What Happens at Work When the AI Goes Away?

A new framework for thinking about artificial intelligence at work starts with a question companies have spent relatively little time asking.

What happens if the AI goes away?

Companies are racing to build generative AI into everyday work, from drafting and data analysis to scheduling and decision support. Vedant Das Swain of NYU Tandon and Koustuv Saha of the University of Illinois Urbana-Champaign argue that organizations should also prepare for the opposite possibility: a budget cut, outage, privacy rule, regulatory action or vendor dispute that suddenly makes AI unavailable.

The team describes their Counterfactual Resilience Framework, or CReF, in a Visionary Paper that they presented at the inaugural ACM AI Leadership Summit, which ran August 30 to September 2 in Atlanta. According to Das Swain, the paper is meant to inspire transformative research agendas and identify grand challenges from the engagement of the entire AI ecosystem.

CReF proposes a way for organizations and researchers to expose dependencies that may be difficult to see while AI is working normally. The framework asks users to construct a "counterfactual anchor" with three parts: a workplace setting in which AI is being used, an event that removes access to it, and the aftermath.

The idea is to treat AI's absence as a kind of stress test.

"A lot of research on human-AI teaming and complementarity assumes that LLMs and generative AI tools are like oxygen,” Das Swain said. “ It will be abundant, uniformly accessible, and continually replenished. We wanted to challenge this assumption by urging organizations to ask how they would look if AI was suddenly snatched away. Our framework helps stakeholders confront who looks productive, who retains expertise or whether basic work can continue in the face of a technological barrier, outage, or attack. Existing AI productivity frameworks excite us by focusing on the best case. Instead, we shift attention to pragmatically planning for the worst case."

The paper explores three potential vulnerabilities: whether AI benefits are distributed unevenly across jobs; whether prolonged AI use could erode skills workers need to operate independently; and whether organizations can recover when an AI-dependent workflow is disrupted.

To illustrate them, the authors construct fictional workplaces.

In one, Alex manages complicated relationships and gets relatively little benefit from AI, while Sam works on a data analytics team whose output increases after adopting it. Management begins interpreting their differing output through an AI-influenced notion of productivity. When rising costs temporarily cut off the organization's AI access, Alex's work changes little. The scenario asks whether AI can distort performance expectations when its usefulness differs dramatically among jobs.

Another scenario involves Tanya, a junior compliance analyst who has spent two years reviewing AI-drafted risk assessments rather than regularly writing them herself. When an audit prompts her employer to suspend the system, she struggles to return to manual drafting. The example is fictional, but the concern is not new. The authors connect it to decades of research on automation showing that removing people from routine practice can erode expertise they may later need during unusual or high-stakes situations.

The paper's third scenario imagines an elder-care facility that can no longer afford an AI system woven into care planning and scheduling. Workers trying to reconstruct the previous workflow discover that institutional knowledge once distributed among spreadsheets, handwritten logs and employees has become difficult to recover.

CReF does not predict that these things will happen, and it does not assign organizations a resilience score. The authors describe it instead as a "generative analysis tool" intended to expose hidden dependencies and potential safeguards.

Companies should not abandon AI, the researchers argue, but they should develop resilience alongside adoption. The question for an AI-dependent workplace, in other words, may not only be how much better people perform when the technology is available. It may also be what remains when it isn't.

Das Swain, Vedant & Saha, Koustuv. (2026). AI Resilient Future of Work: A Counterfactual Lens on Human-AI Collaboration.