workflow software, autonomous vehicles, industrial AI, and eventually robotics can affect logistics, manufacturing, retail, construction, health care, and field service albeit on different timelines.
• The model assumes workers are homogeneous within each group and that a worker moving from cognitive work to another occupation immediately receives the destination-group wage. The authors explicitly acknowledge that this omits lost tenure, occupation-specific human capital, retraining costs, and persistent earnings scars.
• It does not represent worker bargaining power, unions, credentialing barriers, licensing, immigration status, disability, age, gender, race, household constraints, or geographic immobility.

Why this changes the conclusion: the report can estimate aggregate unemployment in an exposed block, but not the incidence of harm. In practice, a transition from a $180,000 software role to a $70,000 role with delayed re-employment, relocation costs, and lost career capital is economically different from a simple move between two homogeneous labor pools even if both workers are counted as employed in 2030.

A more credible extension would use a multi-occupation, multi-skill transition matrix tied to task profiles, wage distributions, regional labor demand, credential transferability, and observed worker-flow data.

2. “Automation versus augmentation” is exogenous rather than an economic choice

A central model primitive is the automation share, the fraction of AI-performed task instances that are automated rather than augmented. The report holds this ratio constant within each scenario.

That is an important weakness because firms do not generally receive an immutable “automation” or “augmentation” technology from the model provider. They choose a workflow design based on:

• Model reliability, evaluation coverage, and cost of error.
• Liability and regulatory accountability.
• Human review requirements.
• Integration and change-management costs.
• Data access, systems-of-record integration, and security constraints.
• Customer trust, brand risk, and service quality.
• Relative wages, AI inference cost, and availability of managerial capacity.

In other words, automation is endogenous to prices, risk, organizational design, and institutional constraints. A firm might initially augment workers, then automate narrow subtasks, then redesign the service, or use AI to expand quality, personalization, and output without reducing headcount.

The model’s distributional results are therefore particularly assumption-sensitive. Its labor-share decline is driven in part by treating automated task instances as a direct reassignment from labor to capital, while newly created labor tasks offset this only through a simple reinstatement ratio. Real organizations may create complementary tasks - evaluation, exception handling, customer escalation, process redesign, data governance, security, audit, compliance, and AI supervision that are neither proportional to automated tasks nor confined to the original occupation.

3. New task creation is represented by a single “reinstatement ratio”

The report represents new work through, defined as the number of newly created labor tasks per automated task. It assumes these new tasks are cognitive, appear immediately, enter at the same unit cost as remaining labor tasks, and offset displacement one-for-one in the affected wage bill.

This is a highly restrictive treatment of what may be the decisive unknown in technology-driven labor transitions.

It does not distinguish between:

• New tasks inside existing firms versus entirely new industries.
• Tasks requiring high expertise versus low-paid monitoring work.
• Human-in-the-loop oversight that exists because models remain unreliable versus enduring high-value complementary work.
• Tasks created by increased output demand versus tasks created by regulation, safety, and governance needs.
• New tasks that arrive quickly versus new occupations that take years to emerge.
• Gross job creation versus net job creation after AI reduces the labor intensity of the new sector itself.

The historical analogy is also tricky. Past technology waves created complementary jobs over decades, often alongside institutional and educational adaptation. The paper’s 2026–2030 window is short, yet its extreme scenario assumes rapid capability and adoption while treating the emergence of entirely new task categories through a static parameter rather than a dynamic process.

4. No endogenous adoption model

The paper uses logistic paths for both:

• The mass of tasks that AI affects.
• The share of affected task instances that actually diffuse into use.

That makes the model easy to operate, but adoption is not simply a smooth function of time. Enterprise deployment is often gated by workflow redesign, data quality, procurement, model evaluation, cybersecurity, legal review, integration with legacy systems, change-management capacity, and the availability of complementary capital.

The omitted adoption economics is especially consequential for the “substantial” and “extreme” cases. A capability threshold does not automatically become economy-wide effective deployment. For a model to lower unit costs at scale, it needs adequate accuracy, uptime, governance, input-data access, human escalation procedures, and favorable total cost of ownership.

A stronger version would estimate adoption by industry from observed deployment data rather specifying a scenario-level logistic curve.

Macroeconomic gaps

5. Aggregate demand, price adjustment, and financial conditions are omitted

The report explicitly states that it does not include price rigidities and associated demand effects, apart from slow adjustment in real wages. It therefore cannot produce a feedback loop where displacement reduces income and consumption, depresses demand, weakens hiring, and worsens labor-market disruption.

This matters in both directions:

• Downside: fast layoffs, high household debt burdens, falling confidence, reduced consumption, and weaker local economies could amplify unemployment beyond the model’s search-and-matching mechanism.
• Upside: AI-driven declines in the prices of professional services, software, design, administrative work, health workflows, or business services can raise real incomes and unlock demand for complementary goods and services.
• Investment channel: data-center construction, grid investment, semiconductor capacity, networking, cooling, and facilities can create large near-term demand shocks. Anthropic’s reviewers specifically noted that the model does not include aggregate-demand effects associated with the data-center buildout.
• Monetary and financial channels: interest rates, credit availability, equity valuations, AI investment booms, asset-price corrections, and fiscal responses can materially alter a four-year transition.

The 2030 horizon is precisely the period in which demand, credit, and investment-cycle dynamics are likely to matter most. A longer-run neoclassical allocation model is not sufficient for short-to-medium-run transition analysis.

6. The generic “capital” treatment masks the real bottlenecks

The report uses a single capital good with an upward-sloping supply schedule and a fixed elasticity. It does not distinguish compute from other capital, nor does it explicitly model household saving decisions.

That shortcut is material because “AI capital” is not ordinary homogeneous capital. It includes:

• Advanced chips and their manufacturing supply chain.
• Data-center land, construction, cooling, networking, and power infrastructure.
• Electricity generation, transmission, interconnection queues, and water constraints.
• High-quality proprietary data and data-rights arrangements.
• Specialized AI talent, integration services, evaluation infrastructure, and organizational capital.
• Cloud-market concentration and access to frontier model providers.

Each component has different adjustment speeds, market structures, rents, geographic constraints, financing needs, and policy dependencies. Treating all of this as one smoothly supplied capital stock risks overstating or misallocating the mechanism behind capital returns and labor-share shifts.

The report’s extreme scenario projects the labor share falling from roughly 60% to 45% by 2030, while capital’s share rises to 54.8%. That result is not merely an implication of AI capability; it is strongly mediated by capital-supply elasticity, ownership structure, competition, energy and chip constraints, and fiscal policy none of which are modeled in sufficient detail.

7. Competitive markets and rent allocation are assumed away

The production block relies on competitive factor markets and factor payments exhausting output. But frontier AI is plausibly characterized by:

• Concentrated compute ownership.
• High fixed costs and scale economies.
• Proprietary data and model weights.
• Cloud-provider concentration.
• Platform economics and API dependence.
• IP rights and contractual control of downstream deployment.
• Strategic behavior by firms and states.

These features determine whether gains appear as competitive factor returns, monopoly rents, lower prices for consumers, higher wages for scarce complementary talent, or retained earnings by a small group of firms. A model with perfect competition cannot answer the politically central question: who owns and captures the AI surplus?

This is especially relevant because the paper frames redistribution as a challenge in high-growth scenarios, but its distributional mechanics do not include the market-power and ownership mechanisms that would generate the distribution in the first place.

AI-specific gaps

8. Capability is treated as a single task-level productivity input

The report translates AI into an affected-task mass, a diffusion rate, and a per-instance unit-cost reduction. This is a reasonable first-order accounting device, but it compresses crucial distinctions:

• Benchmark competence versus reliable autonomous completion in real workflows.
• Median performance versus tail-risk error rates.
• One-shot task execution versus long-horizon agency.
• Tool use, memory, planning, verification, coordination, and physical action.
• Human supervision intensity and exception frequency.
• Domain-specific performance where error costs differ by orders of magnitude.
• Security, adversarial robustness, privacy, and model misuse.

An AI system that is 95% successful at a low-stakes drafting task and one that is 95% successful at tax filing, medical triage, code deployment, or financial approval do not create the same economically usable productivity. The report’s unit-cost formulation can obscure the expensive human-review and risk-management layer required to make nominal capability deployable.

9. The model neglects data, quality, and trust constraints

For many knowledge-work applications, the binding constraint is not raw model ability. It is whether the organization has:

• Clean and permissioned data.
• Structured workflows.
• Accessible systems of record.
• Clear quality criteria.
• A way to attribute responsibility for decisions.
• Customer acceptance of machine-mediated services.
• Reliable measurement of realized benefits.

The report does not model these constraints explicitly. Consequently, it may conflate “AI can do a task” with “firms can profitably and safely substitute AI into a production workflow at scale.”

For enterprise architecture, this is arguably the practical gap with the greatest near-term relevance. Deployment often fails not at inference but at process redesign, data integration, governance, and operating-model change.

10. Robotics is excluded, limiting both downside and upside analysis

The report explicitly limits AI exposure to cognitive tasks and does not allow rapid advances in robotics and physical-task automation; it cites this as a reason not to extend the analysis past 2030.

This produces two asymmetries:

• It can understate displacement if embodied AI, perception, manipulation, autonomous logistics, and industrial automation improve rapidly.
• It can overstate relative wage gains for the “all other” group, because the model treats those jobs as largely unexposed recipients of complementary demand.

The report’s result that non-cognitive wages rise strongly in transformative cases depends heavily on the assumption that physical and interpersonal work remains insulated. That should be modeled as a conditional case, not a default structural feature.

11. Innovation is likely under-modeled and incorrectly separated from deployment

The report says its innovation mechanism is deliberately conservative: AI raises research inputs in a semi-endogenous growth model, but physical tasks remain bottlenecks. It notes that this leads to a small innovation contribution through 2030 and that the model omits feedback between research and automation; the authors characterize their innovation effects as likely a lower bound.

That is candid, but it leaves a major hole because the most transformative AI pathways are not simply “more task automation.” They could involve AI accelerating:

• Algorithmic research and model architecture discovery.
• Automated experimentation and scientific reasoning.
• Drug discovery, materials science, energy technologies, and process engineering.
• Semiconductor design and fabrication optimization.
• Code generation, verification, and software production.
• Robotics design and industrial process improvement.
• The production of future AI systems themselves.

At the same time, these feedback loops may be limited by experimental throughput, energy, regulation, manufacturing, tacit knowledge, and physical-world validation. The key issue is not merely that the model is conservative; it is that it does not endogenize which bottlenecks bind as AI improves.

Empirical and methodological gaps

12. The scenario inputs are weakly identified

The report is clear that the three cases are not predictions and have no assigned probabilities. That restraint is good. However, readers can still over-interpret the numerical outputs as empirical forecasts:

• 1.6%, 8.3%, and 32.4% GDP increments by 2030.
• Specific cognitive wage paths.
• A 45.2% labor share in the extreme case.
• 17.9% cognitive-worker unemployment in the extreme case.

These numbers inherit uncertainty from every scenario primitive, parameter calibration, and model abstraction. Yet the presentation can make the output look more precise than the input evidence warrants.

The report would be stronger if it included:

• Probability distributions rather than only three named cases.
• Global sensitivity analysis across all parameters.
• Threshold charts showing which assumptions drive sign changes in wages, employment, and labor share.
• Confidence intervals tied to empirical calibration uncertainty.
• Historical back-testing against prior automation waves.
• Explicit tests against early evidence from AI-exposed occupations and firms.
• Alternative structural models, not just parameter robustness within one framework.

13. Survey expectations are not evidence about future technical or economic outcomes

Anthropic surveyed 10,980 U.S. adults and fed median responses into its model; the resulting median outlook lies near the substantial scenario. That is useful evidence about public expectations and perceived risk, but it is not a credible estimator of AI capability progress, firm adoption, or macroeconomic response.

The survey approach raises several issues:

• Lay respondents may not distinguish between capability, reliability, autonomy, deployment, and economic substitution.
• Expectations can be shaped by media coverage and question framing.
• Median beliefs do not aggregate into an economically meaningful prior.
• The response distribution may hide correlated misconceptions.
• The survey does not substitute for forecasts from domain experts, forecasters, enterprise buyers, labor-market researchers, or revealed-preference data.

The public-expectations module should therefore be framed as a social-perceptions result, not as a calibration anchor for the most likely macroeconomic trajectory.

14. The 2024 baseline and short horizon deserve more stress testing

The report selects 2024 as the pre-AI base period, describing it as the last year before LLM use spread. That choice may be operationally useful, but it is contestable:

• Generative AI adoption began materially before 2024.
• AI-related productivity improvements can be embedded in software and business processes before they appear in formal macro data.
• Adoption and gains are lumpy; a smooth baseline may obscure early diffusion.
• The economy’s 2026–2030 trajectory will also depend on non-AI forces—demographics, fiscal policy, trade, energy, housing, geopolitics, pandemics, and climate shocks.

The short time horizon also intensifies path dependence.

Missing institutions and policy feedback

15. Distribution is treated as a market outcome, not a political-economic process

The report explicitly omits political economy, policy responses, business cycles, financial-market disruptions, and catastrophic risks. These are not peripheral omissions in a transformative-AI scenario; they may determine the realized economic path.

The model excludes, among other factors:

• AI taxation, capital-gains taxation, payroll taxes, and social insurance financing.
• Universal basic income, wage insurance, negative income taxes, transition subsidies, or sovereign wealth funds.
• Antitrust and interoperability policies.
• Labor-law changes, collective bargaining, worker data rights, and algorithmic-management regulation.
• Workforce training systems and licensing reform.
• Export controls, industrial policy, and cross-border capital flows.
• Immigration policy and labor-supply responses.
• IP, copyright, data-rights, and liability regimes.
• National-security constraints and international coordination.

As a result, “labor share falls to 45%” should be read as a result inside a particular market structure and policy vacuum not as a technology-determined outcome.

16. The model is U.S.-centric despite global AI supply chains

The analysis focuses on the United States. Yet frontier AI production and diffusion are globally entangled through chip supply chains, energy markets, international capital, cloud infrastructure, talent flows, offshoring, trade in digital services, and geopolitical restrictions.

A U.S. model can still be valuable, but it needs external-sector treatment to handle questions such as:

• Does AI cause reshoring or accelerate offshoring?
• Are AI rents captured domestically or by foreign investors and suppliers?
• Do global compute constraints limit U.S. deployment?
• How do exchange rates, trade balances, and international competition alter domestic wages?
• What happens when different jurisdictions impose different data, liability, and labor rules?

What the report should add next

A practical Version 2 research agenda would prioritize the following.

1. Replace the two labor groups with an occupation–task–skill network.
   Model heterogeneous workers, mobility costs, retraining time, wage scarring, regional demand, and demographic distributional effects.
2. Endogenize automation and adoption.
   Make firms choose automation, augmentation, human review, and deployment speed based on quality, risk, cost, workflow fit, market competition, and regulation.
3. Separate AI capital into compute, energy, chips, data centers, software, data, and organizational capital.
   This would make capital-supply, rent, and bottleneck projections much more realistic.
4. Add a Keynesian transition block.
   Include consumption, investment, price adjustment, monetary policy, credit conditions, debt, and financial-market dynamics especially because 2026–2030 is a transition horizon rather than a long-run equilibrium.
5. Model market power and ownership.
   Distinguish competitive returns from monopoly rents; track who owns frontier models, compute, cloud infrastructure, and data.
6. Make task creation endogenous and delayed.
   Separate within-firm complementary work, new industries, human oversight, regulatory work, and demand-induced employment; allow their skill requirements and timing to differ.
7. Model embodied AI as a separate uncertainty axis.
   Rather than treating non-cognitive work as unexposed, include slow-, medium-, and fast-robotics cases.
8. Use probabilistic scenario analysis.
   Publish distributions, sensitivity decompositions, break-even thresholds, and alternative structural specifications instead of point estimates under three stylized cases.
9. Validate against microdata continuously.
   Use firm-level adoption, task-level usage, payroll data, job postings, wage trajectories, productivity measures, and worker flows to update priors and reject implausible model settings.
10. Publish the full parameterization and reproducible code.
    The scenario explorer is helpful, but policy-relevant scrutiny requires versioned code, parameter sources, calibration rationale, and machine-readable outputs.

How to interpret the extreme scenario

The extreme scenario is best interpreted as a conditional stress test, not as a balanced central projection. It assumes very rapid diffusion, near-total autonomous completion of much cognitive work, essentially no new knowledge tasks for people, and an economy that absorbs extraordinary AI-related productivity gains over a few years. Anthropic itself notes that some reviewers regarded the extreme case as a thought experiment rather than a scenario, while others questioned whether the modest case understates emerging effects.

That does not make the stress test useless. It makes it useful for questions like:

• What institutional arrangements fail if cognitive work is displaced faster than workers can transition?
• What capital-ownership arrangements would cause gains to concentrate?
• What safety nets, training systems, tax mechanisms, and public-investment capacity need to be ready before disruption occurs?
• Which empirical signals - adoption speed, autonomous-task reliability, job postings, wage changes, capital returns, and labor reallocation - would tell us that reality is moving toward this tail case?

The report is therefore valuable as a transparent risk map. Its limitation is that it should not be mistaken for an empirically grounded estimate of the likely 2030 economy, or for a sufficiently rich model of how enterprise deployment, labor institutions, demand, market structure, and public policy will shape AI’s economic effects.

 

Reference

1.      Economic Scenarios for Transformative AI - https://www-cdn.anthropic.com/files/4zrzovbb/website/cf58f84d46a4a76bf5a5b039ac695fba6b80041c.pdf

2.      What will our economic future look like? - https://www.anthropic.com/institute/econ-scenarios

The 5 Levels of Running Coding Agents

 

I have classified coding using AI agents into five levels, shifting the focus from simple prompting to the logistical challenges of infrastructure and environment management.

The 5 Levels of Agent Management:

1. Level 1: The IDE: Starting with a single agent inside editors like Cursor or Copilot. While easy to use, it can lock you into specific proprietary workflows.
2. Level 2: The Terminal: Running agents as direct processes in the terminal. This offers more control and flexibility by decoupling the agent from a specific IDE.
3. Level 3: Multi-Agent Management: Scaling up to multiple agents simultaneously. Tools like T-Mux or, more specifically, Herder to manage background sessions and track agent status.
4. Level 4: Remote Machine: Moving agents off your local machine to a desktop or server using Tailscale and SSH. This allows for GPU-intensive tasks and ensures agents continue running when you close your laptop.
5. Level 5: The Cloud: Running agents in disposable cloud sandboxes, such as Control Plane. This provides total environmental control, persistence, and the ability to connect agents to internal infrastructure securely.

Level 1

Level 1 represents the starting point for most users, which is running a single AI coding agent directly inside your IDE. Common examples include tools like Cursor or Copilot or Claude Code or Codex integrated into your code editor.

Key characteristics of this level include:

• Workflow simplicity: You provide a task, the agent modifies files, and you review the resulting "diff" before proceeding.
• Ease of use: It is a highly effective way to "vibe code" and complete a significant amount of work without needing additional configuration.
• Constraints: While convenient, the IDE often dictates the agent's workflow, how you interact with it, and what tools are exposed. Additionally, it frequently involves extra subscription costs on top of the base model fees

Level 2

Level 2 shifts the workflow from the IDE to the terminal. By treating the AI agent as a standalone process rather than a plugin, you gain greater control over the environment and avoid proprietary limitations.

Key advantages of this approach include:

• Flexibility: You can choose your preferred model, specific machine, and custom environment tools without being locked into a single editor's ecosystem.
• Automation: It becomes significantly easier to script interactions, chain multiple tools together, or move your entire setup via SSH to different machines.
• Open Ecosystem: This level often relies on open-source tools such as OpenCode or Claude via command-line interfaces which can be more cost-effective than proprietary IDE subscriptions.

Level 3

Level 3 focuses on the challenges of managing multiple AI agents simultaneously. As you scale your projects, you will likely find yourself needing to run several agents at once, which can quickly lead to terminal clutter and difficulty tracking the status of each task.

Key tools for this level include:

• T-Mux: A traditional terminal multiplexer that allows you to keep multiple sessions alive in the background. It is effective for managing numerous processes through different panes, enabling you to detach and reattach to your workflows easily.
• Herder: A modern alternative to T-Mux specifically designed for agent management. Herder solves the "mental overhead" of tracking which agents are active, waiting for input, or have completed their tasks, making it a more purpose-built solution for this workflow.

Ultimately, this level is about moving from manually juggling terminal windows to utilizing specialized software that organizes your agent fleet, preventing the chaos of having dozens of processes running without clear oversight.

Level 4

Level 4 focuses on decoupling your coding agents from your local hardware. As you scale, you may run into limitations like needing a more powerful GPU for machine learning or simply wanting your agents to continue working without being tied to your laptop.

Key strategies for this level include:

• Remote Access via SSH: By setting up your secondary machine (e.g., a desktop) and your laptop on the same Tailscale network, you can remotely control your powerful machine from anywhere.
• Background Persistence: Using T-Mux or Herder on a remote machine ensures that your agents keep running even if you close your laptop.
• Simplified Workflows: Using the remote flag in Herder to handle the SSH connection automatically, making it seamless to manage your agent fleet across different physical devices

Level 5

Level 5 moves the entire agent infrastructure away from physical hardware and into disposable cloud environments. This stage focuses on gaining maximum control while eliminating dependencies on your local machine.

Key features of this level include:

• Environmental Control: Unlike vendor-managed solutions (like Cursor or Claude's built-in cloud agents), using tools like Control Plane allows you to define custom images containing your specific dot files, tools, and preferred agent workflows.
• Persistence and Flexibility: You can spin up a sandbox in under a minute, hand it tasks to run uninterrupted, and then suspend the sandbox when idle to reduce compute costs to zero, resuming exactly where you left off later.
• Secure Infrastructure Access: A major advantage over IDE-based cloud agents is the ability to securely connect your sandbox to internal infrastructure (e.g., AWS, GCP, or Azure) using identity management, enabling the agents to interact with your private services without compromising security.

Anthropic Predicts 2030's Economy

 

Here is Part 1 detailing the report published by Anthropic regarding the potential impacts of AI on the American economy by 2030. The report was developed by economists Anton Korinek and Charles Jones and reviewed by experts including Nobel laureate Daron Acemoglu and David Autor. It uses a task-based framework to model three distinct economic scenarios.

Part 2 will detail the holes in the report.

Key Concepts and Findings

The Task-Based Framework

The model views every job as a " dynamic bundle of individual tasks." This framework allows economists to model the complex impact of AI on the labor market by breaking the entire $30 trillion US economy down into these parts.

Key components of this framework:

• Task Bundling: A job consists of many different tasks performed daily. For example, a nurse's job includes triage, drawing blood, charting, and ordering supplies. These bundles are not static; they evolve as technologies change which tasks are required.
• AI's Impact on Tasks: When AI is introduced, it interacts with these task bundles in three primary ways:
• Augmentation: AI assists the human worker with a task, potentially making it faster or more efficient.
• Automation: AI takes over a specific task entirely, removing the need for human involvement in that specific action.
• Task Creation: New tasks are generated that require human intervention, such as monitoring or verifying the AI's output.
• Economic Aggregation: By analyzing how AI affects millions of tasks across every sector, the model can project how these individual changes at the task level aggregate into macro-economic shifts, such as changes in GDP, employment rates, and wage distributions.

The Five Dials

The economic model presented in the report is governed by five key dials (variables), which function as controls to project different future scenarios for the American economy by 2030.

1. Capability: Measures what fraction of knowledge work (tasks done with heads, not hands) AI will be able to perform as well as a trained professional, such as writing, bookkeeping, or coding.
2. Adoption: Determines how much of AI's capable work is actually put into practice by companies and individuals.
3. Autonomy: Distinguishes between augmentation (AI helping a person, like cruise control) and automation (AI doing the task alone, like a self-driving car).
4. Productivity: Estimates how much faster a task gets completed when AI is involved.
5. Adjustment: Models how long it takes for a person displaced by AI to find a new job or switch careers.

Additionally, there is a sixth, deeper assumption in the model: for every task AI automates, how many new human tasks are created? While historically this ratio has been one new task for every two automated, the "worst-case" scenario models this ratio at zero

Three Scenarios

The Modest scenario

It represents a world where AI's impact on the economy is significant but aligns with historical patterns of technological progress. In this scenario, the economy grows slowly, and job market adjustments remain manageable.

Here are the specific settings for the five dials:

• Capability: AI is capable of handling about one-fifth (20%) of all knowledge work by 2030.
• Adoption: People actually use AI on only about one-fifth of the tasks it is capable of performing.
• Autonomy: When AI is applied, it is split half and half between helping a person (augmentation) and working alone (automation).
• Productivity: AI makes tasks approximately 35% faster.
• Adjustment: The difficulty of switching careers remains the same as it is today; there is no added friction.

Additionally, the model assumes that for every two tasks automated, one new human task is created. The resulting economic impact is described as roughly similar to the introduction of the internet, showing up slowly and fitting into historical trends

The Substantial scenario

It represents a future where AI's impact is more significant than historical technological shifts like the internet, leading to faster economic growth and significant labor market shifts.

Here are the specific settings for the five dials:

• Capability: AI can perform about half (50%) of all knowledge work tasks.
• Adoption: Adoption is lower than capability, with people actually using AI on about 40% of what it can do. This means roughly one in five knowledge work tasks is touched by AI.
• Productivity: Tasks touched by AI get done more than 50% faster.
• Autonomy: AI operates with higher independence, doing the task alone three out of four times.
• Adjustment: Switching careers becomes more challenging, roughly twice as hard as it is today.

In this scenario, the economy grows at twice its normal speed. Overall, while the economy expands significantly, this path creates more friction in the labor market compared to the Modest scenario, as workers must navigate these changes.

The Extreme scenario

It represents a highly transformative future for economy by 2030, characterized by massive economic growth (up to 15% annually), but nearly 1 in 5 knowledge workers are displaced, and capital owners gain a larger share of the economy.

Here are the specific settings for the five dials:

• Capability: AI is capable of handling the vast majority of all knowledge work tasks.
• Adoption: AI is adopted and used on more than half of all tasks it is capable of performing.
• Autonomy: In this scenario, AI operates almost entirely independently; nine out of ten times, there is no human in the loop.
• Productivity: When AI touches a task, it more than doubles the productivity. This scenario also accounts for the AI improving itself, which is baked into the productivity dial.
• Adjustment: Switching careers becomes four times harder than it is today because a massive segment of the workforce is attempting to transition at once.

Additionally, this scenario assumes that essentially no new human tasks are created to replace those automated by AI

Key Economic Findings

The report identifies four primary economic findings concerning the potential impact of AI by 2030. These findings highlight the tension between overall economic growth and individual worker outcomes:

• Economy Grows: In every scenario modeled, the economy expands, ranging from a 1.6% increase in the modest scenario to a 32% increase in the extreme scenario.

• The "Musical Chairs" of Labor: AI triggers a significant shift in labor demand. Knowledge workers (office, professional, and management roles) face high rates of displacement, while demand for physical, hands-on work (like nurses, electricians, and construction workers) increases. The unemployment rate for knowledge workers could rise significantly (up to 17.9% in the extreme scenario) because wages adjust slowly and switching careers is difficult.

• Wage Divergence: While average wages may technically rise across the economy, this hides a stark reality: knowledge worker pay often stagnates or declines relative to a no-AI baseline, while pay for physical labor increases due to relative scarcity.

• The Capital-Labor Split: For the first time, a larger share of the economy's output flows to capital (owners of machines, data centers, and software) rather than labor (the workers). In the extreme scenario, the labor share drops to 45%, while capital's share rises to 55%.

Model’s limitations

The Anthropic report explicitly acknowledges several limitations and factors that the model does not account for. These omissions are critical to keep in mind when interpreting the projections:

• No Robotics: The model focuses exclusively on knowledge work - tasks done with heads rather than hands. It does not account for the impact of capable robots entering physical labor markets, which could lead to significantly worse job displacement.
• No Policy Response: The model assumes the government takes no action. It does not factor in potential policy interventions or responses to economic shifts.
• No Economic Volatility: The analysis ignores macroeconomic cycles, such as booms, busts, financial crises, or recessions.
• No Demand Boosts: The model does not account for the economic stimulus created by massive infrastructure investments, such as the billions currently being spent on data center construction.

Additionally, reviewers noted that the model does not follow the experiences of individual workers, making it difficult to assess the personal severity of job loss

Watch out for

The Anthropic report emphasizes that the next 18 months are crucial for determining which economic path the U.S. will follow. Because the modest, substantial, and extreme scenarios share nearly identical settings today, the real-world data gathered in the near future will reveal the true trajectory of AI adoption.

Practical indicators to monitor

• Corporate Adoption Strategies: The most critical factor is how companies choose to integrate AI once they adopt it. Observe whether firms are using AI to produce more output with the same number of people (augmentation) or to produce the same amount of output with fewer people (replacement).
• Labor Demand Shifts: Keep an eye on employment data for knowledge workers versus physical labor roles. You should monitor if the 'musical chairs' effect starts occurring, where demand for word-and-number-based jobs softens while demand for hands-on, physical jobs (like nursing, construction, or electrical work) increases due to relative scarcity.
• Wage Trends: Track whether knowledge worker wage growth begins to decouple from the broader economy. Specifically, check if professional and office worker pay starts to flatten or decline compared to historical norms, while wages for physical, human-centric roles rise.
• Capital Investment: Watch for shifts in the 60/40 labor-to-capital income split. An accelerated move toward capital-heavy income, where more revenue flows to owners of AI infrastructure and data centers rather than to employees, will signal that we are moving toward the more extreme scenarios

Reference

1.      Economic Scenarios for Transformative AI - https://www-cdn.anthropic.com/files/4zrzovbb/website/cf58f84d46a4a76bf5a5b039ac695fba6b80041c.pdf

2.      What will our economic future look like? - https://www.anthropic.com/institute/econ-scenarios