← The journal
Analysis / AI safety

The AI Rehearsal Button: See What an Agent Would Do Before It Acts

A scripted prototype shows how one change in a customer-support case can change an AI agent’s proposed action. Support teams can request a private walkthrough and discuss a pilot.

By Henry Chimanzi | FrontierAI.fi

A customer writes, “I ordered two lamps. One arrived broken. Can you sort it out?”

You ask an AI agent to handle it. What have you just authorised?

The agent could draft an apology, send an email, request a replacement for one lamp or refund the whole order. You already know what “sort it out” means in your business. The agent has to work it out from the information and rules you give it.

I built a prototype around a simple idea. Before giving an AI agent permission to act, give the team a way to rehearse its decisions.

I call it the AI Rehearsal Button.

How the Rehearsal Button works

Give the agent a fictional or anonymised case and the rules it should follow. In the proposed product, it works through the case using simulated tools. A simulated email tool shows what it would send and to whom. A simulated order tool shows which item it would replace. A simulated refund tool shows the amount it would request and why.

Then change one fact and run the case again. Put the proposed actions side by side and highlight the first decision that changed. A reviewer can see whether the agent responded to the fact that matters without reading two full transcripts.

One broken lamp, two decisions

Imagine a retailer with a simple policy. Acknowledge the complaint and log the case. Request a replacement for the affected item only after the warehouse confirms the damage.

In the first rehearsal, the warehouse confirms that one lamp arrived damaged. The proposed actions are to draft an acknowledgement, request one replacement lamp and log the case.

In the second rehearsal, the warehouse is still checking. The proposed actions are to draft an acknowledgement, hold the replacement and log the case.

The first difference is the replacement decision. If an agent requests a replacement in both cases, it has acted too soon in the second. If it waits in both, it has missed an authorised replacement in the first.

What the prototype shows today

The private prototype displays both fictional cases side by side. Click “Run both rehearsals” to reveal the proposed actions and see the first decision that changes.

The prototype is scripted. It illustrates the review experience without running a live AI agent, sending emails, changing orders or processing refunds.

What a pilot would test

I’d start with one customer-support workflow, such as damaged deliveries. The support team would supply a small set of anonymised or fictional cases and agree what should happen in each one. For each case, we’d change one fact and decide in advance whether that change should affect the agent’s actions.

The next step would be to connect a test agent to simulated tools and record what it proposes. A support lead could then compare ordinary transcripts with a rehearsal view that highlights changed decisions. We’d measure review time and count which policy mistakes each view helps the reviewer catch.

I’d also include cases where a changed detail should have no effect. If the rehearsal flags every difference as important, it gives supervisors another queue to read.

Why this matters for frontier AI

In my earlier Missing Question Test, I asked whether an assistant could find the fact that changes its advice. The Rehearsal Button takes that question into a workflow. When the fact changes, does the agent propose to send, replace, refund or escalate differently?

A rehearsal gives teams a way to examine proposed actions before going live. In production, the software still needs to enforce permissions and approval rules when the agent acts.

Request a private demo

If your team is testing an AI support agent, email [email protected] with the subject “AI Rehearsal Button”. Tell me which action you’d want to rehearse before letting the agent handle it live.

I’m available for a 15-minute walkthrough and open to discussing a pilot, product partnership or licensing.

Sources & further reading

  1. FrontierAI.fi — The Missing Question Test
  2. FrontierAI.fi — The One-Word Test

Sources reviewed 10 Nov 2026. Analysis and reading guidance are editorial interpretation; provider claims are not independently benchmarked.