Most roundups of the best AI automation tools for small business rank products against each other. This checklist works the other way: eight capabilities to verify before you commit to any platform, grouped into three areas covering where your data lives, what work the platform actually handles, and which actions it must stop and ask you about. Each check is written as a question you can put to a vendor, because the answer tells you more than a feature page does.
Short answer: Working through the eight checks takes about ten minutes. At the end, you have a concrete set of requirements to test any platform against, covering where your data goes, what routine work the platform handles, and which actions it must hold for your approval. A platform that clears all eight runs the day-to-day work of a small team without acting outside the limits you define.
Area 1: Data ownership
A hosted platform can be well engineered and still keep your documents, conversation history, and task records in a database someone else operates. Single-tenant hosting means the opposite: the task queue, the documents, conversation history, and the configuration database run on infrastructure you own or rent, not on a shared system serving other customers. Two checks tell you which of those you are buying.
- Stored data. Ask where documents, conversation history, and task records live. The answer to look for is a server you own or rent, in a database your team controls, not a vendor's shared multi-tenant platform.
- Credentials. Ask how API keys, mailbox passwords, and third-party tokens are stored. They should sit in a local secret store and be resolved at execution time by the tool that needs them, and they should never appear in the text sent to a model.
Area 2: Operational coverage
Reaching your inbox is the floor, not the ceiling. A platform that only answers single-turn questions in a chat window leaves the routing, the follow-up, and the record keeping to you. Check the channels your team already uses (web chat, email over SMTP and IMAP, Telegram bots, voice), then check that multi-step work runs from assignment to completion through a workflow rather than a person nudging each step. Agents on the same node can also be configured with different models matched to their roles, so a drafting agent and a research agent need not share one setting. For a fuller picture of what a multi-agent system covers end to end, see AI workflow automation.
- Channels. Email is the minimum. Web chat, Telegram bots, and voice cover the rest of the places a small team already works.
- Multi-step tasks. The platform should assign work, track progress, and execute repeatable workflows to completion without manual management of every step.
- Document versioning. Documents should carry version history, so an agent acting on a policy, contract, or template reads the current version rather than last quarter's copy.
Area 3: Human oversight
Oversight is where vendors get vague, so make it specific. A human-in-the-loop gate stops irreversible or externally visible actions before they happen: sending external email, publishing content, deleting records, running privileged system commands, and spending money all wait for explicit confirmation. Behind that, agents run in isolated system accounts under declared scopes with least-privilege tool grants, which is what keeps the agent that drafts invoices out of the payroll mailbox.
- Approval gates. Ask for the list of actions the platform will not take without a person confirming. External messages, deletions, and spend belong on it.
- Access scope. Ask what each agent can reach. The answer should be a defined scope: only the tools and data that role needs, not the whole system.
- Cost model. Ask how pricing changes when you add a colleague. A single node serves the whole team with no per-seat fee, and the baseline node runs on 2 vCPU, 4 GB RAM, and 20 GB SSD when reasoning routes to a remote model gateway.
What people usually miss
Warning: Self-hosted does not automatically mean fully private. In a standard configuration, tasks that need complex reasoning send their prompt to an external model provider.
Where data is stored and where reasoning happens are two separate questions, and most evaluations fold them into one. On a self-hosted platform the storage stays local: relational state, the event log, the document repository, and vector memory all live on your server. But when a task routes to an external frontier model gateway such as Anthropic Claude, OpenAI GPT, or Google Gemini, the context payload for that specific request travels over the network to that provider, under that provider's commercial API terms, which are typically zero-retention.
Fully local privacy is achievable, and it is a configuration decision rather than a claim on a landing page. It requires local model inference running on dedicated CPU or GPU hardware provisioned on the host. That pushes the hardware requirement well above the 2 vCPU and 4 GB RAM a gateway-routed node needs, and the question to put to any vendor is direct: where does the reasoning happen, and under what API terms?
The trade-off you accept with self-hosting is operational. Host security, operating system updates, monitoring, backups, and recovery become your responsibility or your managed service provider's. You also keep paying for infrastructure, maintenance, and per-token model calls. Running private AI removes per-seat subscription fees, not running costs.
Frequently asked questions
Do I need a technical person to set this up?
Yes, or someone who plays that role. Deploying and administering a Linux server on Ubuntu or Debian is technical work: provisioning the host, keeping the OS updated, watching it, and running backups. You can do that in house or hand it to a managed service provider. What you cannot skip is having someone accountable for the machine.
What does it cost to run after I have it installed?
Two lines on the bill: infrastructure and model usage. You choose and pay for the server, and 2 vCPU, 4 GB RAM, and 20 GB SSD is the baseline when reasoning routes to remote gateways. There is no per-seat subscription, so adding a team member does not add a licence. Model gateway calls are metered per token by the provider, and if you want inference with no external calls at all, you pay for the CPU or GPU capacity that local model weights require.
Will the AI keep interrupting me to approve things, or can it work on its own?
Routine work proceeds on its own, and the interruptions are limited to the gates you would want a person at. Agents operate under declared scopes and bounded workflows, and they stop for explicit confirmation before sending external email, publishing content, deleting records, running privileged system commands, or spending money.
What is the minimum server size I need to get started?
2 vCPU, 4 GB RAM, and 20 GB SSD on a Linux host running Ubuntu or Debian, which is the base node when remote model gateways handle the reasoning. If you intend to run models locally instead, the requirement is dedicated CPU or GPU compute sized to the weights you choose, and that decision drives the hardware more than anything else on this list.
Next step
The eight checks exist because this platform was built to pass them, and the only way to confirm that is to run the list against an instance you control. Start with Get started to review the architecture and deploy a node on hardware you own.