Most small business owners who try AI tools end up with a stack of disconnected apps and still copy-pasting information between them.
Short answer: AI integration for small business means agents connected to the channels your business already runs on: email, messaging, and internal tasks. It takes a Linux server running the agent platform, at least one connected channel, and human approval gates on any action that sends external messages or modifies records.
What you need before you start
Three classes of prerequisites have to be in place before any configuration begins: the infrastructure, the credentials, and one workflow described narrowly enough to test.
- Rent or set aside a server or VPS running Ubuntu or Debian with at least 2 vCPU, 4 GB RAM, and 20 GB SSD storage. That is the minimum when agents call external model gateways. Running model inference entirely on your own hardware needs more machine than this.
- Get an API key from an external model provider such as Anthropic or OpenAI, unless you are provisioning local model weights on dedicated CPU or GPU hardware.
- Choose the one workflow you will connect first: incoming email handling, a Telegram request channel, or the browser web chat panel used inside the team.
- Write down which actions in that workflow must be confirmed by a person before they run: sending external email, publishing content, deleting records, spending money.
- Confirm who administers the server. Deploying and running Linux takes technical proficiency, so either someone on staff has it, or a managed service provider handles deployment and upkeep.
Steps, in order
- Pick one workflow to start with. Name the channel, the kind of input that arrives there, and the output the agent should produce. Do not connect a second workflow in the first deployment. You finish this step with a one-page scope you can test against.
Tip: Starting with one agent and one channel lets you verify the approval gates and task routing before the surface area multiplies.
- Provision the server. Ubuntu or Debian, 2 vCPU, 4 GB RAM, 20 GB SSD. You finish this step with a running Linux host you can log into.
- Install the platform. Follow the installation runbook at /docs/install. You finish this step with the platform running and an administrator account.
- Create the first agent and give it a role. Define the tools it is allowed to use. Each agent runs in its own isolated system account with only the capabilities explicitly granted to it, and that boundary is enforced by the operating system rather than by a prompt. You finish this step with one agent that can do exactly the job in your scope and nothing else.
- Connect one channel. SMTP and IMAP for email, a Telegram bot token for Telegram, or the browser web chat panel for internal use. The channel adapter sanitizes incoming input and maps each conversation to a persistent thread. You finish this step with one real inbox or chat reaching the agent.
- Configure human approval gates. For every action in your scope that sends external messages, publishes content, or deletes records, require explicit confirmation. You finish this step with a system that cannot take those actions until a person approves them.
- Run a test end to end. Send a real message through the connected channel from a separate account, and watch the whole path: pickup, task creation, and the pause for approval before any outbound output. You finish this step with evidence rather than an assumption.
How to check it worked
- A test message sent through the connected channel is picked up and turned into a task automatically, with nobody moving it by hand.
- An action that would send an external message or modify a record pauses and presents an approval request before it runs. The agent does not route around that gate.
- The event log shows each agent action in order, so you can read back what happened, in what sequence, and what the agent was waiting on.
What goes wrong and how to tell
Warning: An agent whose tool grants are set too broadly can attempt actions outside its intended scope. Review the tool list for each agent before you connect it to a live channel, and grant only the tools that workflow actually requires.
Per-token costs climb faster than expected. When an agent reasons with an external frontier model, each reasoning call routes to that provider's API and is billed per token. There is no per-seat fee, but volume is volume. Watch gateway usage from the first week, before the workflow is carrying real business.
Messages drop silently. A wrong SMTP credential or a misconfigured Telegram bot token can leave messages vanishing with no visible error in the interface. Send a test message from a separate account immediately after connecting the channel, and repeat it after any credential change.
Self-hosting is ongoing work. Operating system updates, backups, system monitoring, and recovery are yours or your provider's responsibility, not the vendor's. Plan who does them before the system carries anything the business depends on.
Frequently asked questions
Does my business data go to OpenAI or Anthropic when I use this?
Partly, and it depends on how you configure inference. The orchestration runtime, task queue, documents, conversation history, and configuration database stay entirely on infrastructure you own or rent. When an agent reasons with an external frontier model, the specific context payload for that request travels to that provider and is covered by their commercial zero-retention API policy. Configuring local model weights keeps that payload on your hardware, and calls for dedicated CPU or GPU capacity beyond the base minimum specifications.
What does it actually cost to run AI agents for a small business?
There are no per-seat licensing fees, so the bill is hosting plus inference. You pay for the server or VPS sized to the minimum specifications (or larger, if you run local models), plus per-token fees to whichever external model gateway you configure. The exact gateway cost depends on request volume and the model you select. This is infrastructure-sized cost, not zero cost: hosting, upkeep, and tokens all remain.
Do I need a developer or IT person to set this up?
Yes, or access to someone who fills that role. Deploying and administering a Linux server takes technical proficiency. Most small businesses either have that skill on staff already or engage a managed service provider for deployment and ongoing host maintenance.
What happens if an agent sends the wrong email or takes the wrong action?
Sending external email is one of the actions that requires explicit human confirmation before it executes, so the agent cannot send one without passing that gate. The same applies to publishing content, deleting records, privileged shell commands, and monetary spend. If you want to know what an agent attempted regardless of the outcome, the append-only event log records every action and state change in order.
Can my whole team use it, or do I pay per person?
A single node serves the entire team, and there are no per-user licensing costs. Headcount does not change the license line because there is no license line; you scale the server, not the seat count.
Set up your node
When your server meets the requirements above, the installation runbook at /docs/install walks through the process from a clean Ubuntu or Debian instance. Set up your node.