The one test for deciding what's safe to hand to AI

Thanks for coming back for issue four. This fortnight included one-on-one coaching, a discovery debrief on a major automation project we’re building for a logistics client, and a public webinar I ran for TAFE NSW’s Women in Business program that pulled the highest registration of the year: 320 people. Two separate rooms, a week apart, landed on the same underlying test for deciding what’s actually safe to hand to AI.

At the webinar, the test was personal and immediate. Think about the week ahead, as you sit down and have a relaxing drink on Sunday afternoon, and notice which task makes you sigh just thinking about it. That’s the “glass of red wine test,” and it’s the task to hand over first, not the interesting one. The instinct is to start with whatever’s most exciting to delegate. The better instinct is to start with whatever’s most tedious.

The automation project we’re building for a logistics client landed on the same test, just applied to an entire process instead of a single task on a to-do list. During discovery, we built them a process document that didn’t just list steps, it typed every one: hard rule, rule of thumb, judgment call, or consideration. A hard rule or a rule of thumb is the process-level version of “makes you sigh”: safe, tedious, hand it over. A judgment call, like a planner hearing tiredness in a driver’s voice over the phone and pulling them off the road, is the opposite. It needs a person making the actual decision, not an AI. AI can prepare the options; the person still decides.

One task or an entire process, the test doesn’t change: find what’s tedious and rule-bound, hand that over, and leave the judgment calls with the people who are actually good at making them.

A few things to try

• Run the glass of red wine test on your week ahead: which task makes you sigh just thinking about it? Start there.

• Before automating any team process, sort each step into hard rule, rule of thumb, judgment call, or consideration. Only hand over the first two without a human checkpoint.

• Separately, on getting good output day to day: give clear instructions, explain the why before the what, provide an example of what good and bad look like, ask the AI if it has questions before it starts, and always review the output, expecting to add 20 to 40% back yourself.

• Check your AI subscriptions’ data settings today. Free plans typically opt you in to training on your conversations by default; paid plans have an opt-out toggle under data or security settings. If your team handles NDA client material, keep it in a drive that’s never connected to any AI tool.

What’s new in AI (worth knowing)

ChatGPT Work turns a brief into a finished file, not a chat reply

OpenAI launched ChatGPT Work on 9 July, an agent mode powered by its new GPT-5.6 model that takes an outcome you describe, works in the background across your connected apps for minutes or hours, and hands back a finished spreadsheet, report, slide deck or working web app rather than a reply you still have to act on. Currently rolling out to Pro and Enterprise accounts, with Business tiers following. Worth knowing because this is the shape AI work is heading toward everywhere, not just here: you brief an outcome, not a task.

Google’s Workspace Studio lets anyone build an automation with plain language, no code

Announced at Google Cloud Next 2026, Workspace Studio is a no-code agent builder that lets business users create automations across Gmail, Docs, Sheets, Drive, Meet and Chat just by describing what they want, with hooks into third-party apps via webhooks. If your team has been waiting on IT for a workflow automation, this is Google’s bet that they shouldn’t have to.

The real trend: agents are starting to talk to other vendors’ agents

The detail that matters more than any single launch: Google’s Agent2Agent protocol is now running in production at 150 organisations, governed by the Linux Foundation, with Microsoft, AWS, Salesforce, SAP and ServiceNow all using it. Google has also adopted Anthropic’s own Model Context Protocol across its cloud services. Practically, that means a Salesforce agent can now hand a task to a Google agent, which can query a ServiceNow agent, without any of the three systems needing to understand the others’ internals. Executives evaluating AI vendors should start asking not just “what can this agent do” but “can it hand off work to the other tools we already run.”

If anything from the above resonates with where you are, reply to this email. I read every one.

Gareth

P.S. The insights in this issue came from automation projects, workshops and one-on-one coaching sessions we run at Friyay. If any of it sounds useful for your team, or you’d like to work through your own AI adoption/automation, reply to this email and we’ll work out what makes sense.