There are two fundamentally different visions of AI in the workplace, and the IS research frame you choose changes everything about what you study and what you find.
There is a framing problem at the center of most AI-and-work research, and I think it is producing conclusions that sound like they are about the same thing but are actually about different things entirely. The problem is that "AI in the workplace" collapses two fundamentally different configurations into a single category, and the IS research frame you apply depends on which one you are actually studying.
The first configuration is automation. AI replaces a human task. A customer service bot handles tier-one support inquiries without human involvement. A document processing system extracts and routes information from invoices without a data entry clerk. A code generation tool writes boilerplate functions without a developer. In each case, the human who previously performed that task is either reassigned or no longer needed. The value proposition is cost reduction and throughput increase. The system does not help the human do the job. It does the job in place of the human.
The second configuration is augmentation. AI helps a human do the job better or faster. A radiologist reviews AI-flagged areas of a scan rather than scanning the full image manually for anomalies. A salesperson uses an AI recommendation system to identify which customers to contact this week and what to pitch. A developer uses an AI coding assistant to generate options and catch errors, but reviews and decides on every piece of code. The human remains in the loop and in the decision. The AI changes the workflow but does not eliminate the human role.
These are not just different points on a spectrum. They are qualitatively different configurations with different governance implications, different equity implications, and different IS research questions. Treating them as the same phenomenon produces muddled findings.
The IS theory implications are clearer than most papers acknowledge. The Technology Acceptance Model, which I have spent considerable time with in my comps preparation and which is confirmed in my study-hub materials as Davis (1989) with the chain from perceived ease of use to perceived usefulness to intention to use, was designed for the augmentation configuration. It models a human deciding whether to use a tool. Perceived usefulness in TAM means "does this tool help me do my job better?" That question only makes sense if the human is still doing the job. In an automation scenario where the system replaces the human task, TAM is not the right frame. There is no one deciding to use the tool. The question is whether the organization decides to deploy the system, not whether the worker decides to adopt it.
Social Cognitive Theory applies differently as well. Bandura's framework, confirmed in my study-hub materials as developed across 1977 and 1986 with IS applications notably through Compeau and Higgins (1995) on computer self-efficacy, is fundamentally about whether humans believe they can successfully perform specific behaviors. In augmentation contexts, self-efficacy is critical: a worker's belief that they can effectively use an AI copilot to enhance their own work shapes whether they actually engage with it. Workers with low AI self-efficacy avoid the tool or use it superficially, which limits the augmentation benefit. In automation contexts, self-efficacy is largely irrelevant to the worker because the worker is not operating the system. The relevant human cognition is the manager deciding to deploy and the compliance officer deciding whether the deployment meets policy requirements.
The political economy dimension is what makes the framing choice consequential beyond academic classification. Automation concentrates the gains from AI at the organizational level and potentially at the ownership level. The productivity gains accrue to the organization, not the workers whose tasks were replaced. If those workers are reassigned to other tasks, they may or may not capture any of the productivity gain in wages or job quality. If they are displaced, they capture none of it. Augmentation has a different distributional logic. When AI makes a skilled worker more productive, the gains can in principle be shared between the organization and the worker, through higher wages, higher commission rates, faster career advancement, or reduced workload. The word "can" is doing significant work in that sentence. Whether the gains are actually shared depends on labor market conditions, bargaining power, and organizational pay practices. But the structural possibility of shared gains exists in augmentation in a way that it does not in displacement automation.
This is not an argument that augmentation is good and automation is bad. Both configurations involve real tradeoffs, and the line between them is not always clear in practice. A customer service bot that handles tier-one inquiries does eliminate tier-one jobs, but it may also create demand for more complex tier-two and tier-three support work that is harder to automate. The net employment effect depends on elasticity of demand for customer service capacity, substitution rates, and labor market conditions that are genuinely hard to predict. I would not cite specific labor market predictions about AI and employment without a current empirical source in front of me.
What I want to argue for is that IS researchers should be explicit about which frame they are using before choosing their theoretical tools. If you are studying a copilot system where humans use AI assistance to make better decisions, TAM extensions and Social Cognitive Theory are reasonable starting points. You are studying individual and organizational adoption of a tool. If you are studying a system that makes decisions autonomously and routes outputs to humans for review or execution, you are studying algorithmic governance and accountability, which calls for different constructs and different methodologies. The agency in the system is different. The locus of acceptance or resistance is different. The governance challenge is different.
The copilot framing has become popular in enterprise AI partly because it is more palatable than saying "we are automating your job." Microsoft's GitHub Copilot, Google's Workspace AI features, and Salesforce's Einstein all use the copilot or assistant frame deliberately. It positions the AI as a tool in service of the human worker, which reduces resistance and aligns with augmentation theory. But the same system can function as an augmentation tool for skilled workers and as a displacement tool for less-skilled workers doing routine tasks within the same workflow. A coding assistant that generates boilerplate augments the senior developer who reviews and adapts the output. It potentially displaces the junior developer whose primary value was writing that boilerplate.
The IS research that I find most useful takes the configuration seriously rather than treating "AI adoption" as a unitary phenomenon. What is the human doing before the AI is introduced? What is the human doing after? What decisions does the human still make, and which decisions does the system make? Those questions reveal the configuration, and the configuration determines which theories are actually applicable.
The frame is not just an academic preference. It shapes what organizations measure, what workers experience, and who is accountable for outcomes. An organization that frames every AI deployment as augmentation without examining whether it functions as displacement is telling itself a story that may not match what is actually happening to its workers. Researchers who use TAM to study an automation system are asking the wrong questions of the wrong people.
About the author
Share
More notes
Related notes