The main safety question is not whether the AI is smart enough, it is whether it can act without you. An agent that sends, posts, or replies unattended can put a wrong message in front of a customer permanently, and that is the failure that actually costs businesses money. The safest setup for a small business is an agent that drafts and waits for approval, which is how Dutch works: nothing sends or publishes until you click approve.
| Risk | What it looks like | How to protect yourself |
|---|---|---|
| Acting without review | A tone-deaf reply posted publicly on a 1-star review | Require human approval before anything sends or publishes |
| Made-up details | An invented price, policy, or promise in a customer message | Review outputs, and prefer tools that work from your real site content |
| Too much account access | A tool with permission to post or email as you, unsupervised | Grant the minimum access, and prefer tools that work without integrations |
| Your data being reused | Your business content used to train models or shown elsewhere | Read the privacy terms before connecting anything |
Every AI system makes mistakes at some rate. The question that decides your exposure is what happens when it does. If a mistake becomes a draft you delete in two seconds, the error rate is close to irrelevant. If a mistake becomes a public reply under your business name, a single bad output can undo months of reputation work. This is why approval-first design matters more than any accuracy claim in a sales page.
Five questions that separate serious tools from risky ones: Can it send or publish without me approving it, and can I turn that off? What access does it need, and what is the minimum it will run with? Where does it get its facts about my business? Is my data used to train anyone's models? Can I see exactly what it is going to send before it goes?
If a vendor is vague about the first question, that alone is a reason to walk away.
Dutch drafts work and waits for your approval. Every message, reply, and post is reviewable and editable before it goes out, and you can change any wording before approving. It works without connecting any accounts to start, learning your business from your public website. Connectors exist for tools like Google, Slack, HubSpot and LinkedIn, but they are optional and you grant them individually.
Approval is a safety feature, and it is also a real cost: it means Dutch does not run completely hands-off. If your goal is to remove yourself from a process entirely, an approval step will feel like friction, and you may prefer a fully automated tool. We think that trade is wrong for anything customer-facing, but it is a genuine trade-off and you should choose it knowingly rather than discover it after buying.
Some tools can, and that is the setting to check first. Dutch cannot: every draft waits for your approval before it sends, and you can edit the wording before approving.
This is a genuine risk with any AI system. Two protections matter: the tool should work from your real published content rather than guessing, and a human should read customer-facing output before it goes. Dutch does both, reading your actual website and requiring approval.
It should not need any to start. Dutch works from your public website with no integrations required. Optional connectors exist for specific jobs, and you grant each one individually rather than all at once.
Use extra care. Keep patient specifics out of anything the agent handles, keep every output under human review, and confirm your own compliance obligations. A review reply that references treatment details is a problem regardless of who wrote it.
Describe the job in plain English. Dutch scopes it and prices it on the spot. 30 days free, no contract.
Build an agent →