Skip to main content
Adiom Crew
Menu

AI agent guardrails

Let your AI agents act in your systems — under your rules.

In Shopify and NetSuite or anywhere else, your AI agents call the same operations your staff do, under the same rules, on the same record — with their own limits.

Book a 30-minute stack review
Operations logExampleToday · your systems
  1. 09:14approve_purchase_orderApproved

    PO 2291 to Northline Packaging, $4,820

    Rule: Purchasing v3 · approved vendor, under $5,000Ops · D. Chen

  2. 10:02reconcile_invoicesReconciled

    Order system and accounting disagreed on 3 invoices from yesterday

    A missed update. Invoices re-posted; totals now match.

    Rule: Order-to-cash v2 · nightly checkScheduled check

  3. 11:37issue_creditEscalated · declined

    $1,250 account credit for Harbor & Co.

    Declined by Maria: credit already issued on invoice 8841.

    Rule: Credits v4 · over $150 escalatesSupport AI agent → Maria R.

The question

What stops an AI agent from doing something it shouldn't? The same thing that stops a new hire: permissions, a written rule, and a record of what happened. See how it works, step by step.

An example

Bring your own AI agents. We're where they're allowed to act.

Your workflows, your bots, your scheduled jobs call the harness as a tool. Same rules, same record, per-caller permissions. An example, not a real customer:

  • 01

    The trigger

    Your returns AI agent tries to refund $340 on order #48213.

  • 02

    The rule

    Rule Refunds v4 — over $150 escalates. Maria R. gets the Slack message with the order, the customer history and the rule. Taps Approve, adds a reason.

  • 03

    The record

    The refund, its inputs, both actors — the AI agent and Maria — Refunds v4, the outcome, Maria's reason.

When your team wants to use AI to do real work inside your systems, it goes through the same rules and the same record. That's what makes it safe.

What a rule is

A rule is a sentence your finance lead can read.

Not a setting buried in six tools. One written policy per decision, enforced at the moment the action is taken — by a person in Slack, a scheduled job, or an AI agent acting on their behalf.

  • Plain English

    Written the way your policy document would say it. What you read is what is enforced; they cannot diverge.

  • Versioned

    Every change has an author and a date. The record says which version applied to every decision.

  • Tested

    Each rule is checked against your own past orders and approvals before it is switched on.

  • Shadowed before enforced

    A new rule runs silently first and reports what it would have done. You decide when it enforces.

Permissions are per caller. A returns AI agent can issue refunds under $150. A reporting AI agent is read-only. A new pilot gets nothing until you say so.

In Slack or Teams

Ask what your rules are. Get the rules.

The same command any AI agent's actions are checked against — generated from the rules actually enforced, so it's never out of date.

  • The version and the date it took effect
  • What it would have done during its shadow period
  • Who last changed it
/rules returnsExample
Returns policy · v3 · enforced since Aug 4

1. A return within 60 days and under $150 is approved
   automatically.
2. A return over $150, or past 60 days, goes to Maria R.
   with the order history and prior returns.
3. A third return from one customer in 90 days goes to
   Maria R., whatever the amount.
4. Warranty claims are exempt from the 60-day window.

Shadowed Jul 21 – Aug 3: would have blocked 9 returns;
8 were declined by hand anyway. Rule 4 added for the 9th.
Tests: 14 passing · Last changed by Maria R.

Where this runs

Sandboxed, isolated, and never used to train a model.

  • AI agent sandboxing

    Every AI agent runs in its own isolated sandbox.

  • Models and training

    We use different AI models for different tasks. Customer data is never used to train any model.

  • Per-customer isolation

    Customers are isolated from each other with an industry-standard multi-tenant architecture.

Questions we get

How is this different from Shopify Sidekick or another AI tool?

It isn't one. Sidekick works inside Shopify; your rules span NetSuite, the 3PL, wholesale and finance. What you're buying is a team and the operating layer they run for you: the operations, rules and record that your staff and your AI agents both act through. An AI agent is only as safe as what it can touch, and this is what makes it safe.

Which AI models do you use?

We use different AI models for different tasks, and customer data is never used to train any model. More on security.

Can we set a spending limit for an AI agent?

Yes — permissions and limits are per caller. A support AI agent might approve returns under $150 with a daily budget; anything over goes to a named person, the same as it does for your staff today.

Book a 30-minute stack review.

Bring the AI agent your team already wants to use, and the rule nobody's written down yet.

Book a stack review