AI Accountability in Companies: Who Reviews, Who Decides?

Why companies automate results without redefining review, decision authority and accountability

AI writes proposals, summarizes contracts, prioritizes applications, assesses risks, answers customer inquiries, produces forecasts and recommends decisions.

What once took hours or days is now available within minutes. The visible productivity gain is real. Results are produced faster, processing times fall and the same organization can generate significantly more output.

And yet, the actual work has not disappeared.

Someone still has to determine whether the result is technically correct, commercially sound, legally permissible and appropriate within the specific context. Someone has to decide whether action should follow. And someone bears the consequences when the AI has overlooked an important connection, processed incorrect data or presented a plausible answer with the authority of a fact.

This is precisely where a new contradiction—and therefore a new corporate risk—emerges.

Contradiction of the Week

Companies are automating the generation of results faster than their review. At the same time, they expect people to assume accountability for those results.

AI has been integrated into the business process, yet accountability remains attached to the old position and its old job description. Between the two lies an organizational gap that appears in no productivity calculation.

AI accountability in the enterprise does not begin with the question of legal liability. It begins with how review, influence and decision authority are distributed throughout the daily business process.

Generation Is Accelerated. Review Remains Human.

With AI, an employee no longer produces three contract summaries in an hour, but thirty. A manager no longer receives five decision papers, but fifty. Customer service no longer answers one hundred inquiries, but one thousand. The procurement system no longer examines individual suppliers; it scans the entire supplier base for deviations.

On a dashboard, this looks like productivity.

Within the organization, however, it creates a new challenge—a second calculation. The more results are generated automatically, the more results must be classified, reviewed, approved, corrected or, in exceptional cases, stopped. The cost of producing output falls. The cost shifts into reviewing that output.

This shift is frequently underestimated or ignored entirely. Companies measure how much faster a result is produced. They measure less often how much additional attention and working time are required to assess its quality. Even less visible is the expertise that disappears when people spend their entire day reviewing proposals whose creation they cannot fully trace.

The employee becomes a reviewer of machine-generated results. Yet the role, position, job description, time allocation and decision authority remain unchanged.

The employee is expected to assume accountability for something that is neither fully understood nor entirely traceable. At the same time, that person may lack the authority to stop the AI application, alter the business process or reject an unrealistic requirement.

This is not a question of individual diligence. It is a question of organizational architecture.

The Numbers Do Not Show Too Little Adoption. They Show Too Little Order.

The introduction of AI is no longer a cautious experiment. According to the McKinsey Global Survey 2025, almost nine out of ten surveyed organizations use AI regularly. Sixty-two per cent are already experimenting with AI agents or scaling them. At the same time, around two-thirds of companies have not yet reached enterprise-wide scale. Only 39 per cent report any impact on EBIT, and for most of those companies, the impact remains below five per cent.

The decisive finding is less visible. The companies achieving the greatest economic impact are far more likely to redesign their business processes fundamentally. They also define when results require human review. According to McKinsey, both are among the strongest distinguishing characteristics of leading companies.

AI alone does not create new organizational capability. It first accelerates whatever it is applied to—and the purpose for which it is applied.

When the business process is clearly defined, acceleration can create economic value. When accountability, data, handovers and decision authority are already unclear, AI scales that lack of clarity.

A poor handover does not improve when it is automated. It becomes faster.

An unclear decision does not become more definitive when the decision paper is produced within seconds.
There are simply more decision papers.

And unresolved accountability is not settled when a person’s name appears in an approval field at the end of an automated process.

People Review Results Whose Creation They Cannot See

The global study “Trust, attitudes and use of Artificial Intelligence 2025” by the University of Melbourne and KPMG is based on more than 48,000 respondents across 47 countries. Fifty-eight per cent of employees knowingly use AI at work. Sixty-six per cent rely on AI output without assessing its accuracy. Fifty-six per cent report having made mistakes in their work as a result of AI. Fifty-seven per cent conceal their AI use and present AI-generated work as their own.

Only 47 per cent say they have received training. Just 40 per cent report having policies or guidance for using generative AI in the workplace.

These figures are quickly interpreted as a question of employee competence or behavior. That interpretation falls short.

When AI use happens informally, quality requirements remain unclear and people simultaneously face greater pressure for speed and output, the resulting behavior is entirely predictable: results are adopted, reviewed only superficially or passed on without disclosing their origin.

The employee is not choosing between diligence and negligence in a neutral environment. The employee is reacting to the reality of the work: performance targets, available time, expectations of higher productivity, uncertainty about when AI may be used and an organization that already measures speed while barely making the quality of review visible.

What appears to be individual misconduct may be the logical consequence of contradictory expectations.

“Human in the Loop” — or the Illusion of Control

Many companies answer the accountability question with a reassuring statement: a human reviews the result at the end.

Yet the presence of a human does not constitute effective control.

A review has commercial substance and relevance only when the reviewer understands the context, receives sufficient time, knows the quality criteria, can identify deviations and has the actual authority to reject the result or stop the process.

When any one of these conditions is absent, “human in the loop” becomes a formal signature beneath a machine-accelerated business process.

The human is declared the final control point without being equipped to act as one. The person sees the result, but not necessarily the sources used. The person is expected to approve a recommendation, but does not know every limitation of the model. The person bears accountability, yet cannot change or overrule the system logic. And as volume rises, approval increasingly becomes routine. The AI probably knows what it is doing.

Gartner describes this issue with considerable precision in relation to AI agents. In systems that act following human approval, control loses its effectiveness under time pressure or through approval fatigue. With autonomous agents, the scope and speed of their actions can outpace human oversight. Accountability for the result nevertheless remains with the company. Gartner expects that by 2027, around 40 per cent of companies will downgrade or deactivate autonomous agents after discovering governance gaps only through real-world incidents. Gartner: AI agent governance, 2026 Gartner: AI agent governance, 2026

The decisive dividing line therefore does not run between automated and human-reviewed.

It runs between visible control and effective control.

Accountability Without Influence Is Not Accountability

Accountability requires influence.

Anyone expected to stand behind a decision must be able to understand what it is based on. That person must be permitted to deviate, must have access to the decisive information and must know where their accountability begins and where it ends.

AI changes precisely this relationship.

A system produces the analysis. An employee reviews it. A manager decides on its basis. IT operates the application. An external provider is responsible for the model. Compliance defines the boundaries. The relevant business function is accountable for the commercial result.

When something goes wrong, every involved party can plausibly explain why the cause lay outside its own area of responsibility.

IT provided the system from a technical perspective. The provider disclosed the limitations. The employee may have reviewed parts of the output. The manager made the decision. And Compliance had issued a policy covering all of it.

And yet no one carried complete accountability for how these pieces of the puzzle worked together.

The cause does not lie in one individual part of the chain. It lies in the connection between the parts within the business process.

AI Encounters an Already Weak Decision Architecture

The Deloitte Global Human Capital Trends Report 2026 reveals the scale of this gap. Sixty per cent of executives regularly use AI to inform their decisions. Sixty-four per cent consider AI-assisted decision-making highly important to their current business situation. Only five per cent regard their organizations as leading in its organizational application.

More importantly, an underlying Deloitte analysis of 574 executives classified 57 per cent of the organizations examined as having a low level of organizational decision capability. Decision authority, inputs, quality standards and escalation paths are therefore often insufficiently defined even without AI.

AI does not enter a “neutral” organization. It encounters existing power structures, unclear responsibilities, function-specific metrics and historically established approval practices. It is inserted into this structure and amplifies its logic.

If it was previously unclear who had the authority to decide, another participant now enters the process—one that generates recommendations or prepares actions. If accountability for data quality was previously unclear, AI now processes that data on a far greater scale. If it was previously unclear when a case required escalation, the unresolved case simply reaches the next function faster.

The technology may be new. The company’s structural challenge is not.

Automation Changes the Business Process—Even When It Formally Remains the Same

A company can introduce AI into an existing business process without changing a single box in the process diagram. On paper, everything remains the same: create, review, approve, execute.

In reality, the process has changed fundamentally.

AI increases the volume of results—the output—and radically shortens the time between creation and decision. The reviewer responsible for checking specific results becomes the bottleneck. Exceptions requiring review suddenly accumulate elsewhere. Expertise shifts from creation to assessment, possibly into a function that does not possess the same depth of knowledge. A single error can spread across hundreds of cases within a short period. And the person at the end of the process retains formally identical accountability for a result over whose creation they have increasingly less influence.

This can be seen in almost every function:

• Finance: AI produces variance analyses and commentary. Who checks whether a plausible explanation identifies the actual driver of the result, rather than merely producing a statistically probable statement?
• Procurement: AI assesses suppliers, contracts and risks. Who recognizes when a formal risk score fails to capture the strategic dependency on a single supplier?
• HR: AI sorts applications and drafts performance reviews. Who is accountable for the criteria through which human uniqueness is translated into standardized categories?
• Customer Service: AI formulates commitments in real time. Who has the authority to restrict speed in favour of an accurate and sustainable response?
• Leadership: AI analyzes and condenses data into decision papers. Who ensures that this condensation does not remove precisely the contradictory signals a leader needs to see?

AI does not merely change one task within a process. It changes volume, speed, information density, dependencies and the distribution of influence across the entire process.

Anyone who merely introduces AI as an application has not thought far enough.

The Accountability and Decision Chain Must Be Clear Before Automation

The question is not simply whether a human or an AI makes the decision. That distinction is too broad and too superficial.

An organization must distinguish what AI actually does within a specific business process:

• Does it produce a draft?
• Does it analyze and condense information?
• Does it issue a recommendation?
• Does it prioritize?
• Does it prepare a decision?
• Does it trigger an action following approval?
• Does it act independently within defined boundaries?

Each of these stages changes the risk and therefore demands a different structure of review, intervention and accountability.

For high-risk systems, the EU AI Act likewise requires companies to assign human oversight to a person who is genuinely equipped and capable of exercising it. Monitoring, responding to risks and enabling human intervention are not abstract governance concepts. They must become practically effective through roles and business processes. European Commission: Navigating the AI Act

A name in an accountability matrix is not enough.

The organization needs clarity on who defines quality criteria, who reviews results, which cases must be escalated, who may overrule an AI recommendation, who can stop the automated process and who bears commercial accountability for the consequences of an incorrect decision.

Only then does accountability regain a real connection to influence.

The CEO Question Is Not: How Quickly Can We Introduce AI?

The pressure for speed is understandable. In the IBM CEO Study 2025, 64 per cent of the 2,000 CEOs surveyed said they invest in technologies before their commercial value is fully established, driven by fear of falling behind. Only 25 per cent of AI initiatives had delivered the expected return on investment, and just 16 per cent had been scaled across the enterprise.

Speed is not proof of organisational maturity. It may equally be an expression of unresolved urgency.

The central task at the top of the company is not to introduce AI into every function as quickly as possible. It is to reorder the system—the organizational architecture—in which AI is expected to create positive results.

Leadership, decision authority, business processes and accountability require the same level of attention as data, models and technical integration. Not sequentially. As one connected architecture.

Otherwise, AI produces results within seconds while the organization continues to decide in days, escalate in weeks and discover only after several months that no one carried complete accountability.

The Unavoidable Consequence

AI can accelerate work. It can condense information, identify patterns and prepare decisions. It can perform clearly defined actions within controlled boundaries.

It cannot create order where accountability remains unresolved.

If a human is expected to stand behind an AI-generated result, that person needs more than an approval button. They need context, time, professional judgement, access to the decisive information and the actual authority to say no.

Without that authority, the human does not carry accountability.

They carry the risk created by an architecture that has automated speed and merely assigned accountability.

Who in your company can actually stop an AI-driven decision—and does that person also carry accountability for the result?