Chat is useful when people need to express an ambiguous intention, explore a subject or refine a request through dialogue. It is a poor default when they need to compare options, inspect state, control an action or recover from failure. The product decision is not how to place a language model inside a chat window. It is which interaction mode makes the user's work, the system's capability and the consequence of action easiest to understand and control.

Chat is an interaction mode, not an AI strategy

A blank message box makes a powerful promise. It suggests that the user can ask for almost anything in their own words and that the product will work out what they mean. That flexibility can remove the cost of learning commands, navigating a complex menu or translating an uncertain goal into a rigid form.

The same openness also hides the shape of the product. A user cannot see which requests are supported, what information the system needs, how results will be structured or whether a sentence will produce advice or trigger an action. The interface can feel simple while moving discovery, specification and error handling into the conversation itself.

This is not an argument against conversation. It is an argument against choosing it before understanding the work. Product teams should treat chat as one interaction mode among several, then earn its place through the user task rather than the presence of a language model.

Machine learning changed what a conversation can conceal

Earlier conversational products were often built around defined intents, scripted routes and limited commands. Their failures could be frustrating, but their operational boundary was usually narrow. Generative models can interpret varied language, maintain context and produce plausible responses across a much wider surface. That makes conversation more useful and its boundary harder to infer.

Microsoft Research found a marked gap between user expectations and the capabilities and goals of conversational agents in interviews with 14 users published in 2016. Current systems are more capable, but the underlying design problem remains. Human language encourages people to attribute understanding and flexibility that the product may not possess. Google People + AI Research similarly warns that human-like presentation can create unrealistic expectations and recommends making capabilities, limitations and adaptation legible.

A fluent answer is therefore not enough evidence that the interaction worked. The product must help the user understand what the system can do, how well it can do it and what to do when it is wrong. Microsoft Research's 18 human-AI interaction guidelines establish that broader requirement across the initial interaction, normal use, failure and change over time.

Choose the surface from the work

Conversation is strong when the user needs to express an incomplete goal, provide rich context, explore alternatives or refine an artefact through successive changes. Language can compress a large amount of intent into one exchange. It can also help when there is no stable sequence of fields because the right question depends on the preceding answer.

A structured surface is usually stronger when people need to scan available actions, compare several options, edit precise values, repeat a task at volume or understand a stable state at a glance. Direct manipulation is often better for spatial, visual and compositional work. A persistent agent workspace is better when delegated work continues across time, tools or approval points and must remain inspectable after the conversation closes.

Many useful products will combine these modes. Language can capture an initial intention. Structured controls can make constraints explicit. A preview can expose the proposed result. A workspace can preserve progress and evidence. The design quality lies in moving between these modes without losing context, not in forcing the whole task through a transcript.

An interface that can act needs more than dialogue

The interaction requirement changes again when the system can invoke tools or alter another system. OpenAI distinguishes agents from simple chatbots by their ability to manage workflow execution and use tools on a user's behalf. That distinction is provider guidance, but it identifies a material product boundary. Once an utterance can change a customer record, initiate a transaction or continue working in the background, the conversation becomes part of an operating system.

Microsoft Research's work on human-agent communication identifies 12 challenges created by systems that communicate in natural language while using tools and acting with greater independence. The research calls for new patterns that support transparency and control. Product teams should respond by exposing the proposed action, affected objects, source information, permissions, approval point, current state and recovery route where each is material.

A confirmation question inside a transcript may be sufficient for a small, reversible action. It is not sufficient when the user needs to compare many changes, understand dependencies or distinguish what has already happened from what is merely proposed. In those cases, conversation should hand over to a surface built for review and control.

Design the boundary and recovery with the main path

A conversational product cannot rely on users discovering its limits by repeated failure. Give examples that reveal the supported work. Ask for missing information at the point it becomes relevant. When uncertainty affects the result, show the evidence or limitation that changes the user's decision. When the AI route fails, preserve a structured or manual way to complete the task where the context allows it.

This is also where multimodal design becomes commercially important. Visible options reduce unnecessary turns. Structured review can prevent a plausible response from becoming an unnoticed error. Clear hand-off and recovery protect the user's time and the organisation's operating confidence. The result may contain conversation, but the product no longer depends on conversation to carry every responsibility.

Measure the task, not the amount of talk

A longer exchange is not evidence of engagement if each turn repairs ambiguity created by the interface. A shorter exchange is not automatically better if the user accepts an unsupported answer. Evaluate whether the interaction advances the user's goal while preserving understanding and control.

Microsoft Research's RUBICON work evaluates task-oriented human-AI conversations against the user's goals and the wider conversational context. Its published application to more than 12,000 debugging conversations is provider research, not a universal product measure. The useful implication is that response fluency and turn-level ratings are too narrow when the purpose of the interaction is to accomplish work.

Track task completion, unnecessary turns, correction, comparison, abandonment, hand-off, recovery and the consequences of accepted errors. Segment the evidence by interaction mode. This reveals whether conversation is reducing effort or merely hiding product structure inside a sequence of messages.

Start with one user task, not one universal interface

Choose a task that matters to the user and map its intent, structure, evidence, consequence, duration and control. Then assign the interaction mode that makes each part easiest to understand and operate. Conversation may lead the exploratory work, support a structured workflow or disappear once the task becomes repeatable.

The aim is not to minimise chat. It is to prevent chat from becoming an attractive container for unresolved product decisions. When the interaction follows the work, AI can feel more capable because the product makes capability, state and control visible in the form best suited to each decision.

Use an interaction-mode test

Before drawing the interface, test one important user task against six questions. The test helps a team decide whether conversation should lead, support another surface or stay out of the task. It is not a fixed taxonomy. Products often need a combination, and the right balance can change as the task becomes better understood.

Start with the user's work rather than the model's capabilities. If the answers reveal a stable, high-consequence or comparison-heavy task, a blank conversational surface is unlikely to carry enough structure. If the goal is genuinely ambiguous and clarification creates useful progress, conversation may deserve the leading role.

  • Intent: Is the user expressing an ambiguous goal or selecting a known action?
  • Structure: Does the task benefit from open language, or from visible fields, options and constraints?
  • Evidence: Must the user compare sources, alternatives or changes before accepting the result?
  • Consequence: Is the output advisory and reversible, or can it create a material external effect?
  • Duration: Will the task finish in the exchange, or continue across tools, people and time?
  • Control: Where must the user inspect, edit, approve, stop or recover the work?

Sources and further reading

  1. Microsoft ResearchGuidelines for Human-AI Interaction
  2. Microsoft ResearchThe Gulf between User Expectation and Experience of Conversational Agents
  3. Google People + AI ResearchMental Models
  4. Microsoft ResearchChallenges in Human-Agent Communication
  5. OpenAIA practical guide to building AI agents
  6. Microsoft ResearchRUBICON: Evaluating conversations between humans and AI systems