Enrique Dans: Two Years of Enterprise AI Show Firms Must Rebuild Processes, Not Just Speed Them Up

Enrique Dans argues that roughly two years of enterprise AI deployments have exposed a structural mistake: Companies tried to accelerate legacy workflows by bolting large language models onto existing systems, and that approach is hitting practical limits.

Enrique Dans says two years of enterprise AI experiments have revealed a hard lesson: companies that tried to speed up existing workflows by grafting large language models onto legacy processes are now confronting scalability and effectiveness limits. Initial enthusiasm for inserting models into operational layers has collided with practical constraints, and Dans warns the result is not merely lower — than-expected productivity but a structural mismatch that prevents reliable, repeatable gains. This matters because it shifts the burden from short — term automation wins to broader organizational redesign — requiring business leaders and technologists to coordinate on rebuilding processes, not just deploying copilots or assistants.

Early enterprise efforts focused on embedding LLMs in operational layers — adding copilots, assistants and automation layers intended to accelerate routine tasks and boost productivity. Those experiments produced mixed results: the models could speed certain activities, but Dans argues the technology itself is not the root failure. Instead, the problem lies in placing LLMs at the wrong architectural layer, where legacy decision chains, brittle integrations and fragmented data landscapes blunt the models’ effectiveness and make outcomes unpredictable.

To frame the debate, Dans revisits business process reengineering (BPR) from the 1990s, which advocated redesigning organizations around information systems rather than layering new software on top of existing workflows. He notes BPR efforts often yielded uneven returns: they could become expensive reorganizations that fell short of transforming how companies operated because underlying systems remained rigid and unable to adapt in real time. That history, he suggests, is a cautionary tale for firms now experimenting with AI.

Dans contends the current moment differs from the 1990s because modern AI capabilities demand a renewed willingness to rethink core processes rather than only optimize them. He urges companies to stop asking solely how to use AI in existing processes and instead ask whether those processes were ever designed for AI. That reframing moves the conversation from incremental productivity gains to structural decisions about where and how models should be integrated to produce reliable value.

For builders and technical leaders, Dans’s practical takeaway is clear: move beyond superficial automation experiments and evaluate whether current workflows and system layers can meaningfully host LLM-driven capabilities. Organizations should reassess process architecture, integration points and the scope of organizational change required so that AI is embedded in layers aligned with its strengths, rather than being bolted onto incompatible systems. The choice is between transient productivity bumps and deeper redesign that can sustain AI-driven change.