Why Everyone Keeps Extending TAM Instead of Replacing It

Davis built the Technology Acceptance Model in 1989. Perceived ease of use leads to perceived usefulness, which leads to intention, which leads to use. The chain is clean. Every IS doctoral student learns it in their first semester. And for a model designed around email, word processors, and spreadsheet software, it did its job.

But that was 37 years ago. Today the same chain, with the same core constructs, is being applied to AI agents that make decisions autonomously, blockchain platforms where governance is distributed across nodes, and virtual reality environments where the boundary between tool and environment dissolves. I keep reading papers that take TAM, add one or two constructs, slap "AI" into the title, and call it a theoretical contribution. The papers are not wrong in any formal sense. The path coefficients are significant. The R-squared values are acceptable. The reviewers sign off. And yet something is off, and I think it has been off for decades.

Venkatesh et al. (2003) consolidated eight prior acceptance models into UTAUT: performance expectancy, effort expectancy, social influence, and facilitating conditions. UTAUT became TAM's big sibling. It was more comprehensive, accounted for moderators like gender and age, and fit the data better. Then came UTAUT2. Then came context-specific extensions for mobile, for healthcare, for e-government, for e-learning. The pattern was always the same. Find a new technology context. Add a construct or two. Test the model with survey data. Report significant paths. Publish.

Every IS researcher knows this pattern. We have been doing it for three decades.

What I did not have until recently was a precise vocabulary for what was happening. Burton-Jones et al. (2021) gave IS the language it needed (and if you want to understand what theory even is before asking whether we are doing it right, I covered that separately). They proposed four next-generation theorizing strategies. Extend adds boundary conditions to existing theories to cover new contexts. Reformulate applies new logics to existing constructs without abandoning them. Replace introduces new constructs to substitute for inadequate ones. Envision adopts entirely new ontological assumptions about how the world works.

Extension is by far the most common strategy. And once you understand the incentives, it is hard to blame anyone for choosing it.

Extension preserves the existing construct network. Perceived usefulness and perceived ease of use are known quantities. Their measurement instruments are validated, cited thousands of times, and trusted by reviewers. You do not have to argue for the validity of your measures. You do not have to explain what your constructs mean. You just add a new pathway, a moderator, a context variable, and the reviewers can follow the logic because they have seen it a hundred times before.

Extension preserves citational momentum. Every TAM extension cites Davis (1989). Every UTAUT extension cites Venkatesh et al. (2003). These papers have tens of thousands of citations. Citing them is safe. The reviewers wrote papers in the same tradition. The editors built their careers on the same research program. Rejection is less likely when you are playing within the established paradigm.

Extension preserves the reviewer's comfort zone. Sarker et al. (2019) documented what this comfort looks like empirically. Fifty-six percent of published IS research is Type I on their sociotechnical axis: purely social, with IT only as background context. The methods are familiar, the theory is imported, the review process is predictable. Only thirteen percent achieves Type IV, where social and technical genuinely interact. Extension is a Type III or Type I move. It is the path of least resistance.

Replacement is harder for exactly the opposite reasons. A replacement strategy requires building a new nomological net from scratch. You must define new constructs, develop new measures, establish discriminant validity against the old ones, and argue that the old model was not just incomplete but fundamentally wrong for the new context. That is a multi-paper program. That requires patience. That gets rejected by reviewers who ask why you did not just add a moderator to TAM instead.

Reformulation sits between extension and replacement, and it is rare precisely because it demands more than either. It requires you to accept the old constructs but rewire the logic connecting them. When Baird and Maruping (2021) argued that delegation should replace use as the central construct for agentic IS, they were not extending TAM. They were reformulating the very relationship between human and system. Delegation involves appraisal, distribution, and coordination. It is bidirectional: the human delegates to the AI, and the AI can delegate back. It cannot collapse into "intention to use" because the system is no longer a passive tool waiting to be operated. I wrote about this shift in more detail before, but the theoretical move matters here for a different reason. Reformulation is not a TAM extension, and that distinction is precisely what the Burton-Jones framework was built to capture.

The result is that almost every "AI adoption" paper I read follows the extension playbook. Take TAM or UTAUT. Add "trust in AI" as a new construct, or "algorithmic transparency," or "perceived autonomy." Test it on a sample of 300 MTurk respondents. Report that the new construct has a significant path to behavioral intention. The core model remains unchanged. The theoretical contribution is marginal.

And here is where I need to say something that might not make me popular at an IS conference. I think extending TAM to agentic AI is the IS equivalent of adding epicycles to Ptolemaic astronomy. Each new construct is another circle on top of circles. The model keeps fitting the data because you keep adding parameters. But the core assumption is wrong, and no number of additional constructs fixes it.

The core assumption of TAM, UTAUT, and every model in that lineage is that the human operates the tool. The system is passive. It has no agency, no goals, no capacity to initiate action. Perceived usefulness means "does this tool help me do my job?" Perceived ease of use means "is this tool hard to learn?" Both constructs assume the tool sits still and waits for the human to act.

Agentic systems do not wait. They recommend, filter, reroute, allocate, and sometimes overrule. The human is no longer the sole operator. In many cases, the human is not even the primary decision-maker. The relationship shifts from operation to delegation, and delegation requires an entirely different set of constructs. Appraisal: can I trust this system with this task? Distribution: who has decision rights over what? Coordination: how do we manage the ongoing relationship when both parties can act? These are not new variables to add to a regression. They are a different nomological net.

I understand why the field keeps choosing extension. The incentives are stacked that way. A TAM extension gets published. A radical replacement gets desk-rejected. I am not being cynical here. The institutional structure of academic publishing rewards incremental contribution over foundational rethinking. The Burton-Jones et al. framework diagnoses this problem, but diagnosing it and fixing it are different projects.

What I keep thinking about is how long this can continue. The Ptolemaic model of astronomy worked for over a thousand years. It predicted planetary positions with reasonable accuracy. It just added more epicycles to handle the anomalies. Kepler did not improve the Ptolemaic model by adding a hundred circles. He replaced the core assumption, circular orbits, with elliptical ones. The data fit better, but that was not the real contribution. The real contribution was changing the question from "what circles explain this motion?" to "what shape does this motion actually trace?"

I do not think the IS field needs to abandon TAM. For the class of problems it was built to solve, people deciding whether to adopt a tool they will operate, it works fine. But applying it to systems that operate themselves is not extension. It is category error disguised as empirical rigor.

The next time I see a paper titled "Extending TAM to AI-Powered [Insert Application]," I will ask myself the same question Sarker et al. would ask. If you remove the IT artifact from this model, does anything change? If the answer is no, the paper is not about AI. It is about survey scales and path coefficients, with AI as the noun in the title.