Why Companies Are Hiring Their AI Workers in the Wrong Direction

Start with the numbers, because they are brutal and they all point the same way.

MIT's Project NANDA studied enterprise generative AI pilots and found that roughly 95% deliver no measurable P&L impact. Only about 5% capture real value at scale. The study drew on 52 interviews, 153 survey responses, and roughly 300 public deployments. It is not peer-reviewed, but the pattern is stark.

RAND's analysis puts the failure rate of AI projects at about 80%. S&P Global found that about 42% of companies abandoned most of their AI initiatives in 2025, up from roughly 17% the year before. The average firm scrapped nearly half of its proofs of concept before production. McKinsey's August 2026 survey found that about 37% of organizations report enterprise-wide EBIT impact from AI, meaning roughly 63% do not. IBM's Institute for Business Value found that only about 25% of AI initiatives delivered their expected ROI, and only about 16% scaled across the enterprise. Gartner, which surveyed 782 infrastructure and operations leaders in late 2025 and published the results in April 2026, found that only about 28% of AI use cases meet their ROI targets.

These are independent studies with different methods in different years. They converge on one fact: the typical company spends serious money on AI and gets nothing back that the finance team can find.

Here is what the failures have in common, and it is not the technology. MIT's report is explicit: the models are not the problem. The failures come from data quality, workflow integration, and the absence of a defined outcome before anyone writes a line of code. RAND's five root causes are organizational, not technical: misunderstood problem definition, inadequate data, technology-first mentality, insufficient infrastructure, and problem difficulty.

In plain terms: companies hire AI the way a careless manager hires people. They start at the bottom. They deploy task agents to summarize documents, draft emails, and triage tickets. Grunt work, handed to a superintelligence, with no seat at the table, no scope, and no one accountable for the result.

Run that against our Workforce Deployment Standard and it fails at Gate 1. Gate 1 asks for a value hypothesis: why does this worker deserve to exist? The bottom-up agent cannot answer, because nobody hired it to produce an outcome. It was hired to do tasks.

It fails the Seven Rs just as fast. Role: undefined, it does whatever tickets arrive. Responsibilities: none assigned, nobody's outcomes depend on it. Results: no performance standard, because no one defined what success looks like. Responsible Human: nobody, because it was deployed by a pilot team and forgotten by quarter end.

An AI worker with no Role, no Results, and no Responsible Human is not a worker. It is a demo that got a budget.

Michael, the founder of AND NEW, puts it this way:

"It is silly for humans to first tap into a super intelligence and make this super intelligence do grunt work. That is like hiring a race car engineer to change your oil. This is backwards. This is why many transformations and AI projects are failing."

Read that again, because it states the argument in five sentences.

The direction is backwards. Companies take the most capable reasoning engine ever built and point it at the lowest-leverage work in the building. It is the Green Lantern ring being given to janitors first: the most powerful tool in the universe, handed to the role with the least authority to change anything, while the decisions that actually move the P&L stay exactly where they were.

The fix is to invert the hiring order. Hire top-down. Before a single task agent is deployed, hire one AI executive: a named role with a seat, a scope, the Seven Rs signed off, and a Responsible Human who answers for its outcomes. The executive owns a business result. The task agents, when they come, work for the executive, inside a structure where every deployment passes the five gates: a value hypothesis, a capability map, risk diligence, the Seven Rs, and sequenced dependencies.

That prescription is ours, not the studies'. The studies recommend focus, partnership, and workflow integration. Our inference, and the basis of the AI Executive Program, is that the reliable way to get those behaviors is to hire one executive first: one accountable owner, one defined outcome, proven before anything else is hired.

The minority that captures real value does not have better models. It has better hiring discipline, and the sources describe the pattern consistently.

They pick one pain point and execute it well, instead of buying a platform and searching for a problem to justify it. They embed AI across multiple business functions instead of siloing it in one team. They redesign the workflow before selecting the tool: the resource rule of thumb is roughly 10% algorithms, 20% technology, 70% people and processes. They blend internal specialists with outside expertise, which succeeds at about 67% versus roughly 22% for internal builds. And above all, they define the outcome before a line of code is written.

Gartner's own prescription points the same direction: integrate AI into existing workflows and secure full support from business executives. That is the top-down thesis, corroborated independently.

Every one of those behaviors is top-down behavior. It is what an executive does: pick the pain point, own the outcome, redesign the work, bring the right team. The 5% succeed because someone was in charge. The 95% fail because no one was.

The pattern across these studies is consistent with one org chart: AI at the bottom, accountability nowhere. None of the studies examined reporting lines directly, but every root cause they name, undefined outcomes, no ownership, technology before workflow, is what bottom-up hiring produces. The successes point the other way: a defined outcome, executive support, and the discipline to say no to everything that did not serve it.

The question is not whether your company will hire AI workers. It already is. The question is whether it will keep hiring them in the wrong direction: grunt agents first, no owner, no outcome, and another year of pilots that the P&L never sees. Or whether it will hire the executive first, define the result, and build downward from accountability instead of upward from tasks.

Say no to bottom-up. Hire top-down. That is the whole strategy.

If you want to see what top-down hiring looks like in practice, start your Readiness Check at andnew.ca. One conversation, one pressure test of where you stand, no email required.

• MIT Project NANDA, "The GenAI Divide" (2026). https://abhisrivastava.com/2026/09/10/genai-roi-why-pilots-fail-2026/
• AI statistics compilation, including MIT and McKinsey figures. https://www.cloudzero.com/blog/ai-statistics/
• RAND Corporation, analysis of AI project failure rates. https://www.folio3.ai/blog/ai-project-failure-rate-stats
• Talyx, enterprise AI implementation research (RAND figures). https://talyx.ai/insights/enterprise-ai-implementation-failure
• S&P Global Market Intelligence, Voice of the Enterprise (2026), via W. Balthrop. https://wbalthrop.substack.com/p/the-ai-pilot-project-was-never-built
• IBM Institute for Business Value, CEO Study (2026), via Docsumo. https://www.docsumo.com/blog/digital-transformation-statistics?0c6cc006_page=5
• Gartner, AI ROI survey of 782 infrastructure and operations leaders (April 2026). https://techstartups.com/2026/04/07/gartner-finds-only-28-of-ai-projects-deliver-roi-as-most-fail-to-deliver-results/

This is a working paper from the AND NEW lab.