> ## Documentation Index > Fetch the complete documentation index at: https://docs.dmly.io/llms.txt > Use this file to discover all available pages before exploring further. # Reply with AI > Add an AI Reply step to an automation so an assistant answers your contacts in your own words. The **AI Reply** step lets an assistant read the conversation and write the reply, instead of you scripting every message. You give it instructions, pick a provider and model, and choose what it is allowed to do. It works on every channel your automations run on. Two things to know before you start: * **You need a connected AI provider.** The step is useless on its own. Connect **OpenAI**, **Claude**, **Gemini** or **DeepSeek** under **Integrations** first. Until you do, the step's settings show *No AI provider is connected. Connect OpenAI, Claude, Gemini or DeepSeek under Integrations to configure this node.* In the [template library](/automation/templates), the four AI bot templates always carry a **Requires AI** chip. * **The AI Agent is a plan feature.** If your plan doesn't include it, you can still see the step and the provider tiles, but you can't add, publish or connect anything. See [What a plan without AI looks like](#what-a-plan-without-ai-looks-like) below. ## Add the step Open an automation in the [flow builder](/automation/flow-builder), or start from one of the four ready-made AI bots in the [template library](/automation/templates): **AI Customer Service bot**, **AI Lead Qualification bot**, **AI Appointment Booking bot** or **AI Class Booking bot**. Each is a three-step flow — trigger, AI Reply, end — with a prompt and tools already filled in. Add the **AI Reply** step from the palette and connect it to the step before it. It has one output, so whatever comes next follows on from the AI's turn. Open the step and use the **Prompt** tab. A new step arrives with a full customer-service prompt already written, not a blank box — edit it rather than starting over. On the **Configuration** tab, choose the **Provider** and **Model**. Use the **Playground** tab to send a test message without messaging anyone, then publish the automation. ## The prompt The **Prompt** tab is where you tell the assistant who it is and what it should do. **System prompt / instructions** carries the whole personality. The prompt a new step starts with is a working customer-service assistant covering its role, greeting, what it does, its identity, staying grounded in what it knows, its tools, its style, and what to do when the question is off-topic. Edit the parts that are wrong for your business — usually the greeting, what you sell, and the off-topic rule — and leave the rest. You can use tokens in the prompt. `{{business_name}}` becomes your workspace name, and `{{first_name}}` becomes the contact's first name, so the same prompt reads correctly for everyone. An AI Reply step never runs without a prompt. If you clear the box, the built-in customer-service prompt is used instead. The rest of the tab handles the edges of the turn: * **Fallback message (on error / empty reply)** — what gets sent if the provider errors or returns nothing. The default is *Sorry, I couldn't process that just now — a team member will follow up shortly.* * **Follow-up messages** — extra messages sent after the AI reply, every time. Good for a fixed closing line; bad for anything the AI might have already said. * **Typing delay before reply (seconds)** — a pause before the reply lands, so it doesn't arrive instantly. ## Provider and model The **Configuration** tab decides which service answers. **Provider** lists only the AI integrations you have actually connected. **Model** is a free-text box with suggestions — pick one from the list, or type any model id your provider supports. | Provider | Document search | Web search | Reasoning effort | | ---------------------- | --------------- | ---------- | ---------------- | | **OpenAI** | Yes | Yes | Yes | | **Claude (Anthropic)** | No | Yes | No | | **Google Gemini** | Yes | Yes | No | | **DeepSeek** | No | No | No | All four can call [tools and custom functions](/automation/ai-agent-tools). Those are the only four providers the AI Reply step can use. **Reasoning effort** only appears when you pick OpenAI — no other provider reads it. DeepSeek offers a `deepseek-reasoner` model, but the effort setting still doesn't apply to it. Two more settings: * **Max tokens** — roughly how long a reply can get. New steps start at 500. * **Conversation history (messages)** — how much of the chat the assistant sees. New steps start at 12, and the history is capped at 50 messages however high you set it. Only real messages count; internal notes are never shown to the AI. If the assistant answers the current message well but forgets what was said three messages ago, raise **Conversation history (messages)**. If replies get cut off mid-sentence, raise **Max tokens**. If the provider you picked is later disconnected, the step doesn't break — it falls back to the first AI provider you still have connected. ## Tools and knowledge The **Built-in Tools** tab is where you decide what the assistant can *do* — look up a contact, add tags, hand over to a human, check availability and book an appointment — and the **Functions** tab lets it call your own URLs. Both are covered in [AI tools and functions](/automation/ai-agent-tools). Two settings on that tab are worth knowing about here: * **Document search (File Search)** lets the assistant answer from your own documents. It only appears for OpenAI and Gemini; for Claude and DeepSeek you'll see *Document search is not available for this provider.* It's switched on by default for new steps, but it does nothing until you create a store and tick it — set that up under [Knowledge base](/integrations/knowledge-base). * **Web search** lets the model search the web to ground its answer. Not available on DeepSeek. ## Test before you publish The **Playground** tab runs your configuration against a test message and shows the **Reply** plus any **Tool calls** it made. Nothing is sent to a contact and nothing is saved to the [Inbox](/inbox/overview) — tool calls run against demo data, so a test that "adds a tag" doesn't touch a real contact. ## When it keeps chatting, and when it doesn't The step works out for itself whether the conversation should continue: * If the AI Reply step leads to an **End** step, a note, or nothing at all, it stays in the conversation and answers the contact's next message too. This is what you want for a support assistant. * If it leads to another step, it replies once and the flow moves on. If the assistant hands the conversation to a human during its turn, it stops immediately rather than talking over your team. ## What a plan without AI looks like If your plan doesn't include the AI Agent, nothing disappears — it stops short instead. Here is exactly what you get: The OpenAI, Claude, Gemini and DeepSeek tiles are still listed, but they don't open. Each shows a **Plan locked** badge instead of a status, and the button reads **Upgrade** instead of **Connect** or **Manage**. Selecting one shows: *The AI Agent isn't included in your plan. Upgrade to connect OpenAI, Claude, Gemini or DeepSeek.* **AI Reply** is still in the palette so you can see it exists, but adding it doesn't place the step on the canvas. You get: *The AI step isn't included in your plan. Upgrade to use the AI Agent in your automations.* Installing one of the AI bot templates, or publishing any flow that contains an AI Reply step, is refused with: *This automation uses an AI step, which your plan does not include. Upgrade your plan to use it, or remove the AI step.* If a workspace is downgraded while an AI flow is running, the flow keeps working — the AI step is skipped rather than failing. A system note appears in the conversation reading *AI step skipped — your plan does not include the AI Agent.* and the flow carries on to the next step. Your contacts won't get an error, but they won't get an AI reply either. AI suggested replies in the Inbox are a separate feature from the AI Reply step, but the same plan feature switches them off. They simply don't appear. To check what your plan covers, see [Plans and billing](/billing/plans). ## When the AI doesn't reply The flow never gets stuck on an AI step. If something is wrong, it moves on to the next step rather than failing. Three causes. *AI step skipped — no AI provider is connected* means you never finished connecting OpenAI, Claude, Gemini or DeepSeek under Integrations. *AI step skipped — your plan does not include the AI Agent* is the plan gate above. *AI step skipped — the provider returned no reply* means the provider errored or returned nothing **and** your **Fallback message** was empty — so the contact got nothing, and this note was left instead. The provider errored or returned an empty reply, so your **Fallback message** was sent. This is usually a bad model id, a provider outage, or an expired API key — check the provider under Integrations and try the same message in the **Playground**. If you have cleared the **Fallback message**, nothing is sent to the contact at all — you get the skip note above instead. There are two limits, and a step is subject to one or the other — never both. Which one applies depends on where the step sits, exactly as described under **When it keeps chatting, and when it doesn't** above. A **mid-flow** step — one whose output leads to another step — can fire three times in a run. That's a guard against a loop back into the step running up your provider bill. Past three, the step is skipped and a system note reads *AI step skipped — per-conversation AI limit reached.* If you see that note on a normal flow, check for a loop back into the AI step. A **conversational** step — one whose output leads to an **End** step, a note, or nothing — answers up to 40 turns in a run instead, and never gets the three-call limit. At 40 it quietly moves past the step: no note, no error, the contact just stops getting replies. **Document search (File Search)** does nothing unless you have created a store and ticked it in the step. If you see *No vector stores yet — create one on the provider's Integration page (Vector stores tab).*, that's the cause. Note that Claude and DeepSeek can't do document search at all. For anything else, see [Automation not triggering](/troubleshooting/automation-not-triggering) and [Common mistakes](/automation/common-mistakes). What the assistant is allowed to do, and how to add your own. Upload PDFs and web pages the AI can answer from. Add an API key so the AI Reply step can run. Start from a ready-made AI bot instead of a blank canvas.