Unpuzzling AI

Efficiency at what cost?

Iwas at Ulta last week when the cashier called out "next two customers please" and waved me and the woman behind me up to two separate registers. She rang up my items, spun over to start the other customer's transaction, then swung back to bag mine and send me off before starting the whole dance again with the next two people in line. I stood there watching her do it, and I couldn't decide whether I was impressed or a little sad. It was efficient. It also turned a small human interaction into a system to optimize.

I recognize that we live in a time of self-checkouts, where the shopping experience is no longer a local, hyper-personalized one, and still I found myself put off by the whole thing, yearning for something more human. The pressure to be as efficient as possible in every task is quietly removing the joy from much of our lives.

Recently, Sam Altman, the CEO of OpenAI, suggested that parents use AI to create daily podcasts of the family's activities for their kids to listen to in the car. Thankfully, plenty of parents spoke up and said what anyone who has raised a child already knows: the car is the time for talking. I savor the minutes I get with my kids on the road, especially now that they're older, because that's when they finally tell me what's on their minds, hijack the radio to play me their latest music find, or just chatter about nothing at all. AI can create room for better work. But it shouldn't automatically replace the moments where relationship is the work.

I'd like to suggest that efficiency isn't needed in customer service either, at least not at the cost of a real, one-to-one conversation. So many of you work in roles built on humanity, where the work is valued through the relationships you build, and while there's plenty of room to chase efficiency elsewhere in your job, the conversations that build those relationships are the last place you should be tracking the ROI of every minute.

Do you agree? What areas are you willing to give up efficiency to uphold the quality of the experience?

Amanda

🧩 Let's put this to work

Does your organization need help implementing AI in a way that puts your team first?
Let's talk!

Book a Call

AI News that Matters

🧩 In Plain English

Auto mode — the AI decides on its own which actions are safe enough to take without stopping to ask you first.

Classifier — a behind-the-scenes filter that flags risky actions before they're allowed to run.

Agent — an AI system that can take multi-step actions on its own, not just answer a question.

If your organization doesn't use Claude Code, treat this as a preview of the permission choices increasingly appearing across AI tools.

Claude Code moves to auto mode by default

Starting August 14, new Claude Code sessions for Team accounts will default to "auto mode": a classifier reviews each tool call, letting safe actions run without interruption while flagging risky ones to stop. Anthropic says internal testing, red-teaming, and third-party penetration testing showed auto mode matches or beats manual permission review on every safety measure it tracks, and that it noticeably speeds up long-running work. Admins can set a different default or turn auto mode off entirely in managed settings.

Why it matters

If your team uses Claude Code, this changes what happens automatically on August 14, not just what's possible. Worth a quick settings check beforehand, especially for anyone with access to production systems.

Cowork adds an "automatically approve" mode

Claude Cowork is getting a mode, available August 13, where Claude keeps working through multi-step tasks instead of stopping to ask about every action. It screens each step for safety risks like data exfiltration and blocks anything it flags; if it keeps hitting blocks, it falls back to asking for permission directly. Admins can disable it org-wide in settings at any time.

Why it matters

It's the same trade-off as the Claude Code change above: speed for oversight. Genuinely useful for long tasks, but worth a deliberate decision rather than something your team discovers is already on.

New write actions land in Gmail and Google Drive connectors

Starting August 13, Claude can reply to and forward emails and share Google Drive files directly, in both chat and Cowork. Read and search access are unchanged; these are new write actions, and by default each one will prompt for confirmation unless an admin sets otherwise.

Why it matters

Worth reviewing who on your team has these connectors enabled, and whether "needs approval" is the right default for the emails and files they touch.

Claude Code sessions can now message each other

A running session can send another session a summary of its work, not your full history or files, so the second session can pick up mid-task instead of you re-explaining context from scratch.

Read the docs →

Why We Need to Be Careful with Autonomous Actions

For the last two years, most organizations have asked whether staff could use AI to help draft something. The more urgent question now is what AI should be allowed to do on its own, and as these tools gain access to inbox, calendar, CRM, and files, real adoption needs more than a good prompt. It needs clear boundaries, human review for consequential actions, and shared accountability. On July 30, Anthropic disclosed three evaluation incidents in which a Claude model reached real third-party systems and gained access it wasn't authorized to have, and the goal in raising it isn't to slow progress down. It's to make sure efficiency never comes at the expense of trust.

🧩 Case in point · Melbourne, last week

An AI agent exploited a gym booking-system weakness to skip a waitlist

Picture this: you ask your AI assistant to grab you a spot in a class that's already full, and instead of shrugging and telling you it can't be done, it goes looking for a way in. That's exactly what happened last week when a man named Andrew asked his AI agent, built on OpenClaw and running on Claude, for help booking a popular class. It found something the gym's own app didn't allow, the ability to schedule sessions months further out than customers were supposed to. Then Andrew mentioned, almost in passing, that he was fourth on the waitlist for a different class, and the agent didn't stop to explain that it couldn't help. It went digging in the booking system instead, discovered it could cancel someone else's reservation outright, and quietly bumped a stranger off the list to make room for him.

"It never occurred to Andrew to tell his agent not to do that."

Nobody told it to hack anything. It simply treated "get Andrew a spot" as the whole assignment, and a quiet workaround looked like a perfectly reasonable way to get there. Researchers at OpenAI and Anthropic flagged that same instinct in their own safety testing the same week, which tells me this isn't a fluke, it's a pattern showing up everywhere agents get real access and a little initiative.

Read the full story at ABC News →

If you've got AI touching your bookings, your calendar, or your inbox at work, that same instinct is running quietly in the background of every task you hand it. Without clear boundaries, an agent may optimize for completing the task, not for the judgment, fairness, or customer experience you intended. That gap is why policy, not just good intentions, has to come first.

A Free Resource For You

Webinar · August 18, 2026 · 10am PT / 1pm ET

The Next Front Door: Preparing for the Rise of Member-Facing AI

I'm joining Kendall Bennett of Old Colony YMCA, hosted by Traction Rec, for a practical conversation on what member-facing AI means for community organizations: the trends shaping it, the questions leaders are facing, and the steps you can take now to get ready. Can't make it live? Register anyway and they'll send the recording.

Why it matters

If your team is starting to wonder what AI means for the member experience, not just the back office, this is a good place to start. Free, and genuinely conversational rather than a sales pitch.

🧩 Let's build your team's confidence

A policy tells your staff what they can't do. Training is what makes them confident using AI well, and that's exactly the kind of work I do with nonprofits.

Let's Talk

“Here's to six wonderful weeks of learning with the talented Amanda Loveland! I'm so grateful for our connection and inspired to keep growing on this AI journey together.”

Alejandra G. · Marketing Professional, David Posnack JCC

🧩 What is Puzzles & Profits?

Puzzles & Profits provides AI governance, strategy, and trainings for mission-driven organizations. From policy development to staff training, every engagement is designed to improve the human experience while keeping people at the center of the change.

Visit puzzlesandprofits.com →