McKinsey & Company · Published 2026-09-22

Improving the economics of agentic AI: A guide for business leaders

privateReview: unreviewedConfidence: high
Agentic AItoken economicsoperating modelroiworkflowsagent-economicstoken-costverificationcost-varianceagentic-workflowsworkflow-redesigncontext-ownershipagent-roi
Open on YouTube ↗

Summary

Overview

  • Speaker: Lucia Rahilly (Host), Tanguy Catlin, Lari Hämäläinen
  • Channel: McKinsey & Company
  • Main topic: Improving the economics of agentic AI and ROI for enterprise business leaders
  • Purpose: To provide business leaders with research-backed guidance on improving the economic returns of agentic AI through deliberate workflow redesign, cost management, and strategic capability ownership. McKinsey leaders discuss the current state of AI adoption, the gap between individual productivity gains and enterprise-level return on investment (ROI), token usage and pricing dynamics, agentic workflows, and a five-step CEO to-do list for building an agentic operating model.

Topic Map

The State of AI Adoption and ROI Gap

  • Explanation: Survey data reveals that while 89% of enterprises have adopted AI, only 6% report meaningful EBIT impact ('AI high performers'). Spend is increasing rapidly, but ROI remains elusive for most.
  • Key claims:
    • 89% of enterprises claim to have adopted AI.
    • Only 6% of organizations are classified as AI high performers.
    • 60% plan to significantly increase AI investments in the coming 12 months.
  • Examples:
    • One-third of organizations spend more than 10% of their total ICT budget on AI.
  • Terminology:
    • AI high performers
    • EBIT impact
    • token costs
  • Why it matters: Highlights that simple adoption of AI tools does not equal value creation; structural redesign and focused execution are required.

Token Usage and LLM Pricing Dynamics

  • Explanation: Analysis of token consumption and pricing shows that while raw model inference capability has become cheaper per token, token consumption has grown exponentially due to agentic workflows and complex coding assistants.
  • Key claims:
    • Token usage has grown by roughly 2,000% CAGR.
    • Frontier model inference capability is significantly cheaper per unit, but total cost is driven up by massive token consumption.
    • 93% of CIOs report exceeding their AI token budget.
  • Examples:
    • Coding agents write 10x more code than humans, driving up token consumption.
  • Terminology:
    • LLM price index
    • token usage
    • frontier models
    • inference cost
  • Why it matters: Enterprises face unexpected budget constraints as agentic workflows consume exponentially more tokens than simple chat interfaces.

Economics of Intelligence and Agentic Workflows

  • Explanation: Framework for evaluating agentic ROI based on probability of success, verification time, and human task time. Agentic workflows show a 30x cost variance on identical tasks.
  • Key claims:
    • Agentic workflows introduce up to 30x cost variance on identical tasks due to probabilistic paths.
    • An AI agent lifts productivity when its probability of success clears the ratio of verification time to do-it-yourself time.
    • 60% of agentic tokens are spent on refinement, not initial generation.
  • Examples:
    • Comparing single-agent vs multi-agent swarms and tool call protocols (MCP vs direct API).
  • Terminology:
    • agentic workflows
    • cost variance
    • refinement tokens
    • probability of success
  • Why it matters: Helps leaders understand when an AI agent is economically worth deploying based on success probability and verification overhead.

The CEO To-Do List for Agentic Operating Models

  • Explanation: Five strategic steps for CEOs to successfully capture value from agentic AI, including defining differentiation, owning proprietary context, redesigning workflows, managing costs, and rethinking firm boundaries.
  • Key claims:
    • AI transformation at its heart is a people and process transformation.
    • Organizations must build agentic capability as a discipline and redesign workflows end-to-end rather than inserting agents into existing ones.
  • Examples:
    • Running 2-3 lighthouses in the company that truly matter and rebuilding them with AI.
  • Terminology:
    • agentic operating model
    • proprietary context
    • firm boundary
    • lighthouse
  • Why it matters: Provides a practical action plan for executives to transition from tool-level productivity to enterprise-scale value creation.

Key Points

Adoption vs. Value Creation

  • Explanation: Most companies are using AI for personal tooling (chatbots) rather than enterprise-wide re-engineering.
  • Evidence: 89% adoption rate versus only 6% reporting positive EBIT impact.
  • Practical implication: Companies must move beyond individual productivity gains to redesign end-to-end business workflows.

Token Consumption Surge

  • Explanation: Agentic workflows and automated coding assistants generate massive token usage, causing budgets to be exceeded.
  • Evidence: 93% of CIOs report exceeding their AI token budget; token usage growing at ~2,000% CAGR.
  • Practical implication: Establish strict governance and token efficiency measures to control operational costs.

The Economics of Intelligence Formula

  • Explanation: An AI agent pays for itself if its success probability clears the verification-to-task time ratio.
  • Evidence: P(success) > verification time / do-it-yourself time.
  • Practical implication: Evaluate agent deployments based on error recovery costs and verification overhead.

Frameworks, Models & Processes

Economics of Intelligence / Agent ROI Framework

  • How it works: Compares the probability of agent success against the time required to verify the output versus doing the task manually.
  • Components:
    • Probability the agent succeeds (P)
    • Time to check the agent's work (verify)
    • Time for human to do the task (do)
  • When to use: When deciding whether to deploy an AI agent for a specific enterprise task.

The CEO To-Do List for Building the Agentic Operating Model

  • How it works: A five-step roadmap for enterprise leaders to successfully integrate and scale agentic AI.
  • Components:
    • Set differentiation strategy before technology strategy
    • Own your data on proprietary context
    • Build agentic capability as a discipline and redesign workflows
    • Reshape the boundary of the firm (deliberate build, buy, partner)
    • Run economics as tightly as deployment
  • When to use: When planning or executing an enterprise AI transformation.

Examples & Case Studies

Coding agents writing code autonomously in software development.

  • Illustrates: How agentic workflows consume significantly more tokens than human developers.
  • Lesson: Software development has high token usage because agents rewrite large codebases iteratively.

Actionable Takeaways

  • Immediate:
    • Audit current enterprise AI token consumption and budget overruns.
    • Identify 2-3 critical 'lighthouse' workflows to redesign end-to-end with AI.
    • Implement rigorous measurement of verification versus generation costs.
  • Strategic:
    • Transition from individual productivity tooling to agentic workflows and 'agentic factories'.
    • Own proprietary context and data to maintain competitive advantage.
    • Reshape the firm's boundaries by deciding what technology capabilities to build in-house vs. outsource.
  • Questions to investigate:
    • Are we deploying AI on top of existing broken workflows or redesigning them from scratch?
    • How do we measure the true cost of agentic error recovery and verification?
    • What proprietary context do we own that gives our AI agents a unique advantage?

Claims Worth Verifying

  • 89% of enterprises have adopted AI. (survey statistic)
  • Only 6% of companies report positive EBIT impact from AI. (survey statistic)
  • Token usage has grown by roughly 2,000% CAGR. (market metric)
  • Agentic workflows show up to 30x cost variance on identical tasks. (research finding)

Notable Quotes

"Every AI transformation at its heart is a people transformation." (at 1:57) "Redesigning an operating model is a full-contact sport. It requires leaders to test, learn, and change how the organization works in practice." (at 2:07) "Everyone is racing toward AI, but the ROI is still elusive." (at 11:42)

Compressed Summary

  • 89% AI adoption, but only 6% achieve significant EBIT impact.
  • Token consumption is growing exponentially at ~2,000% CAGR.
  • Agentic workflows exhibit up to 30x cost variance on identical tasks.
  • CEOs must redesign end-to-end workflows and own proprietary data context.
  • Keywords: agentic ai, token economics, operating model, roi, workflows
  • Core insight: Enterprise AI success requires moving beyond individual productivity tooling to end-to-end workflow redesign and rigorous economic management of agentic systems.

Core insights

7
Mental Modelhigh noveltymoderate evidence

Agent deployment has an explicit economic threshold: an agent only lifts productivity when P(success) exceeds the ratio of verification time to do-it-yourself time. Verification cost, not raw model quality, is the gating variable for whether an agent is worth running.

Why it matters

Gives engineers a computable go/no-go criterion for a given task instead of intuition, and reframes the design goal from 'maximize model accuracy' to 'make verification cheap relative to the task' (cheap checkers, deterministic assertions, tight output contracts).

Generalization

Any automated system with a probabilistic success rate and a required human review step should be evaluated on the review-cost-to-manual-cost ratio, not just accuracy; this applies to code review bots, data pipelines, and document generation alike.

An AI agent lifts productivity when its probability of success clears the ratio of verification time to do-it-yourself time.
Open source video
P(success) > verification time / do-it-yourself time.
Open source video
Empirical Resulthigh noveltymoderate evidence

Identical agentic tasks can vary up to 30x in cost because agents take probabilistic execution paths, whereas classic software has near-deterministic unit cost.

Why it matters

Cost becomes a distribution rather than a fixed unit price, which breaks per-task budgeting, capacity planning, and any SLAs that assume predictable spend. It implies retry budgets, per-task cost caps, and step-count/loop-count limits as first-class runtime primitives.

Generalization

Any system with probabilistic control flow (retries, tool-call loops, self-repair) needs variance-aware cost controls, not average-cost accounting.

Agentic workflows introduce up to 30x cost variance on identical tasks due to probabilistic paths.
Open source video
Mechanismhigh noveltymoderate evidence

60% of agentic token spend goes to refinement rather than initial generation, making the repair/verify loop — not the first-shot generation — the dominant cost center of an agent architecture.

Why it matters

Optimization effort should target the loop structure (how many refinement passes, what triggers a retry, how much context is re-sent each pass) instead of first-token generation quality. It also explains why naive 'bigger model' upgrades may not reduce total cost.

Generalization

In iterative generation systems, cost is governed by iteration count and context re-transmission per iteration; architectural choices that bound iterations usually beat model-quality improvements for cost control.

60% of agentic tokens are spent on refinement, not initial generation.
Open source video
Tradeoffmedium noveltymoderate evidence

Unit inference price is falling while total token consumption grows roughly 2,000% CAGR, so cheaper per-token models invert into higher absolute spend; 93% of CIOs report exceeding their AI token budget.

Why it matters

Classic cloud optimization intuition (wait for prices to fall) fails here: agentic workflows and coding assistants expand consumption faster than price declines. Cost observability, per-workflow token attribution, and consumption guardrails become required production infrastructure.

Generalization

For any technology where capability improvements unlock new consumption patterns (Jevons-style effects), track total cost of consumption, not unit price, as the governing metric.

Token usage has grown by roughly 2,000% CAGR.
Open source video
93% of CIOs report exceeding their AI token budget.
Open source video
Frontier model inference capability is significantly cheaper per unit, but total cost is driven up by massive token consumption.
Open source video
Practicemedium noveltymoderate evidence

Inserting agents into existing workflows fails; value appears only when workflows are redesigned end-to-end around agent capability, piloted via 2-3 'lighthouse' processes rebuilt from scratch.

Why it matters

Legacy workflow structure encodes assumptions (human latency, handoffs, batch steps) that agents either inherit as waste or break; rebuilding a few high-value processes end-to-end is a better architectural strategy than broad shallow insertion.

Generalization

Automation layered onto an unchanged process inherits the process's design constraints; end-to-end redesign of a bounded, high-value process is the reliable path to structural gains.

Organizations must build agentic capability as a discipline and redesign workflows end-to-end rather than inserting agents into existing ones.
Open source video
Running 2-3 lighthouses in the company that truly matter and rebuilding them with AI.
Open source video
Failure Modemedium noveltyweak evidence

Autonomous coding agents produce roughly 10x more code than humans, shifting the bottleneck from generation to human review capacity — a generation/verification asymmetry that mirrors the agent ROI threshold.

Why it matters

If generation throughput outpaces verification throughput, downstream review queues become the real constraint and token spend grows without delivery throughput. Architecturally this argues for automated verification, static analysis, and test harnesses co-deployed with any high-volume generation agent.

Generalization

Whenever generation is cheap and verification is expensive, the system's effective throughput is bounded by verification capacity, not generation capacity.

Coding agents write 10x more code than humans, driving up token consumption.
Open source video
Empirical Resultlow noveltystrong evidence

89% of enterprises report AI adoption but only 6% report meaningful EBIT impact, indicating that tool-level deployment does not translate into enterprise-level return without structural change.

Why it matters

Signals a measurement and attribution gap between productivity tooling and value capture; engineers building internal agents should instrument outcome-level metrics rather than adoption or usage counts.

Generalization

Adoption metrics are not value metrics; for any capability rollout, instrument the business outcome directly and expect a wide gap between usage and realized return.

89% of enterprises claim to have adopted AI.
Open source video
Only 6% of organizations are classified as AI high performers.
Open source video

Deep dives

5

Verification cost as the gating variable in agent ROI

Research question

Can the agent ROI threshold P(success) > verification_time / do_time be operationalized into a per-task deployment gate, and how much does including failure-recovery and re-run cost move the threshold in practice?

Why

It converts agent deployment from intuition into a computable go/no-go criterion and reframes design goals from maximizing model accuracy to minimizing verification time (deterministic assertions, cheap checkers, tight output contracts).

An AI agent lifts productivity when its probability of success clears the ratio of verification time to do-it-yourself time.
Open source video
P(success) > verification time / do-it-yourself time.
Open source video
Source video

Bounding cost variance in probabilistic agent execution paths

Research question

Do runtime cost governors (per-task spend ceilings, retry budgets, step/loop caps) reduce cost variance on identical tasks without materially degrading task success rate?

Why

Cost behaves as a distribution rather than a unit price in agentic workflows, which breaks average-cost budgeting, capacity planning, and SLAs; bounded variance is a prerequisite for production budgets and predictable spend.

Agentic workflows introduce up to 30x cost variance on identical tasks due to probabilistic paths.
Open source video
Source video

Refinement loop topology as the dominant token cost lever

Research question

Which loop topologies (single-shot with strong output contracts, bounded self-repair, critic-verifier loops) minimize refinement-token share while preserving end-to-end task success, and how much of refinement cost is context re-transmission versus additional passes?

Why

Most token spend sits in the repair/verify loop, not first-shot generation, so optimization should target control flow, iteration bounds, and context caching rather than bigger models.

60% of agentic tokens are spent on refinement, not initial generation.
Open source video
Source video

Tool protocol and topology as economics decisions (MCP vs direct API, single agent vs swarm)

Research question

Do multi-agent swarms and MCP-style tool protocols earn their added token overhead and path variance relative to single agents with direct API calls on identical tasks?

Why

Tool integration choices are usually made on convenience alone; treating protocol and topology as cost/variance variables forces multi-agent designs to justify their coordination and token cost.

Comparing single-agent vs multi-agent swarms and tool call protocols (MCP vs direct API).
Open source video
Source video

Separating attribution failure from structural failure in the adoption-to-EBIT gap

Research question

Is the gap between near-universal AI adoption and rare EBIT impact driven by organizations measuring differently, or by organizations architecting differently (workflow redesign and context ownership vs tool insertion)?

Why

It determines whether the fix is outcome-attribution instrumentation or architectural redesign, and it tells engineers to instrument business outcomes rather than adoption or usage counts.

89% of enterprises claim to have adopted AI.
Open source video
Only 6% of organizations are classified as AI high performers.
Open source video
Source video

Article ideas

4

Stop optimizing tokens per unit and start optimizing tokens per verified outcome

The classic cloud cost playbook — wait for unit prices to fall — is inverted in agentic AI: cheaper per-token inference unlocks more ambitious consumption, so total spend is governed by loop count, context re-transmission, and verification cost rather than model price. Cost control is therefore a control-flow and verification design problem, not a procurement problem.

Angle

Argue that cost observability, per-workflow token attribution, and retry/step limits are production infrastructure, and that unit-price metrics are actively misleading when consumption grows faster than price declines.

Source video

Verification is the product: designing agents around the cheapest possible check

Because an agent only pays off when success probability exceeds the ratio of verification time to manual task time, verification is the cost-determining subsystem of any agent architecture. Teams that treat human review as an informal afterthought will deploy agents that lose money on tasks a person would have finished faster.

Angle

Derive concrete design consequences: output schemas chosen for cheap checkability, deterministic assertions and static analysis co-deployed with generation, and bounded human review surfaced as an explicit line item.

Source video

Rebuild the process or inherit the waste: why agent insertion fails

Inserting agents into legacy workflows inherits the original design assumptions — human latency, handoffs, batch steps — as new waste. Structural gains come from end-to-end redesign of a few high-value processes rebuilt from scratch, piloted as lighthouses rather than spread as shallow insertion.

Angle

Contrast shallow tool insertion against end-to-end redesign of 2-3 lighthouse processes, and connect redesign to proprietary context ownership as the reason the result is defensible rather than replicable.

Source video

When generation outpaces verification, throughput is capped by review

Autonomous coding agents produce far more output than humans, which shifts the bottleneck from generation to human review capacity. Any high-volume generation agent deployed without automated verification and test harnesses will increase token spend without increasing delivered throughput.

Angle

Frame the generation/verification asymmetry as the same ROI threshold seen from the delivery side, and prescribe co-deployed automated verification as a release requirement rather than a follow-up project.

Source video

Project ideas

4

Agent ROI Gate

gatehouse

A deployment gate that scores candidate tasks on measured generation time, verification time, and manual baseline time will predict which tasks clear P(success) > verify/do, and tasks above the threshold will show measurable productivity lift while those below will not.

Proof of concept

Instrument 20-30 candidate tasks across coding, research, and transactional workflows; log manual time, agent time, success rate, and verification time; rank tasks by the computed ratio and run a controlled A/B on the top and bottom ranked tasks.

Measurement

Predicted-vs-actual productivity delta (agent time + verification time vs manual time), false-positive deployment rate, and correlation between the pre-computed ratio and observed lift.

Source video

Variance Governor Runtime

new

Enforcing per-task spend ceilings, retry budgets, and step/loop caps at runtime will cut the spread between p50 and p95 task cost on identical tasks substantially while keeping task success rate within a small tolerance of ungoverned execution.

Proof of concept

Run a fixed task suite with identical inputs across an ungoverned agent and a governed agent with configurable loop bounds, retry caps, and hard per-task spend ceilings; log per-task token cost distribution and success rate.

Measurement

Ratio of p95 to p50 task cost, share of tasks hitting the ceiling, change in success rate, and reduction in budget-overrun incidents at the workflow level.

Source video

Refinement Budget Benchmark

beyond-evals

Loop topologies that separate generation from bounded refinement and minimize context re-sent per pass will reduce refinement-token share materially versus an unbounded emergent self-repair loop, without reducing end-to-end task success.

Proof of concept

Implement three topologies on the same task set — single-shot with strong output contracts, bounded self-repair (fixed pass count), and critic-verifier loop — with token accounting that splits generation, refinement, and re-transmitted context.

Measurement

Refinement-token share of total spend per topology, total tokens per successful task, and task success rate; test whether refinement share drops below the ~60% baseline.

Source video

Lighthouse Outcome Attribution Harness

movement-lab

Rebuilding 2-3 high-value processes end-to-end with agentic workflows and instrumenting outcome-level metrics will show materially larger EBIT-linked impact than agent insertion into the same processes, and the gap will be invisible under adoption/usage metrics alone.

Proof of concept

Select two lighthouse processes; redesign one end-to-end and instrument the other with shallow insertion; deploy shared outcome instrumentation (cost-to-serve, cycle time, error/rework rate) alongside usage telemetry.

Measurement

Outcome metric delta between redesign and insertion arms versus usage/adoption metric delta between the same arms, quantifying the usage-to-value gap.

Source video

Architectural implications

5

Verification is treated as the cost-determining step in agent economics rather than an add-on.

Before

Agents are designed around generation quality; human review is an informal afterthought.

After

Verification is a first-class subsystem: cheap automated checkers, assertions, and bounded human review surfaced as an explicit cost in the agent's design.

Consequence

Task selection, output schemas, and tool interfaces are chosen to minimize verification time, making cheaply checkable outputs a design requirement rather than a nicety.

Source video

Cost is a distribution rather than a per-task constant in agentic workflows.

Before

Cost planning assumes a deterministic unit cost per task and average-based budgeting.

After

Runtime enforces retry caps, step limits, and per-task spend ceilings, with cost variance instrumented per workflow.

Consequence

Budget overruns surface at the workflow level and can be contained by architectural controls (loop bounds, context pruning) rather than by policy memos.

Source video

The refinement loop dominates token spend, so loop topology is the primary cost lever.

Before

Iterative self-repair is an emergent property of the agent loop, unconstrained and unmeasured.

After

Refinement passes are explicitly bounded, and context re-sent per pass is minimized or cached.

Consequence

Cost control shifts from model selection to control-flow design, and the separation between generation and refinement becomes an explicit architectural boundary.

Source video

Tool-call protocol choice (MCP vs direct API) is treated as an architectural cost/variance question.

Before

Tool integration is chosen on convenience and compatibility alone.

After

Protocol and topology choices (single agent vs multi-agent swarm, MCP vs direct API) are evaluated against their token overhead and path variance.

Consequence

Tool integration becomes an economics decision, and multi-agent designs must justify their added token and coordination cost.

Source video

Proprietary context is named as the defensible asset, not the model.

Before

Capability is assumed to live in the model and its API.

After

Context, data ownership, and firm-boundary decisions (build/buy/partner) determine where differentiation resides.

Consequence

Memory, retrieval, and context pipelines become strategically owned components rather than outsourced infrastructure.

Source video

Tradeoffs and failure modes

5

Falling per-token price vs. exploding token consumption

Benefit

Cheaper frontier inference enables more ambitious agentic workflows and coding assistants.

Cost or risk

Total spend rises faster than unit price falls, producing budget overruns for most organizations.

Frontier model inference capability is significantly cheaper per unit, but total cost is driven up by massive token consumption.
Open source video
Source video

Agent autonomy vs. verification overhead

Benefit

Agents execute tasks without continuous human involvement.

Cost or risk

The agent only pays off if success probability exceeds the verification-to-task time ratio; low-probability tasks impose review costs that can exceed manual execution.

An AI agent lifts productivity when its probability of success clears the ratio of verification time to do-it-yourself time.
Open source video
Source video

Generation throughput vs. review capacity

Benefit

Coding agents produce code far faster than humans.

Cost or risk

Generated volume outpaces human review, and the extra iteration drives token consumption upward.

Coding agents write 10x more code than humans, driving up token consumption.
Open source video
Source video

Broad adoption vs. realized value

Benefit

Near-universal adoption of AI tooling spreads familiarity and usage.

Cost or risk

Adoption without workflow redesign yields little measurable EBIT impact, so investment concentrates without return.

89% of enterprises claim to have adopted AI.
Open source video
Source video

Probabilistic agent paths vs. predictable unit economics

Benefit

Agents can handle variable, open-ended tasks without brittle hand-coded branches.

Cost or risk

Cost per identical task can swing up to 30x, undermining budgeting and capacity planning.

Agentic workflows introduce up to 30x cost variance on identical tasks due to probabilistic paths.
Open source video
Source video

Open questions

5

How can cost variance from probabilistic agent paths be bounded without destroying the flexibility that makes agents useful?

Why unresolved

The summary states the 30x variance figure but gives no mechanism for controlling it (retry caps, budget-aware routing, deterministic checkpoints are not discussed).

Research direction

Build and benchmark runtime cost governors (per-task spend ceilings, retry budgets, loop-count limits) and measure variance reduction versus task success rate.

Source video

What task properties predict that an agent will clear the P(success) > verification/do ratio?

Why unresolved

The formula is given as a criterion, but no empirical breakdown by task type or workflow class is provided.

Research direction

Instrument generation time, verification time, and manual time across a portfolio of candidate tasks to produce a task-selectivity model for agent deployment.

Source video

What architectural changes actually reduce the 60% refinement share of token spend?

Why unresolved

The summary quantifies the refinement share but does not attribute it to loop structure, context size, or stopping criteria.

Research direction

Compare loop topologies (single-shot with strong contracts, bounded self-repair, critic-verifier loops) on refinement-token share and end-to-end task success.

Source video

Is the adoption-to-EBIT gap a measurement problem or a genuine structural failure of tool-level deployment?

Why unresolved

The summary presents the 89%/6% contrast without separating attribution difficulties from actual absence of value.

Research direction

Design outcome-attribution instrumentation for agentic workflows to test whether high performers are measuring differently or architecting differently.

Source video

Do multi-agent swarms and MCP-style tool protocols earn their added token and coordination cost versus single agents with direct API calls?

Why unresolved

The summary raises the comparison but reports no outcome or cost result.

Research direction

Benchmark single-agent vs. multi-agent and MCP vs. direct API on identical tasks along success rate, token cost, and path variance.

Source video

Key claims

8
comparativeVerification needed

Only 6% of organizations report meaningful EBIT impact from AI despite 89% claiming adoption.

Evidence

89% of enterprises claim to have adopted AI.

Question

What survey population, definition of adoption, and definition of 'AI high performer' produce the 89% vs 6% figures, and do they replicate?

Source video
factualVerification needed

Token usage is growing at roughly 2,000% CAGR.

Evidence

Token usage has grown by roughly 2,000% CAGR.

Question

Over what measurement window, across which providers or workloads, and how are token units normalized across models?

Source video
factualVerification needed

93% of CIOs report exceeding their AI token budget.

Evidence

93% of CIOs report exceeding their AI token budget.

Question

Is this self-reported budget variance or measured overrun, and what sample and timeframe underlie it?

Source video
causalVerification needed

Agentic workflows show up to 30x cost variance on identical tasks due to probabilistic paths.

Evidence

Agentic workflows introduce up to 30x cost variance on identical tasks due to probabilistic paths.

Question

Which tasks and agents produce the 30x spread, and how much of it traces to retry count versus context growth?

Source video
factualVerification needed

60% of agentic tokens are spent on refinement rather than initial generation.

Evidence

60% of agentic tokens are spent on refinement, not initial generation.

Question

How is 'refinement' defined versus generation, and does the ratio hold across coding, research, and transactional agents?

Source video
causalVerification needed

An agent lifts productivity only when P(success) exceeds verification time divided by do-it-yourself time.

Evidence

An AI agent lifts productivity when its probability of success clears the ratio of verification time to do-it-yourself time.

Question

Does the threshold hold empirically when failure recovery cost and re-run cost are included, not just single-pass verification?

Source video
opinionVerification needed

Organizations must redesign workflows end-to-end rather than insert agents into existing workflows to capture value.

Evidence

Organizations must build agentic capability as a discipline and redesign workflows end-to-end rather than inserting agents into existing ones.

Question

Is there comparative evidence that end-to-end redesigned workflows outperform incremental agent insertion at equal investment?

Source video
comparativeVerification needed

Coding agents write about 10x more code than human developers, driving token consumption.

Evidence

Coding agents write 10x more code than humans, driving up token consumption.

Question

Is the 10x measured in generated lines, merged code, or total output including discarded iterations, and does it net out against human revision cycles?

Source video

Connections

5