refactorPARTNERS
All Insights

Security & Compliance

Grok Bot Can Now Spend Money. Your Firm's Agents Need Better Rules Than That.

Refactor Partners||
4 min read

Agents that act, not just answer

Consumer AI took another step from "assistant" to "agent" this fall. xAI's Grok Bot, which recently gained access to users' phone calendars, reminders and contacts, can now complete online purchases. Users connect a payment method and the bot checks out on their behalf.

How it decides when to ask permission is the interesting part. According to xAI, every action passes through an automated review, plus rules the user writes. "Allow automatically" rules let the bot proceed. "Ask first" rules trigger an approval prompt, and when both match, "Ask first" wins. That's a sensible default.

There's a catch in the details, though. A payment card saved to the bot's built-in vault fills in at checkout without a separate approval each time. If you want to approve every payment, you have to remember not to save the card. And xAI is clear that completed actions can't be undone.

None of this is unique to Grok. It's simply the most visible example of a question every firm deploying agents now faces.

The real question is what happens when no one is watching

When a human associate drafts an email to a client, there's a natural pause before it's sent. When an agent drafts it, that pause only exists if someone designed it in.

The failure modes that worry us aren't dramatic. They're ordinary: an email goes to the wrong client because two names were similar, a vendor invoice gets paid twice, a document with another client's data gets attached to a reply. Each is an annoyance when a human catches it. Each is a real problem when nothing stands between the agent's decision and the outcome.

Sort actions by whether they can be undone

The most useful framework we've found is simple. Before an agent goes live, list every action it can take and put each one in one of three buckets:

  • Reversible and internal. Drafting a document, tagging a file, updating an internal tracker, summarizing a call. Let the agent act freely, and review samples after the fact.
  • Reversible but visible. Scheduling a meeting, updating a CRM record a client can see, posting to a shared channel. Let the agent act, but log every action somewhere a person checks daily.
  • Irreversible or external. Sending anything to a client, moving money, filing anything with a regulator, deleting records. Require explicit human approval, every time, with no "always allow" option.

That last bucket is where Grok's vault behavior would worry us in a business setting. Convenience features that quietly move an action from "ask first" to "just do it" are exactly the settings that need to be locked down by policy, not left to individual preference.

Make approval fast, or people will route around it

There's a reason consumer products lean toward fewer approval prompts: people hate them. If your approval step is slow or confusing, staff will find ways to skip it, which is worse than having no approval step at all, because now you think you're protected.

Good approval design for a professional services firm looks like this: the reviewer sees exactly what the agent intends to do, why, and what it's based on, in one screen. Approving takes one click. Rejecting with a correction takes two. And the queue lives where people already work, not in a separate dashboard they have to remember to open.

This is the same principle behind human-in-the-loop review as a trust mechanism. Approval isn't a speed bump. Done well, it's what makes people comfortable letting the agent do more over time.

Build the off switch first

One more lesson from this round of agent launches: every production agent needs a way to stop it immediately. That means a single control that halts all outgoing actions, which anyone on the responsible team can use without asking permission. You will rarely need it, but when you do, you'll need it in seconds, not after a support ticket.

The firms getting the most from agents aren't the ones with the most permissive settings. They're the ones that defined the boundaries clearly enough that everyone, including clients, trusts the agent to stay inside them.

agentic workflowsAI securityhuman in the loopapprovalsAI governance