AI Transformation Is a Power Redesign

AI decision rights are becoming one of the most important questions in AI transformation. Major advisory firms have published frameworks for redesigning organisations around AI. Each one addresses who owns which decisions, how authority moves between humans and agents, and how oversight should work. 

But none of them names what they are really helping organisations do. 

Redesigning decision rights means redesigning power. It determines who can act, whose judgment carries weight, whose override matters, and who is accountable when an AI system gets something wrong. AI transformation is forcing organisations to answer these questions again—often for the first time in a generation. The frameworks used to answer them treat power as a neutral design variable. It is not.

What’s the Right Question on Decision Rights?

The standard AI decision-rights framework asks who should own each decision in the new operating model, then maps roles, oversight, and escalation paths. That work is necessary, but it starts from the wrong question. 

The better question is whether the model puts authority where the best judgment actually sits. 

The first question assumes today’s authority structure is broadly right and translates it into the AI era. The second treats AI redesign as a chance to test whether authority sits with the people who hold the most relevant knowledge and judgment, or simply with those who inherited it through seniority and the old operating model.

MIT Sloan’s research on intelligent choice architectures explains why this is structural. As AI systems begin shaping decision environments - what options appear, which trade-offs surface, and what information reaches human judgment - the authority to design those environments becomes one of the organisation’s most consequential powers. MIT calls this meta-decision rights: the power to shape the system within which everyone else decides. 

In many AI transformations, those rights are being allocated by the same senior groups that held authority before. The process producing the new architecture is the process that produced the old one. The system is being built by people furthest from the work it governs.

The Commercial Cost

MIT Project NANDA’s July 2025 research (300 disclosed AI initiatives, 150 leadership interviews, and 350 employee surveys) found that despite US$30–40 billion in enterprise generative AI spending, 95% of organisations show no measurable business return. Two composite scenarios, drawn from documented failure patterns, show why.

A large insurer automates first-line claims triage. The model handles standard cases well but fails on non-standard ones: irregular income documentation, losses spanning coverage categories, or documents that do not match expected formats. These route to exception handling. But the experienced adjusters who knew the informal workarounds are no longer in the workflow; the headcount reduction that justified the investment has already happened. 

Complaints rise. Remediation costs follow. A regulatory enquiry opens. The knowledge that would have prevented this was inside the organisation, held by the adjusters. It was not treated as a design input because the designers did not know it existed, and the adjusters lacked the authority to be asked.

A healthcare network deploys AI scheduling across its nursing workforce. The system optimises visible variables: headcount, cost, and award compliance. It cannot see what experienced coordinators know: which nurses work well together in high-acuity situations, or which wards need experienced staff on particular shifts. Patient safety incidents rise on two wards. Experienced nurses leave, citing schedule inflexibility. 

The World Economic Forum’s 2025 analysis identifies this failure mode: algorithmic scheduling can improve predictability, but if poorly designed it can lock workers into rigid patterns or penalise them unfairly. The local knowledge of what “poorly designed” means was held by the people the system replaced.

How Inclusion Improves AI Transformation

Liberty Mutual’s LibertyGPT deployment shows the alternative. The system was built to enhance new claims adjusters’ judgment by giving them access to experienced colleagues’ accumulated knowledge. It saved more than 200,000 person-hours in its first year. 

The design question was not “how do we replace what they do?” It was “what do experienced adjusters know that new ones lack, and how do we make it available?” Authority followed knowledge. 

UKG’s global survey of 8,200 frontline workers found that employees using AI report lower burnout - 41% versus 54% among those not using it. BCG’s AI at Work survey found frontline adoption stalled at 51%, while leaders exceed 75%. That gap is the difference between AI designed with workers and AI designed around them.

Six conditions of inclusion directly shape AI decision-rights performance:

• Contribution: the knowledge existed, but the conditions for contributing it to the design did not.

• Decisions: meta-decision rights determine whose inputs shape the decision environment, and whose are absent.

• Trust: automation bias is partly a trust failure. Overrides matter only when people feel able to use them.

• Inclusive Leadership: leaders who share meta-decision rights build different architectures; leaders who consolidate them reproduce the hierarchy.

• Norms: if the unwritten rule is that challenge only comes from inside the project team, predictable blind spots remain.

• Systems: override mechanisms, escalation paths, and exception handling either reinforce genuine human judgment or merely simulate it.

The Hard Question for AI Operating Model Design

The headcount imperative is real. For many organisations, AI transformation is primarily a cost-reduction exercise. Arguing that workers should help design systems that may replace them is awkward. It is also unavoidable. 

Capturing workers’ knowledge before eliminating their roles is not a solution to the power problem; it is a more efficient version of it. The people are gone. Their knowledge remains in the system. Their authority does not. 

The honest inclusion argument is this: organisations serious about performance must ask which automated roles contain judgment the model cannot replicate, and what different role those people should occupy in the new architecture.

Frenkel’s 2026 peer-reviewed research in Sociology Compass shows what happens when this calculation fails: reduced job discretion under algorithmic management increases stress, dissatisfaction, and turnover. The savings and the cost are real.

Advisory frameworks treat AI decision-rights redesign as governance and efficiency work. It is also power redistribution. Organisations that fail to ask where power is going, and whether it is going where judgment sits, are building a faster version of an architecture that was already underperforming. The 95% failure rate is the measure of that.

The Bottom Line

Two questions are worth putting to the leadership team before the AI operating model is finalised.

First: do the AI decision rights this architecture creates reflect where the best judgment sits, or where authority already was? If it is the latter, the redesign is a faster hierarchy, and its gaps will compound at machine speed.

Second: for roles being automated or restructured, has anyone mapped the difference between the procedure being automated and the judgment the role also contained? The procedure may be replaceable. The judgment often is not.

TROY RODERICK & PARTNERS works at this intersection. Our CONDITIONS framework and “Power Play” program assess and strengthen the conditions that determine whether the knowledge, judgment, and perspectives needed for effective AI governance and AI architecture can surface, and whether the design process is producing a genuinely different architecture or a faster version of the old one.

If you are reviewing AI decision rights, AI governance, or your organisation’s AI operating model and want to know whether your architecture is building those conditions or undermining them, that is the work we do.

SOURCES & FURTHER READING

• BCG. 'AI at Work 2025'

• CAIC. 'AI in Claims Operations: A 2026 Playbook'

• Challapally A et al. 'The GenAI Divide.' MIT Project NANDA (July 2025)

• Deloitte. '2026 Human Capital Trends: Decision-Making with AI'

• Frenkel S. 'New Technology and Work.' Sociology Compass (2026)

• Kawakami A et al. 'AI Failure Loops in Devalued Work.' Sage Journals (2026)

• Liberty Mutual LibertyGPT: cited in Schrage & Kiron, MIT Sloan Management Review (2025)

• Schrage M, Kiron D. 'The Great Power Shift.' MIT Sloan Management Review / TCS (2025)

• SHRM. 'Navigating AI in the Workplace 2026'

• UKG / Workplace Intelligence. 'More Perspectives from the Frontline Workforce' (8,200 workers, 10 countries, 2025)

• World Economic Forum. 'Beyond the Desk: How AI Is Transforming the Frontline Workforce' (October 2025)