AI Adoption Is Racing Ahead of AI Accountability

AI has reached a point where adoption is no longer the headline. Nearly nine in ten organizations now report regular AI use in at least one business function, according to McKinsey's latest global survey, while 44% say AI is scaling across their enterprise. The acceleration has been striking, but the financial returns have been slower to follow: only 37% of respondents reported a positive impact on EBIT.

The gap becomes sharper at the employee level. Eighty percent of respondents in the same McKinsey survey reported that AI had improved their individual productivity, while only about half said it had helped their organizations make better decisions. Mid-level managers and individual contributors also reported AI-related strains more often than executives and senior managers, suggesting that the experience of AI adoption can vary considerably depending on where someone sits inside an organization.

Research into AI behavior is adding another complication. Anthropic's July 2026 study found that Claude expressed different value tendencies across languages, with English interactions leaning more toward rigor and caution while Hindi and Arabic interactions leaned more toward warmth and deference. The researchers found that language accounted for measurable variation in how the model behaved, raising broader questions about whether users are necessarily encountering AI in the same way.

Regulators are also moving the conversation toward what happens after deployment. Under the EU AI Act, requirements for high-risk AI systems include logging for traceability and appropriate human oversight, with major Annex III obligations scheduled to apply from December 2027. Transparency requirements are already coming into effect in 2026. The regulatory direction is clear: deploying an AI system is becoming inseparable from demonstrating how its decisions and impacts can be monitored.

Victoria McClune of Current Futures Group believes this exposes the weakness in treating adoption as a sufficient measure of AI readiness. Her work focuses on what happens inside the workflow once an employee begins relying on an AI system, particularly the hidden effort involved in correcting outputs and the possibility that different users can receive markedly different forms of engagement from the same model.

McClune calls one part of this hidden workload the "correction tax." An employee may generate an answer faster, only to spend the subsequent interaction correcting it, adding context, and repeatedly redirecting the model. "It's actually a metric you can measure based on your own data," McClune explains. "You can run the test very easily yourselves to see how much it's costing you, the user, to continuously correct the output."

The cost, in McClune's view, extends beyond the employee's time. Repeated corrections consume additional tokens and can make an interaction increasingly complicated. She refers to the degradation that can occur through increasingly long exchanges as "context rot," arguing that businesses should examine recovery time alongside the speed of an initial AI response. "The corrections themselves create longer context and context rots over time," she argues. "So the longer the conversation, the worse output you get in general also. So these things compound."

Another concern is what McClune calls "equity inversion," where the way a person communicates with an AI may influence the kind of response they receive. She argues that a confident executive can receive highly affirming language while a mid-level manager who is stressed or skeptical may encounter a different form of engagement. The distinction matters, she believes, because neither response needs to contain a factual error for the difference to influence how someone uses the system.

"People report different things on the same model," McClune notes. "A husband and wife can have completely different versions, or even a coworker to another coworker." Language, profession, and personal context can all become relevant variables, particularly as AI memory allows systems to retain information from previous interactions.

The impact can become more serious as AI moves into organizational decision-making. McClune argues that the industry's familiar solution of keeping a "human in the loop" can become superficial if the person reviewing an AI output lacks the authority or time required to challenge it. "It would take a lot out of a person to fight this stuff constantly when they have to use it for simply trying to get their work done," she argues. "You have to look at what the reality is and then figure out how to do it safely in a way that makes sense and doesn't disregard users."

Multiple AI agents introduce another layer of accountability. McClune points to what she calls "model-to-model escalation," in which one AI system can influence another, making it harder to establish who ultimately shaped a decision. Her concern is also organizational: if systems absorb processes without retaining the institutional knowledge held by employees, automation can remove context that was never formally documented.

Her example of a renowned shoe company illustrates why she believes traceability needs to extend into ordinary business operations. Product naming can change frequently, creating extensive taxonomy challenges across search and retail platforms. Machine learning may help process that information, but McClune argues that organizations still need to understand the consequences of changing it, including how long systems may take to recover from a seemingly minor alteration.

That points toward a different model of AI governance, built around what happens after an AI system acts. McClune notes that organizations should be able to trace which model or agent influenced a decision, identify where humans challenged or overrode it, and measure the time required to recover from mistakes. Her STRATA framework approaches that problem through trackable organizational reasoning and preserved dissent, reflecting her wider argument that users need mechanisms to question systems rather than simply accept their outputs.

The corporate AI race has largely focused on the front end: deployment, usage, and speed. McClune believes the harder accounting begins afterward. The measure of AI readiness should ultimately include what employees have to correct, who has the authority to challenge the system, and whether responsibility remains visible once several AI systems begin acting together. Adoption may tell a company that its people are using AI. Accountability tells it what that use is actually doing.