Setting Guardrails for Your AI Agent in BearScope
You decide what your AI agent is allowed to do. Here's how to configure guardrails that keep it safe.
An AI agent that can act in your systems needs limits, and those limits should be yours to set — explicit, readable, and enforced before anything runs. In BearScope, you decide exactly what your AI agent is allowed to do. This is a how-to for setting guardrails that keep it safe without making it useless.
We'll go through the four things you configure, in the order you should configure them: allowed actions, value limits, escalation rules, and what happens when something's uncertain.
AI agent guardrails setup starts with allowed actions
The first decision is the simplest and the most important: what is the agent allowed to do at all?
Every capability the agent could use — reply to a customer, tag a conversation, issue a refund, start a return, post to your team chat — is off until you turn it on. Connecting a tool does not grant action in it. You enable actions one at a time, deliberately.
A sensible starting point for most teams:
- Turn on reading everywhere. The agent should see the full picture — orders, history, policy, conversation context. Reading carries no risk and makes every decision better.
- Turn on low-stakes actions first. Tagging, drafting replies for a person to send, opening a status trace. These are easy to allow and easy to trust.
- Hold the consequential actions — refunds, replacements, cancellations — until you've set the limits below.
Start narrow. It's far easier to widen a confident agent's permissions than to claw back an over-eager one's.
Set value limits on the actions that move money or inventory
Some actions are fine in principle but need a ceiling. A refund is allowed; a $2,000 refund should probably involve a person. So for any action with a cost, you set a value limit — the threshold below which the agent acts on its own and above which it must escalate.
You can set these per action and per situation. A common pattern:
| Action | Agent acts alone | Routes to a person |
|---|---|---|
| Refund | Up to your per-refund cap | Above the cap |
| Replacement | In-policy, standard items | High-value or final-sale items |
| Goodwill credit | Small, within budget | Above the credit limit |
| Cancellation | Standard orders | Already-shipped or custom orders |
The limit is checked before the action runs, every time, on real values pulled from your systems. The agent can't talk its way past a cap, and it can't act on a guessed amount — the check enforces the number you set against the actual order data.
A guardrail you configure but can't see enforced is a hope. In BearScope the check runs before every action and shows up on the receipt — so you can prove the limit held, not just that you set it.
Write escalation rules for the cases that need a person
Value limits handle "how much." Escalation rules handle "what kind." Some situations should go to a person regardless of dollar amount — and you say which.
Good escalation rules to set early:
- Sentiment. A clearly angry or distressed customer routes to a person, even for a small request.
- Sensitive topics. Anything touching legal, safety, billing disputes, or account security.
- Repeat contacts. A customer back for the third time on the same issue needs a human eye, not another automated pass.
- Policy edges. Requests that fall outside clear policy — the agent shouldn't improvise on gray areas.
When a rule fires, the agent doesn't just stop — it hands off with full context: everything it sensed, what it was about to propose, and the conversation so far, so the person picks up mid-stride and the customer never repeats themselves.
Decide what happens when something's uncertain: fail closed
The last setting is the most important and the easiest to get wrong elsewhere. When the agent is unsure — missing data, an ambiguous request, a check it can't complete — what should it do?
BearScope's default is fail closed: when in doubt, don't act. The agent stops and routes to a person rather than guessing and acting anyway. This is the opposite of the risky default, where uncertainty quietly resolves into action. A fail-closed agent's worst case is "a person had to step in," not "the agent did something irreversible on a hunch."
You keep this on. It's the safety net under every other guardrail: if a value limit, an escalation rule, or the data itself is unclear, the action doesn't run. A blocked action is recoverable. A wrong refund or a mistaken cancellation is harder to undo.
Putting it together
Guardrails in BearScope are four decisions you own: which actions are allowed, the value limits on each, the situations that always go to a person, and the rule that uncertainty means stop. Set them narrow, watch the receipts, and widen as your confidence grows. Every action the agent takes is checked against these limits before it runs and leaves a receipt showing the check held — so you're never trusting on faith.
That's the whole idea: your AI agent does real work inside limits you control. Configure it on the product page, read how the checks are enforced, or book a walkthrough to set guardrails against your own policies.
See it on your own conversations.
Bring your busiest day. We'll score every conversation in it.
Book a walkthrough →