The IS Success Model and What It Misses Today

I have been sitting with the DeLone and McLean IS success model for a while, and the more I work through it for my comps, the more I notice what the model assumes versus what the world it is applied to actually looks like now.

DeLone and McLean (1992) built a model with six interrelated dimensions: system quality, information quality, use, user satisfaction, individual impact, and organizational impact. The logic is that quality dimensions influence use, use and satisfaction influence each other in a feedback loop, and both lead to individual and then organizational impact. The 2003 update kept the six-dimension structure by adding service quality, reflecting the shift toward IT departments as service providers rather than just system builders, and replaced the individual and organizational impact dimensions with the broader construct of net benefits. That update also acknowledged that the model needed to handle mandatory use scenarios, where employees use systems because they have to, not because they chose to. Net benefits was defined broadly enough to apply at the individual, organizational, or societal level.

The model is impressive as a synthesis effort. Petter, DeLone, and McLean (2013) mapped the antecedents of IS success across task, individual, social, project, and technology factors, extending the framework substantially. These papers are among the most cited in the IS literature for a reason. The model gave researchers a shared vocabulary for talking about IS success and a structure for organizing a genuinely fragmented empirical literature.

But here is what I keep thinking about. The original model was designed for transactional enterprise systems. Think ERP, payroll, inventory management, order processing. Systems where use is relatively well-defined, where the information the system produces can be evaluated against known accuracy standards, and where the link between system use and organizational outcomes is conceptually clear even if empirically complex to measure.

Most software being deployed today is not that. It is SaaS delivered on someone else's infrastructure with update cycles you did not choose. It is AI-embedded, which means the information the system produces is generated by a model that can be confidently wrong. It is used by people who never had a say in the procurement decision. And it operates inside ecosystems of connected platforms, where the boundaries of any one system are blurry.

Take information quality. In the DeLone and McLean model, information quality means accuracy, completeness, timeliness, and format. Those are reasonable criteria for a report generated from a relational database. But what does information quality mean when the information is generated by a large language model that hallucinates citations, summarizes contracts incorrectly, and does so with the same confident tone it uses when it is right? The construct is not simply applied. It requires rethinking. Accuracy becomes harder to verify, not because the underlying data is bad but because the generation process is opaque. And the user often has no reliable way to distinguish accurate outputs from inaccurate ones without independently verifying the information, which defeats part of the purpose.

The use dimension has a parallel problem. The original model assumed use was voluntary and measurable through frequency and duration. The 2003 update acknowledged mandatory use but did not fully resolve it. Today, many enterprise systems are so deeply embedded in daily work that use is essentially unconscious. Nobody counts their Slack messages. Nobody measures their time in Salesforce as a deliberate choice. Use is ambient. Measuring it through frequency or duration captures something real, but it does not capture what Burton-Jones and Grange (2013), writing from a representation theory perspective, called effective use, which is about whether use is faithful to the domain the system was built to represent. Logging in is not the same as using well. The IS success model, in its standard operationalizations, often cannot tell the difference.

Net benefits is the dimension where I think the gap is most visible. Who are the beneficiaries? For a transactional ERP system, the answer is roughly: the organization that bought it and the employees whose work it supports. For TikTok, the beneficiaries are the platform, the advertisers, the creators, and the users. But the platform's interests and the user's interests are not aligned. The algorithm is designed to maximize engagement, which produces outcomes that are beneficial to TikTok's revenue model and often not beneficial to users' time allocation, mental health, or information quality. Success for the platform. Harm for some users. How does net benefits handle that? The original model has no framework for situations where the organizational benefits of a system come at a cost to users, or where societal harms and individual benefits coexist in the same system.

The consumer platform case also breaks the user satisfaction dimension in interesting ways. Satisfaction in the IS success model is about whether users are satisfied with the information, the system, and the service. But on platforms like TikTok or Instagram, satisfaction in the hedonic sense, which is something closer to pleasure or engagement, is engineered rather than discovered. The recommendation system is specifically designed to maximize time-on-platform, which correlates with certain kinds of satisfaction and directly contradicts others. You can be satisfied in a compulsive, undesired way. The model was not built for that distinction.

I am not saying DeLone and McLean were wrong. The model worked well for the empirical landscape it was built for. But the landscape changed, and the changes are not incremental. AI-generated information, mandatory ambient use, multi-sided platforms with misaligned stakeholder interests, and SaaS systems with invisible infrastructure all create situations where the standard operationalizations of system quality, information quality, use, and net benefits either underdetermine the answer or point in conflicting directions depending on whose perspective you take.

The interesting IS research question is not whether the model fails for these cases. It is what a revised model would look like that could handle them, and whether such a revision would still be recognizable as the same theoretical structure or would require building something new.