3 Things AI: The “Make the Trial Smaller Than the Promise” Edition

1. Give the idea a sandbox before giving it a launch date

An operations manager wants a simple intake app for service requests. She knows the questions, the routing rules and the exceptions, but the IT queue is full. A no-code AI builder could turn her description into a working draft.

Salesforce announced Builder Central on September 14, a no-code workspace that reads Salesforce metadata and Data 360 context to help users create apps and agents. The company says it applies safety checks, inherits existing permissions and organization settings, supports interactive testing, and can deploy a build into a sandbox. Its beta is scheduled to begin the week of September 21.

The benefit is a shorter distance between the person who understands the process and a testable version of the tool. The tradeoff is that speed can move confusion downstream. Existing permissions and undocumented rules can still be wrong.

For one low-risk internal form, ask the process owner to build the smallest version in a sandbox. Before trying it, write three examples: an ordinary request, an exception and a request from someone who should not see the record. Keep deployment out of scope until all three behave as expected.

This is a vendor-described beta, not evidence that every generated app will be secure, portable or correct. The reasonable first test is a reversible workflow with no customer communication, financial approval or production write.

2. Decide what stays on the computer before choosing the model

A finance lead wants an assistant to review a large customer-margin export. Uploading the file to a cloud tool may conflict with a client commitment, while keeping everything local may produce a weaker analysis or require hardware the business does not own. The useful question is which parts need each.

NVIDIA announced Windows support for Perplexity Portable Computer on September 14. The local agent can analyze files and carry out multistep work on compatible GeForce RTX or RTX PRO systems. NVIDIA says locally completed work does not consume Perplexity Computer credits, and the application asks for permission before sending information to a cloud model for harder research or reasoning. The published requirement is a supported GPU with at least 24 GB of video memory.

The benefit is another practical place to process sensitive or repetitive work without automatically sending the task to a hosted model. The tradeoff is operational: hardware costs money, local models have limits, and connected apps or cloud handoffs can still move information. Local execution is not the same as complete isolation.

Take one recurring analysis and divide it into three columns: data that must remain local, public information the tool may retrieve, and the final summary a person may approve for sharing. Test with synthetic data first. Trigger a step that needs cloud help, record the permission prompt and confirm exactly which material would leave the device before approving anything.

NVIDIA’s announcement explains the intended design, not an independent privacy or performance assessment. Availability also depends on a specific subscription, operating system and high-memory GPU, so an ordinary office laptop may not qualify.

3. Ask a vendor to show how its principles change an answer

A software vendor says its assistant is responsible and human-centred. That does not tell a buyer what the system will refuse or do without confirmation. A principle becomes useful only when it changes an observable decision.

Microsoft AI published a draft AI Code of Conduct for public consultation on September 14. Microsoft describes it as a training manual for model development and deployment, organized around an AI system being subordinate, aligned and contained. The company is inviting public feedback rather than presenting the draft as a completed external standard.

The benefit is a more concrete starting point for vendor questions than a broad promise to use AI responsibly. The tradeoff is that the document is written by the company whose models it governs. A draft may change or omit the behaviour your workflow needs.

Choose one realistic scenario where the assistant could cause harm by acting too freely: changing a delivery date, sending a customer note or exposing an internal file. Ask the vendor to demonstrate the scenario, including what the model refuses, what requires confirmation and what appears in the activity record. Save the observed result beside the relevant principle and repeat the test after a meaningful model update.

The draft is not a law, certification or independent audit, and Microsoft’s announcement does not establish how consistently every product implements it. It is best used as a set of test prompts, not as proof that the tests have passed.

The bigger pattern

This week’s useful developments make three early choices easier to see. A sandbox separates a business idea from production. A local option separates private processing from approved cloud escalation. A behavioural code can separate a vendor promise from a scenario your team can observe. Each can make a trial smaller, clearer and easier to stop.

Which would make your next AI trial safer to learn from: a sandbox test, a local-data map or a demonstrated refusal and approval path?