Running Before Crawling

McKinsey's 2025 survey found that 88% of respondents reported regular AI use in at least one business function, yet nearly two-thirds said their organizations had not begun scaling AI across the enterprise. Only 39% reported any enterprise-level EBIT impact. Organizations described as high performers were nearly three times as likely to have fundamentally redesigned workflows and were more likely to define when human validation was required.^3.03

The Wall Street Journal reported that only 26% of companies in a KPMG survey had a comprehensive view of AI costs. KPMG also said it was working with companies that had exhausted annual token and cloud-computing budgets in a matter of months.^3.04

A demo is a first date. Production is the joint checking account.

These numbers do not prove that AI lacks value. They show that usage can become common before workflow redesign, cost visibility, and verified enterprise return become common.^4.03

Gartner forecast that more than 40% of agentic-AI projects would be canceled by the end of 2027 because of escalating costs, unclear business value, or inadequate risk controls. That is a forecast, not a measured failure rate. It is still a useful warning about scaling an architecture before proving the work.^4.04

UBER: The meter moved faster than the value link.^5.03

The Verge reported that Uber had exhausted its annual AI budget four months into 2026. President and COO Andrew Macdonald said the connection between rising Claude Code token consumption and useful consumer features was not yet visible, and that token cost would need to be weighed against headcount. His candor is valuable. My question is why that line of sight was not part of the approval logic before the budget was reportedly gone.

KLARNA: The cost objective became the correction.^5.04

Klarna company-reported that its assistant performed work equivalent to 700 full-time agents. Reuters later reported that CEO Sebastian Siemiatkowski said the company had over-indexed on AI cost cutting, was course correcting, and had returned to hiring. This is not hypocrisy. It is a public correction from fewer people toward better products and service.

ZILLOW: Forecast confidence reached the balance sheet.^5.05

Zillow shut Zillow Offers, announced a 25% workforce reduction, and held or listed about 7,000 homes representing roughly $2.8 billion in inventory. CEO Rich Barton said forecasting unpredictability exceeded expectations. WIRED reported that the business depended on forecasts three to six months ahead and that third-quarter sales were 5% to 7% below forecast. Zillow was not generative AI. It belongs here because algorithmic confidence became balance-sheet exposure.

1. What verified business outcome are we buying?
2. What is the full cost per verified outcome after review and correction?
3. Which standard paths and exceptions are inside the scope?
4. What is the authoritative source of truth, and who owns it?
5. Where must a human approve, override, or escalate?
6. How does cost behave as usage, context, and integrations grow?
7. What happens when the model is wrong in front of a customer?
8. How do we pause, roll back, switch vendors, or shut the system down?

The Wall Street Journal reported that token-based pricing is creating new budgeting challenges even for experienced finance teams. NIST recommends empirically validating capability claims, testing before deployment, managing third-party components, documenting overrides, and monitoring systems in real-world operation.^7.03

This is why payroll can appear to fall while total cost merely changes categories. A visible salary becomes a less visible mix of cloud consumption, APIs, engineering, consultants, evaluation, security, rework, and the remaining people who still have to catch the mistakes.

The agent never sleeps. Neither does the meter.

In a field study of 5,179 support agents, access to a generative-AI assistant increased productivity by 14% on average and by 34% among novice and lower-skilled workers, with minimal impact among experienced and highly skilled workers. The authors found suggestive evidence that the system spread the practices of stronger workers.^8.03

In an experiment involving 758 knowledge workers, participants using AI completed 12.2% more tasks and worked 25.1% faster on tasks inside the tested capability frontier. On a task outside that frontier, AI users were 19% less likely to produce the correct solution.^8.04

NIST recommends evaluating capability claims with empirically validated methods, sharing pre-deployment testing with release authorities, documenting the use of domain knowledge, involving operators in prototyping and testing, monitoring performance in practical settings, and managing third-party dependencies and fallbacks.^10.03

A leader does not need to fine-tune a model. A leader does need to know what the system is allowed to see, what it is allowed to do, who corrects it, how that correction is learned, and what the verified outcome costs.

CRAWL: Document the work.

Map the workflow, exceptions, source of truth, baseline cost, cycle time, quality, escalation, and customer impact. Exit only when the real work can be explained without a demo.

WALK: Assist the specialist.

Use bounded tasks, capture corrections and reasons, compare models and non-AI alternatives, and measure time saved after review. Exit only when the evidence shows leverage rather than hidden rework.

RUN: Automate proven paths.

Automate high-volume, low-ambiguity work first. Add confidence thresholds, approval, monitoring, rollback, and escalation. Exit only when defects and exceptions remain inside agreed limits.

EARN: Scale after the economics survive contact with reality.

Expand only when full cost per verified outcome beats the baseline and leadership has stated how created capacity will be reinvested or changed.

McKinsey's high performers were nearly three times as likely to redesign workflows and more likely to define human-validation processes. The sequence above turns those practices into an operating gate.^11.07

ACCOUNTABILITY RECEIPT: Air Canada argued the chatbot was not the company. The tribunal disagreed.^12.03

In Moffatt v. Air Canada, the airline argued that it could not be held liable for information provided by its chatbot. The British Columbia Civil Resolution Tribunal called the suggestion that the chatbot was a separate legal entity a remarkable submission, found that the chatbot was part of Air Canada's website, found negligent misrepresentation, and ordered Air Canada to pay $812.02. The decision is jurisdiction-specific, but the operating lesson travels: the customer experiences one company.

Gartner's forecast of widespread agentic-project cancellations is one more reason to require explicit value, risk controls, and a stop path before scale.^12.04

AUTHORSHIP AND AI DISCLOSURE

This is my work: the thesis, thinking, writing, judgment, original frameworks, creative direction, and management perspective are mine. I used AI as a research and production tool to organize sources, pressure-test claims, refine language, develop visualizations, and strengthen the digital presentation. I directed the work, made the editorial and analytical decisions, reviewed the evidence, and approved the final publication. AI accelerated and strengthened the process; it did not originate, own, or authorize the point of view.

TRUTH AND CITATION METHOD

Every manuscript block was classified as external fact, author inference, author analysis, author experience, original framework, hypothetical, disclosure, or call to action. Every external factual block carries a footnote. Figures containing external facts carry a source footnote. Company-reported figures are labeled. Gartner's number is labeled as a forecast. Zillow is labeled as an algorithmic precedent rather than a generative-AI case. The Air Canada decision is labeled as jurisdiction-specific.

Research was checked against the cited sources through July 11, 2026. The detailed paragraph truth audit and source ledger are included in the publication package. This paper is an operating perspective, not legal, financial, employment, cybersecurity, or regulatory advice.

ABOUT SUDEEP ARYA

Sudeep Arya is a commerce transformation operator with more than 20 years of experience connecting ecommerce strategy, platform delivery, Amazon, DTC, marketplaces, retail media, analytics, customer experience, product data, integrations, and operating-model change. His background includes work across Burt's Bees Baby, Mars, Randa Accessories, David Yurman, PVH, and Kate Spade, along with independent advisory work for emerging consumer brands.^14.07