Enterprise AI Needs Memory and Process Redesign, Not Bigger Chatbots: what changed

Enrique Dans argues enterprise AI failures trace to architecture: LLMs are text predictors and lack the persistent memory, context, feedback loops and constraints organizations need.

Enrique Dans argues that enterprise AI is underdelivering because it’s built on the wrong architecture: large language models were designed to predict text, not to run company operations. That mismatch matters because organizations require persistent memory, cross — session context, enforceable rules and measurable feedback to make reliable decisions — capabilities that off-the-shelf LLMs and session — based assistants typically do not provide.

Dans contrasts the internal mechanics of LLMs with the operational needs of businesses. Language models excel at generating plausible text but lack durable state and explicit constraints; they do not, by themselves, preserve context across workflows, enforce business rules, or capture the feedback loops necessary to improve outcomes. Without those elements, AI responses remain isolated answers rather than components of repeatable, auditable processes.

Over the past two years much industry effort has focused on the visible layer — bigger models, smoother interfaces, more polished copilots and emergent agents — but Dans says the clearest signals of value come from projects that redesign workflows. Deployments that embed AI into the fabric of operations, he notes, behave less like improved chatbots or prompt chains and more like integrated infrastructure that supports ongoing work rather than a disposable assistant.

Market evidence aligns with that diagnosis. McKinsey’s latest global survey finds AI adoption is widespread but rarely embedded deeply enough into workflows to produce material enterprise — level benefits, and it identifies workflow redesign as one of the strongest drivers of meaningful business impact. In other words, simply increasing model size or adding interfaces does not automatically translate into measurable value without changes to process and integration.

For builders the practical implications are concrete: move beyond prompt — first designs toward systems that preserve state, enforce constraints, and create feedback channels tied to measurable outcomes. Dans emphasizes that enterprise AI must ‘remember’—support persistent context, business rules and audit trails — so it can operate across business processes rather than within isolated user sessions. If Dans is correct, the next frontier for enterprise AI is engineering systems of action: integrating models with workflow logic, persistent memory, feedback mechanisms and operational constraints. He points to organizations that have already begun redesigning their operations around embedded AI capabilities as early evidence that architectural changes, not marginally better models, are the path to enterprise — level impact.