Chat Was the Demo. The Workflow Is the Product.

The default interface for a new AI product used to be an open text box: type anything, get a response. That interface was excellent for demonstrating what a model could do and has turned out to be a mediocre way to actually accomplish recurring work. Salesforce's Agentforce and Intercom's Fin gain real traction now by hiding the chat box almost entirely, embedding model calls inside approval queues and support ticket flows where a user barely notices an AI system is involved.

The reason is straightforward once named: an open chat interface asks the user to remember what to ask and what to do with the answer, every time. A workflow does that remembering for them. Fin sits inside a support queue and resolves a ticket end to end rather than waiting for someone to paste a question into a box; Agentforce updates a Salesforce record automatically rather than requiring a rep to ask it to.

This is a return to an older lesson in software design: users do not actually want a general-purpose tool that can do anything if it also requires them to figure out how. Chat interfaces are general-purpose almost by definition, which makes them powerful for exploration and weak for repeated, well-understood work — exactly the kind of work Fin and Agentforce were built to absorb.

The workflow-first products also solve a trust problem pure chat struggles with. A chat response lives in isolation, hard to verify against anything else. A Fin resolution sits inside a ticketing system with an audit trail a support manager can review; an Agentforce update sits inside a Salesforce record history anyone on the deal can check. The structure around the AI output does real work in making that output trustworthy, independent of the model's raw quality.

There is a real cost to this approach that pure-chat products avoid: building Agentforce or Fin into an existing workflow requires mapping out the process and the connective infrastructure between steps, which is slower and more expensive than shipping an open chat box. Chat is easy to build and hard to make durably useful. Workflow is hard to build and, once built, tends to stay useful in exactly the way chat struggles to.

The products that will define this next stretch of AI adoption are unlikely to be remembered as chatbots at all. They will be the support ticket Fin already resolved and the Salesforce stage Agentforce already advanced — with the model doing real work exactly where a workflow needed it, and nowhere the user had to ask.