Safety Built Into Every Agent
Agents do the work. You stay the boss. Here's exactly how we keep it that way.
Why this page exists
An agent that can act needs rules that can't be talked around
A chatbot that says something wrong is embarrassing. An agent that does something wrong (books a job you can't do, promises a refund, emails the wrong client) costs real money and real trust. So safety isn't a feature we bolt on at the end. It's how every agent is designed from day one.
Courts and regulators are paying attention, too. In one widely reported 2024 case, a Canadian tribunal held an airline responsible for incorrect refund information its own website chatbot gave a customer (Manatt summary of Moffatt v. Air Canada). The lesson for every business: what your AI says on your behalf, you own.
The guardrails
Eight rules every agent we build follows
1. One job
Each agent has a single written job with clear limits. Narrow agents are easier to test, trust and fix.
2. Least access
Agents only reach the tools and data their job requires. A review agent never sees your bank account.
3. Human approval
Anything that spends money, makes a commitment or touches a sensitive subject waits for a person to say yes.
4. Instant hand-off
Customers can always reach a human. Agents escalate the moment something is outside their job.
5. Honest about being AI
We recommend agents say they're automated assistants. They never pretend to be a person.
6. Everything logged
Every action is recorded so you can see what happened, when, and why.
7. Tested first
Agents practice on realistic scenarios, including tricky and adversarial ones, before they meet customers.
8. Off switch
Any agent can be paused immediately, and work falls back to your normal process.
Approval checkpoints
You decide where the line is
During the Blueprint we sort every action an agent might take into three buckets. You can move things between buckets any time.
| Bucket | Examples |
|---|---|
| Agent handles | Answering routine questions, booking open slots, sending reminders, sorting the inbox, logging calls |
| Agent drafts, you approve | Quotes, refunds, replies to negative reviews, published content, unusual discounts |
| Human only | Legal, tax or medical advice, hiring decisions, negotiations, anything you name |
Your data
Handled like it's ours, because we run businesses too
- Agents get the minimum data their job needs, and nothing else.
- We prefer running agents on accounts you own, so you keep control of your data and can revoke access any time.
- We review each AI and software provider's data terms with you during the Blueprint, so you know exactly where information goes.
- Credentials are stored with the access controls of the tools themselves, never pasted into an agent's instructions.
- Calls and texts follow consent rules: agents only contact people who've asked to hear from you, and honor opt-outs.
The pattern in that data is the opportunity: lots of interest, low trust in fully autonomous agents. Narrow agents with human checkpoints are how you get the benefit without the risk. More on telling real agents from hype.
Want to see the guardrails for your business?
Every Blueprint includes a written approval map for each agent: what it handles, what it drafts and what stays human.
Start building How we buildFAQ
Frequently asked questions
Can an AI agent spend money or make promises on my behalf?
Only if you allow it. By default, anything that spends money, offers a refund or discount, or makes a commitment is drafted by the agent and waits for a person to approve it.
Will customers know they are talking to AI?
We recommend every customer-facing agent say it is an automated assistant, and agents never claim to be a person. Customers can always ask to reach a human.
What data can an agent see?
Only the data its single job requires. Access is set during the Blueprint, preferably on accounts you own, and can be revoked at any time.
What if an agent gets something wrong?
Every action is logged so the issue can be traced, the agent can be paused instantly, and Agent Care includes tuning the agent so the same mistake does not repeat.
Do your agents give legal, tax or medical advice?
No. Those topics are set as human-only. Agents can gather information, schedule appointments and organize documents for the professionals who give that advice.