All workEssay 04 / 12

Conversations, interruptions & second chances

People don’t speak
in flowcharts.

They hesitate, correct themselves, leave and come back. Building conversational products meant learning to design for all of that—not just the perfect reply.

“Actually, can we go back?”

A perfectly ordinary request. A revealing test of the system.

Listen

An interruption is information.

A conventional interface can constrain the order of a task. A conversation has fewer visible boundaries. Someone may answer yesterday’s question, introduce a new subject or send only half the information a service needs. Treating those messages as invalid input misses what the person is trying to do.

During six months of conversational and community product work, I worked across messaging integrations, workflow automation and the services supporting them. The difficult part was keeping the journey understandable when a message did not fit the expected next step.

Thinking in states helped. The system needed to know what had already been established, what remained uncertain and which actions were still reversible. The words shown to the person had to express that without turning the conversation into a debugging console.

One task, an unexpected turnIllustrated exchange

Help me find a time.

Morning or afternoon?

Actually, I need to change the day.

Please select morning or afternoon.

Illustrative example · not live product data
Remember

Keep context that earns its place.

Stored history can make a return visit feel continuous. It can also bring back information that no longer matters. Remembering everything is not a substitute for deciding what the current task needs.

I separated reusable product rules from channel delivery. The intent behind an action should remain stable even when one channel uses a button and another uses a short message. Delivery details belong near the channel; the meaning of the action belongs in the product.

That separation also makes failure easier to describe. If a service cannot finish a task, the conversation should retain what is useful and offer a next step. Asking the person to start again may be technically convenient, but it makes them pay for an integration problem they cannot see.

Continue

A good reply leaves a useful way out.

A fallback should do more than say that something went wrong. It can narrow an ambiguous request, offer the previous choice again or explain why a human handoff is needed. Recovery is part of the normal journey.

More flexible language brings more ambiguity. More context brings more responsibility for retention and relevance. Automation helps with common paths, while unusual or sensitive requests still need boundaries. Those trade-offs do not disappear because the interface sounds friendly.

There is also a difference between recovering the task and repeating the message. If a person has already supplied a date, asking for the date again may technically restart the flow while still damaging the experience. A better recovery identifies what is trustworthy, keeps it, and asks only for the missing or changed part.

The illustrative exchange above makes this visible with a single interruption. The second reply does not become useful by sounding more cheerful. It becomes useful because it acknowledges the change of direction and gives the person a concrete choice that the system can act on.

I came away less interested in a chatbot that produces a polished first answer and more interested in one that remains respectful on the fourth correction. The quality of the conversation lives in that patience.

Let people change their minds without losing their place.
CONTEXT / LINKS

Related work & references

Selected public entry points and related portfolio pages. A company website or profile is context, not a claim that private source code is available.

Scope, context & further questions

Grounded in public messaging, chatbot, community, chat-history, and embedded-app work plus verified internship experience.

  • Industry internship: 6 months. Hands-on delivery of conversational and community experiences with cross-functional teams.
  • Delivery surface: Multi-channel. Messaging APIs, lightweight embedded apps, workflow automation, and web services worked as one journey.
  • Core design problem: Stateful. History and context had to survive beyond a single message-response pair.

Questions I am still working through

How much context is enough before continuity becomes surveillance?

Only context that is necessary for the current relationship and task should be retained by default. Short-lived conversational state and durable domain data should be treated differently, with visible controls for anything persistent.

Different product categories will need different retention boundaries, so the answer cannot be one universal token window or time period.

How should a conversational product express uncertainty without becoming frustrating?

State uncertainty briefly, explain what is missing, and immediately offer the next useful action: clarify, choose from options, retry, or hand off. Uncertainty should move the conversation forward rather than become a disclaimer wall.

The remaining design problem is calibrating how much explanation each risk level deserves.

Stories, products & useful placesEsc to close