AI Reply steps: letting the LLM handle the unexpected

Updated 2026-09-19 · 5 min read · By the Convo360 team

The problem with dead ends

Every guided flow, however well designed, eventually meets a reply it was not built for. A customer types a question in their own words instead of tapping a button, or asks something adjacent to your menu but not covered by it. Without a fallback, the flow either loops the same menu again or simply stops responding, both of which push the customer to give up and call your competitor instead, or worse, call you and add to your team's phone load.

Convo360's AI Reply block exists for exactly this gap. It is not meant to replace your structured flow, it catches what the structure misses, so the conversation keeps moving instead of stalling on a reply nobody anticipated.

Where the AI Reply block fits in a flow

The most common placement is the default or 'no match' branch of a List Menu, Buttons or Ask a Question block. When a reply does not match any expected option, route it to AI Reply instead of an error message. You can also place AI Reply as the entry point for an open-ended intent, such as a general enquiry keyword trigger, before narrowing back into a guided path.

A second, less obvious placement is mid-conversation inside a guided sequence: if a customer answers a scripted question with something the flow was not expecting, for example replying with a question of their own instead of the requested detail, an AI Reply step can address that one aside and then hand control back to the next scripted step, rather than the whole flow breaking stride.

Scoping the knowledge base

An AI Reply step draws on a knowledge base you upload as documents or URLs, giving retrieval-augmented answers grounded in your own content rather than the model's general knowledge. Keep the knowledge base attached to a fallback step narrow and relevant, for example only your FAQ document and pricing page, not your entire website. A narrow scope reduces the chance of the AI answering questions it should not, like personalised medical, legal or financial advice, which the AI should always decline and hand over to a human instead.

You also control the system prompt for this step: tell the AI its role, its tone, and explicitly what it should not attempt to answer. A well-written system prompt for a fallback step is usually three short instructions: who you are, what you can help with, and what you must not answer or promise (discounts, guaranteed outcomes, medical or legal advice).

Convo360 accepts both uploaded documents and URLs as knowledge base sources. Uploaded PDFs or text documents work well for content that changes rarely, a price list, a terms document, a service catalogue. Linking a URL works better for a page you update often, like a live FAQ page, since the AI Reply step can be set to refresh from the source rather than you re-uploading a file every time a price changes. Whichever source you choose, keep a habit of checking it after any business change, a new price, a paused service, a changed address, so the fallback never repeats outdated information confidently.

Setting temperature for fallback replies

Convo360 lets you set temperature per AI agent. For a fallback step answering factual questions about your business, a low temperature keeps answers consistent and grounded in the knowledge base, so the same question asked twice gets essentially the same answer. Save a higher temperature for creative tasks like drafting marketing captions, not for a step that customers rely on for accurate information about price, availability or policy.

Bringing the contact back to the menu

1. After the AI Reply block answers, add a Buttons block offering 'Back to menu' or 'Talk to a person'.
2. Route 'Back to menu' with a Go To block pointing at your main flow entry point.
3. Route 'Talk to a person' through live-chat team inbox handoff so a human takes over and the bot pauses automatically.
4. If the AI cannot find an answer in the knowledge base, set the system prompt to say so honestly and offer the human handoff immediately, rather than guessing.

Parking the conversation instead of looping forever

Some conversations genuinely need open-ended back and forth, a jewellery customer comparing designs, or a coaching enquiry asking follow-up questions. Rather than forcing every reply back through a menu, you can let AI Reply steps chain for a few turns, then automatically offer the menu again after a set number of exchanges or once a clear intent (like 'book' or 'buy') is detected, using a Condition block on the reply text.

Set a sensible upper limit, for example after five or six back-and-forth turns without a clear next action, have the flow gently offer the menu or a human handoff anyway. Open-ended AI chat with no exit point tends to wander, and a customer who genuinely wants to book or buy should not have to ask for a human explicitly.

Fallback for voice calls, not only chat

The same idea applies to AI voice agents. A caller who says something outside the expected menu of DTMF or speech options should not hear a dead-air pause or a repeated prompt. Route unclear speech input to the same kind of AI Reply logic, scoped to a voice-appropriate knowledge base and a shorter, spoken-style system prompt, and offer a clear path to transfer to a human if the caller asks for one or the AI cannot help after a couple of attempts.

Step by step

1. Identify your no-match branchesLook at every menu or question block in your flow and check what happens on an unmatched reply.
2. Add an AI Reply blockRoute the no-match branch into an AI Reply step instead of an error message.
3. Attach a scoped knowledge baseUpload only the documents relevant to this fallback's purpose, such as an FAQ or pricing sheet.
4. Write a clear system promptState the AI's role, tone and what it must decline to answer or advise on.
5. Set a low temperatureKeep factual fallback answers consistent by choosing a conservative temperature.
6. Add a return pathOffer 'Back to menu' and 'Talk to a person' buttons after every AI Reply.
7. Review Flow Logs weeklyCheck what questions triggered the fallback and add common ones as proper menu options.

Example conversation

Frequently asked questions

Related reading

Available across India — see WhatsApp & AI calling automation by state and city.