On 11 October 2026, The Wall Street Journal reported Amazon–Decart acquisition talks at roughly $7 billion. The business question reaches beyond the potential price: where can AI practise before its mistakes affect customers, equipment or money? Reuters coverage.

What happened

Reuters relayed the Journal’s account. Its initial dispatch said it could not immediately verify the information, while subsequent coverage kept completion conditional. The figure is a reported potential deal value, and the talks could still fail. Initial Reuters dispatch, subsequent Reuters coverage.

Decart’s products give the simulation thesis substance. The company says Lucy transforms live video, while Oasis 3 generates interactive environments for training and testing autonomous systems, initially vehicles. These are company descriptions; buyers still need to test suitability for their own requirements. Lucy, Oasis 3.

There is also an infrastructure story. Decart markets DOS as an optimisation layer for training and inference across chip architectures. Barron’s highlights software that helps developers switch between AI chips. The inspected reports therefore provide several possible strategic lenses; they do not establish simulation as Amazon’s acquisition rationale. Decart’s optimisation platform, Barron’s report.

Why it matters

Rehearsal becomes valuable when it reveals a failure that would be costly to discover during live operations. Applied beyond physical AI, this thesis suggests a practical business experiment: let a customer service agent encounter an ambiguous refund request, missing delivery evidence or conflicting instructions in a controlled environment before it handles a real account. That is an editorial application of the idea, rather than an announced Decart deployment.

Start with one workflow. Define the actions the agent may attempt, the evidence it should consult and the conditions requiring escalation. Use representative cases, including awkward exceptions. Record the proposed action and its reasoning, then have the process owner assess whether it would satisfy the actual business rules.

Measure incorrect actions, missed escalations, review time and the cost of maintaining the test environment. A programme that produces impressive demonstrations but consumes more supervisory effort than it saves has not yet earned expansion. The relevant return is better decisions and fewer consequential mistakes, rather than the volume of generated scenarios.

The bigger shift

The strategic asset could be a repeatable learning loop: representative scenarios, recorded actions, clear scoring and controlled changes. Access to a capable model does not automatically provide that loop. It requires operating knowledge, maintained test cases and someone accountable for interpreting failures. This is a concrete reason to let AI strategy guide tool selection.

Simulation quality also deserves scrutiny. Decart itself describes Oasis 3 as a world generation model rather than a physics engine. Visual realism alone cannot establish that a test captures every constraint relevant to a real operation. Oasis 3’s product FAQ.

For business workflows, the equivalent danger is a beautifully coherent scenario built on incorrect assumptions. If the rehearsal omits a contractual restriction, an unusual customer behaviour or a system failure, repeated success could create misplaced confidence. Limited live trials should test whether lessons transfer beyond the simulated setting.

My take

The idea of manufacturing experience is useful when experience means evidence from controlled trials. The advantage would come from discovering consequential errors early and improving the process around them. Repeating an unrealistic scenario thousands of times could simply reinforce a mistaken assumption.

Leaders should fund a small rehearsal programme with a workflow owner, a defined testing budget and explicit criteria for moving into a limited live trial. Human-in-the-loop review should determine what the evidence permits, with clear authority to stop the system.

The leadership question is practical: which costly decision could we rehearse this month, and what result would make us withhold permission for AI to act?