Not everything should be a chatbot

Read these first

• Problem selection

What you will know

• Propose a non-chat surface for a request that arrived as "add a chatbot"

  Chat pushes the work of figuring out what to ask onto the user, especially when required fields and approval steps are predictable.

• Compare inline suggestion, background job, structured form, and batch enrichment against a concrete request and record the selection

  The useful discriminators are ambiguity, repetition, urgency, human approval, and whether results can be corrected asynchronously.

• Trace the work a user must do before and after a conversational request, then remove one unnecessary turn

  Unresolved inputs often surface only at runtime as clarification loops, malformed outputs, or abandoned flows.

• Show a feature that got better by removing the conversation

  A before-and-after flow or prototype makes reduced interaction cost and improved completion visible to reviewers.

• Defend the choice to a stakeholder who wanted a chatbot

  Completion rate, latency, correction effort, and failure recovery give the discussion evidence beyond novelty or demo appeal.

Chapters

1. 1

   Chat hands the user the job of discovering what the system can do

   A form shows its fields; a text box shows nothing, so the user has to guess what the system understands, phrase it, and try again when the guess is wrong. That is why the same job often completes worse in chat than in a form even when the model answers well. This chapter names the work a conversation transfers to the user, and how to spot it in a request that arrives as "add a chatbot".

   Read chapter
2. 2

   Five surfaces, and the questions that choose between them

   Inline suggestion, background job, structured form, batch enrichment and conversation are the options, and six discriminators pick between them: ambiguity, repetition, urgency, whether a human must approve, whether a wrong result can be corrected later, and whether a paraphrase is an acceptable answer at all or the user needs the governing artefact itself, with a record of what they were shown. That last one rules on legal, medical, compliance and pricing answers before any of the others get a vote. Lead enrichment on every new record is a background job; an expense report whose eight fields all sit in a photo is a prefilled form. You will run a real request through the grid and record the selection.

   Read chapter
3. 3

   An inline suggestion must be dismissible; batch enrichment trades freshness for cost

   Two of the five surfaces usually survive only as names on a list, and each is chosen on properties the general discriminators do not measure. An inline suggestion has to be dismissible, non-blocking and fast enough to beat the user's own typing, and its acceptance rate is the feedback signal that says whether it is earning its place on screen. Batch enrichment buys unit cost by giving up freshness, works only where a wrong value can be corrected on the next run, and is incomplete without a backfill and re-run story. Judge both against a concrete request rather than in the abstract.

   Read chapter
4. 4

   Conversation earns its place when users cannot state what they want

   Chat fits where the request space is genuinely open-ended and the user cannot name what they need until they have seen a response, so the turns do real work. That property, not task complexity, is the separator, and it rules only once the paraphrase question has been cleared: where the user needs the governing artefact itself, the grid has already decided and open-endedness does not reopen it. A long multi-step job with a known shape belongs in a wizard or a background job. You will place two features on each side of the line and state the property that divides them.

   Read chapter
5. 5

   When the answer must be the source clause, a paraphrase is a liability

   A nurse asking whether two drugs can be co-administered does not want a fluent restatement of the policy. They want the governing clause, at the moment of administration, with a visible refusal when the policy does not cover the case. Wrapping that in chat inserts a paraphrase between the nurse and the text that actually decides, so the surface is a policy lookup inside the workflow with the passage itself as the answer.

   Read chapter
6. 6

   The turn worth deleting is the one asking which project you meant

   An assistant that opens by asking which project the user means charges every session for a fact the interface already holds; infer it from context and let the user override. Here you trace the work a person does before and after a conversational request, spot unresolved inputs that only surface at runtime as clarification loops, malformed output or abandonment, and remove one unnecessary turn from a real flow.

   Read chapter
7. 7

   Completion rate settles the argument that novelty cannot

   Forty percent of sessions being a single "what happened last week" is not efficiency, it is a scheduled digest your product has not shipped. Build both surfaces and measure completion rate, time to result, correction effort and recovery after a failure, then bring those numbers instead of an opinion. A before-and-after flow is what makes reduced interaction cost visible to the stakeholder who asked for a chatbot.

   Read chapter