Migrating to Agents? One Question to Ask First

AI Agents are all the buzz right now. Wherever you look, teams are either building agents, scaling agents, or worrying about how far behind they are.

When agents came into the picture, Google came up with protocols (A2A) for agents to find and interact with each other. Marketplaces have launched so agents can be discovered. A whole genre of work is forming around agent orchestration patterns.

An AI agent is a single LLM-driven system that performs a defined task using tools. An agentic system layers architecture on top: an orchestrator agent decides which sub-agents to invoke, stitches their outputs into a result, and adapts based on context and instructions.

All of this is reshaping how teams work, and I have one question I haven’t seen many people ask out loud or answer yet. Gartner predicts more than 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls. Product leaders are approving migrations to agents faster than we’ve set the bar to evaluate them.

There’s a pattern playing out everywhere right now. A team ships an agent that resolves a specific user or business problem. The proof of concept is convincing, the demo runs clean, executives are supportive, and the project gets the green light to scale. No one asked what the existing workflow was already resolving or at what cost. Even teams that do measure value typically measure it on the new system alone, not against the workflow it replaced. Value delivered is not the same as incremental value delivered.

Consider a specific case. Migrating an existing conversational AI experience from LLM responses to an agentic architecture means rebuilding the workflow end-to-end and redoing instrumentation. To be worth it, the new architecture has to deliver incremental value over what the LLM workflow already does. Would you rather improve LLM response accuracy and latency on the system you already have, or rebuild it on a new architecture for an unproven delta?

Agentic systems cost more to operate than workflows do in terms of evaluation, observability, escalation paths, and retraining. If the delta over the workflow is zero, you’ve imported cost without delivering incremental value.

So here’s my challenge to product leaders: if you already have a workflow, does moving to an agentic architecture offer incrementality? Could you reach the same outcome by staying on an LLM-based workflow? What does the system deliver above and beyond your existing workflow, now, in a year, in five years? Will agentic systems even be part of the architecture two years from now, or will they consolidate into something else?

Before approving any workflow-to-agent migration, ask one question: compared to what? Every well-built system delivers value. What matters is the delta over the system you already have. If you cannot name the workflow being replaced and write down the delta you expect, you are running a land grab dressed as a migration.