Before you commit to any AI assistant for your business, nine questions determine whether the platform fits your operations or creates new problems you did not sign up for.
Short answer: Applying this checklist to a candidate platform takes under two hours and returns a verdict in four areas: data ownership, cost structure, oversight controls, and infrastructure responsibility. There is no single best AI assistant for business, only the one whose answers to these nine questions you can verify against your own operations. For the wider context of how small businesses are structuring AI adoption, see the small business AI overview.
The checklist is vendor-neutral by design. Every item names something you can check in a configuration file, a contract, or a hardware quote rather than a claim you have to take on faith.
Data and privacy: where your records live and where prompts travel
1. Does your data live on your infrastructure or the vendor's?
A self-hosted platform keeps conversation history, documents, and relational records on hardware you own or rent. A vendor-hosted platform keeps them in the vendor's environment. The answer decides who can read your operational data, who is accountable for it, and what happens to your access if the vendor changes its terms or its pricing. Ask which party holds the database account, not which region the status page lists.
2. When the AI reasons through a task, where do those prompts go?
Most self-hosted platforms still route high-complexity reasoning to an external model gateway such as Anthropic Claude, Google Gemini, or OpenAI GPT. The runtime and storage stay on your server, but the specific context payload in that request crosses the network to the provider. Three things to ask:
- Is that egress explicit and visible in the system configuration?
- Is it metered, so you can see what volume leaves and what it costs?
- Is it covered by the provider's commercial zero-retention API policy?
Then ask to see the setting that turns it off.
Warning: A claim that zero data ever leaves the server is only honest for a deployment running exclusively local model weights with no outbound network path. Any platform that implies total isolation without stating that condition is describing a configuration you may not have.
3. Are your credentials and API keys kept out of the AI's prompt context?
Third-party API keys, mailbox passwords, and database tokens should sit in a local secret store and be resolved by local tool wrappers at the moment of use, never injected into a model prompt. If the vendor answers "we encrypt them," follow up with "are they ever passed as text into a model call?" The first answer is about storage. The second is the one that determines your exposure.
Oversight and control: what the system must ask before it acts
4. Does the platform require human confirmation before irreversible actions?
These five classes of action cannot be undone or cannot be recalled once a third party sees them:
- Sending an external email
- Deleting a record
- Publishing content
- Executing a privileged shell command
- Authorizing monetary spend
A platform with human-in-the-loop gates intercepts those actions and holds them until a person confirms. One without them acts on your behalf and reports afterwards. Ask for the list of gated actions, not a yes or no.
5. Can you limit each agent to only the tools and permissions it actually needs?
An agent that answers customer questions has no reason to reach billing records or a shell. Least-privilege tool grants, declared per agent and enforced by the platform, mean that a confused or manipulated agent cannot reach past its own scope. Ask whether those boundaries are enforced outside the model or merely requested inside a prompt, because a prompt instruction is a suggestion and a permission check is not.
6. Is every action the AI takes written to a permanent audit log?
An append-only event log records each state change and agent action, so you can reconstruct what the system did, when, and in response to which task. Without one you are operating on trust alone, and the gap shows up during an incident rather than before it. Ask whether the log can be edited or deleted by an administrator, because an audit trail that can be rewritten is not an audit trail.
Cost and infrastructure: what you pay for and who maintains it
7. Does the licensing cost grow with your headcount?
Per-seat licensing scales with every hire. A node-based license serves an entire team from one deployment with no per-user fee. At ten people the two models land in a similar range. At fifty, the difference is a standing budget line. Ask for the pricing model and a three-year projection, not the current quote.
8. What are the minimum infrastructure requirements to run the platform?
A remote-gateway configuration runs on 2 vCPU, 4 GB RAM, and 20 GB of SSD storage. Running local model inference with no external API calls requires dedicated CPU or GPU compute sized to the model weights you choose. Decide which mode you are pricing before you collect hardware quotes, because the two answers differ by an order of magnitude.
9. Who is responsible for maintaining the host server?
Self-hosting transfers these obligations to your organization or to a managed service provider you engage:
- Host security and operating system updates
- System monitoring
- Backups and recovery procedures
That is an allocation decision, not a defect. If nobody on your team holds that capacity today, budget for a managed host or an integrator before deployment day.
What people usually miss
Tip: The hybrid inference boundary is not a disqualifier, it is a fact to manage. What matters is whether egress to external model gateways is explicit, metered, and governed by the provider's zero-retention commercial terms, not whether it exists at all. A vendor that names the boundary plainly is easier to trust than one that implies isolation without stating its conditions.
Tip: Per-seat licensing is invisible until the team grows. At five users it rounds to nothing. At twenty-five it reappears in every quarterly budget review. Test the licensing model against your one-year and three-year headcount plans, not today's roster.
For a broader look at how small businesses are structuring AI adoption, see the small business AI overview.
Frequently asked questions
Can an AI assistant for business run without sending any data to the cloud?
Yes, but only in a deployment configured with exclusively local model weights and no outbound network path. In the common hybrid configuration, the runtime and storage stay on your server while the context payload for complex reasoning travels to an external model gateway under that provider's commercial zero-retention API policy. Treat full isolation as a configuration you choose and verify, not a default you assume.
What is the minimum hardware needed to run a private AI assistant?
A node using remote model gateways runs on 2 vCPU, 4 GB RAM, and 20 GB of SSD storage. Fully local inference changes the answer completely, because it requires dedicated CPU or GPU compute sized to the model weights you intend to run. The hardware follows the model choice rather than a fixed minimum.
How do I stop an AI assistant from sending emails or taking actions without my approval?
Require human-in-the-loop gates at the platform level rather than in workflow configuration. Sending external email, deleting records, publishing content, running privileged shell commands, and authorizing spend should be intercepted before execution and held until a person confirms. Ask specifically whether the gate is architectural, meaning no workflow can bypass it, or a per-workflow setting that someone can forget to enable.
Does self-hosting an AI assistant eliminate all ongoing costs?
No. Self-hosting removes per-seat licensing, so a team of fifty pays the same software cost as a team of five, but infrastructure hosting, server maintenance, and per-token model gateway costs remain. Ask for a metered estimate of gateway usage at your expected task volume and add it to the hosting line before you approve the project.
How many people can use one AI node?
A single node serves an entire team with no per-user licensing, so the practical ceiling is compute capacity rather than a seat count. Heavier concurrent workloads and local model inference raise the hardware requirement. Adding users on remote model gateways usually does not.
Next step
If the checklist above describes the platform you want, the requirements are stated up front and the path is documented: Deploy a private node.