Most small businesses lose leads not because the product is wrong, but because nobody was there to answer the inquiry at 11 pm on a Tuesday. A configured AI sales agent for small business answers in that moment, asks the questions you would ask on a call, and passes the contact to you when the answers are worth your time.
Short answer: An AI sales agent is a configured software agent that receives inbound inquiries through chat or email, works through a defined set of qualifying questions, and routes the contact to a human when your qualification criteria are met. You set one up by deploying a self-hosted platform on a Linux server, giving the agent a bounded role and a fixed question sequence, connecting it to an inbound channel, and placing a human confirmation gate on the hand-off, so nothing goes out in your name without your approval.
What you need before you start
- Confirm you have a Linux server running Ubuntu or Debian with at least 2 vCPU, 4 GB RAM and 20 GB of SSD storage, which is the base size for a node that calls an external model gateway. Fully local inference, with no calls to an outside provider, needs dedicated CPU or GPU hardware sized to the model weights instead.
- Have an API key ready for the model gateway you intend to use. Smart Node supports Anthropic, Google Gemini and OpenAI as external providers, and the key goes into the local Secret store during setup.
- Decide which channel receives the first inquiry: a web page where you can paste a chat embed, or a mailbox you can attach SMTP and IMAP credentials to.
- Install Smart Node and confirm the service is running. The runbook is at /docs/install.
- Write the qualification brief: the three to five questions you ask today, the answers that disqualify a contact, and the condition that makes one worth a follow-up call.
The brief is human work, and no configuration screen replaces it. The agent executes it literally, including the parts you left ambiguous, which is why the failure table further down keeps pointing back to this document.
If you are still working out how much of the sales week an agent should carry, AI for small business covers where agents fit in a small operation before any specific workflow exists. This page assumes the decision is made and the next move is deployment.
Steps
- Write the qualification brief before you open any software. Put three to five questions on paper, the answers that disqualify a contact, and the exact condition that triggers a hand-off. You finish this step with one document that the rest of the setup only transcribes.
- Install Smart Node and confirm the service is running. The runbook at /docs/install lists prerequisites and the install sequence. You finish this step with the platform up and reachable on your own server.
- Create the sales agent and give it a role. Name the agent, write a scope statement that limits it to inbound sales qualification, and grant it the channel tool and the workflow tool. Grant nothing else. Each agent runs under its own isolated system account, so a sales agent cannot read the files or the mailbox of another agent, and a tool it was not granted cannot be called, whatever a prospect asks it to do.
- Store the credentials in the Secret store. The model gateway key goes in, and the mail passwords too if you use an email channel. Agents resolve secret references at runtime, and the values are never placed in the text the agent sends to a model.
- Connect the inbound channel. For web chat, paste the embed on the page prospects already land on. For email, configure the SMTP and IMAP connector for the address the agent will monitor. Nothing reaches the agent until the channel is enabled. Channel adapters clean up incoming input and map each conversation to a persistent thread.
- Build the qualification workflow and place the human gate. A workflow here is a state machine: each state asks one question, each answer moves the contact to a defined next state, and the agent does not reason freely about what to ask next. Encode the sequence from your brief, set the hand-off condition, then place a human-in-the-loop gate on the hand-off action. The agent pauses and waits for your confirmation before it routes the contact or sends anything outward.
- Run three test conversations before going live. One walks the full qualification path to hand-off, one disqualifies early, one asks something outside the agent's scope. Each should route exactly as your brief specifies. If one does not, change the brief or the workflow, not the test.
Steps 1 and 6 decide how the rest behaves. A vague brief produces a vague agent, and a workflow without the gate produces an agent that acts without you.
How to check it worked
Run all five checks before the first real inquiry arrives. The Event log, which records every agent action as an append-only sequence, is the ground truth here: it shows what the agent did, not what the workflow says it should have done.
- Send one message through the live channel, the chat widget or the monitored mailbox, and confirm the agent replies within the window you expect.
- Walk the qualification flow to the hand-off condition and confirm a confirmation prompt appears before anything external happens.
- Ask something outside the agent's scope and confirm it is declined or redirected, not answered.
- Run the disqualifying path and confirm the conversation ends without a hand-off.
- Open the Event log and confirm there is an entry for every step of every test conversation, with no gaps.
What goes wrong and how to tell
Warning: If the gate was not placed on the hand-off action, the agent will send messages and route contacts on its own. Run the hand-off check above before any live inquiry reaches the agent.
Four failures show up in the first week. Each has a symptom specific enough to point at the cause.
| Symptom | Likely cause | Fix |
|---|---|---|
| The agent answers questions outside sales qualification, such as legal or pricing policy questions | The role scope is too broad, or it was never written | Rewrite the scope statement so it names the permitted topics explicitly, then re-run the out-of-scope test |
| The hand-off never fires, even on good leads | The qualification condition in the workflow is set too strictly for the wording real prospects use | Compare the condition against ten real inquiries and widen the matching, or lower the threshold |
| An inquiry arrives and the agent never responds | The channel connector is off, or the mailbox credentials in the Secret store are wrong | Confirm the channel is enabled, then re-test the SMTP and IMAP credentials |
| Model costs climb above what the test volume predicted | A workflow state loops because an exit condition is missing | Find the state that can re-enter itself and add the missing exit branch |
Frequently asked questions
Will the AI agent send emails without my approval?
No. Sending an external email is one of the actions Smart Node puts behind a human-in-the-loop gate, so the agent stops and waits for your confirmation before the message leaves. If a test hand-off does not produce that prompt, the gate is missing and the agent should not see live traffic yet.
Where does the conversation data go, does it leave my server?
Conversation history, contact records and configuration stay in local databases on your own server. The one piece that leaves is the model call: when the agent uses an external gateway such as Anthropic, Google Gemini or OpenAI, the context payload for that specific request goes to that provider under its commercial API terms. Credentials from the local Secret store are not part of that payload.
Can I run this without sending data to OpenAI or Anthropic?
Yes, provided the server has the hardware for it. Local open-weight models run on your own CPU or GPU, and that configuration removes the external model egress completely. It also raises the machine requirement far above the 2 vCPU, 4 GB base, so size the server to the model you intend to run.
Can one agent handle multiple inquiries at the same time?
Yes. The agent takes conversations as they arrive, and the practical ceiling depends on your server and on how long your model provider takes to answer each turn. No fixed concurrency number is published for the platform, so measure it on your own hardware with your own workflow rather than assuming one.
Do I need a developer to set this up?
You need someone comfortable administering a Linux server, and that covers the install, OS updates, backups and monitoring. An integrator or managed service provider can take that layer. The qualification brief and the scope statement stay with you, because they encode what you know about your own leads.
What happens when the agent gets a question it cannot answer?
It declines or redirects the question instead of improvising an answer, when the role and workflow are written properly. Two settings control this: the scope statement in the agent's role, which lists the permitted topics, and the fallback branch in the workflow, which decides whether the contact goes to a human or to a general contact address.
Where to start
Deploy a private node on the server you control and point it at the channel you already answer: Set up your node, then work down the steps above with the installation runbook from the prerequisites list open beside them.