It still blows my mind how much workflow is just email.
We have modern systems everywhere, yet so many teams still run their day through inbox threads, forwarded PDFs, and someone chasing someone else for an update. Even when there is a "proper tool", a lot of the real work sits outside it, in human back and forth. This is one of the reasons I care so much about AI for schools being practical. If AI doesn't respect the real flow of work, it won't land. If it does, it can quietly remove the friction that wastes everyone's time.
This week, a few different things I read and watched all circled the same point from different angles: the future isn't "AI replaces people". It's "AI supports the messy human loop".
Email is still the operating system for real work
If you want to understand why change is hard, look at where work actually happens. It isn't always in the official platform. It's in the emails, the side conversations, the forwarded attachments, the "just checking in" messages, and the mental load of tracking what's missing.
For schools, this is the same story. Teaching and learning might happen in the classroom, but the operational glue is often a mix of email, documents, and follow-ups. That's why I'm increasingly sceptical of AI that demands a perfect workflow. The winning tools will be the ones that step into the existing mess and make it easier, without pretending the mess doesn't exist.
Flexport is a logistics story but it felt like an education story
I watched a YouTube with Ryan Peterson from Flexport. It's about logistics, but it felt very familiar.
One part that jumped out: logistics contracts often arrive as giant Excel files. Thousands of rows. Heaps of tabs. Humans then translate that into something usable. Often by email. Often by hand. Sound familiar?
What I liked was how human-centred their AI approach is. It takes the boring glue work out of the loop. It also reminded me of kintsugi. You don't pretend the break wasn't there. You reinforce it and make it part of the design.
A few things they talked about that map cleanly to how many organisations work (schools included):
- Using AI to parse messy contracts and unstructured messages into structured data
- Optimising choices at scale, many more times a day than humans can
- An agent that confirms addresses and appointments via email and voice, especially when something is unfamiliar
- Natural language reporting, so customers ask questions without needing dashboards or SQL
- Detecting unhappy sentiment and escalating early
- Training non-engineers one day a week for 90 days so domain experts can build automations themselves
This is the bit I keep coming back to. AI works best when it respects the real flow of work. The messy human flow. The email flow. That's not a compromise. That's the design target.
https://www.youtube.com/watch?v=KTmxaMdUbHA
The AI gap between students and teachers isn't universal
I read a piece from NBC News describing a growing gap in the US, with students moving fast on AI and teachers playing catch up.
Maybe this is shaped by my work and the interface I have into schools, but it isn't what I'm seeing in Australia. In many schools I work with, teachers are running at the same speed as their students. Not perfectly, not uniformly, but the idea that teachers are always "behind" doesn't match what I'm observing on the ground.
My bigger concern isn't speed. It's consistency. Some schools have leaders creating shared language and safe norms. Others have everyone improvising alone. That's where the risk creeps in.
https://www.nbcnews.com/tech/tech-news/ai-school-teacher-student-train-chatgpt-rcna248726
If we want safe AI for schools we need to understand how the model got there
I spent time this weekend reading OpenAI's paper on chain of thought monitorability, and it landed hard.
When implementing AI, we can't just judge it by the final answer. We need some way of understanding how it got there. The paper makes the case that looking at a model's reasoning is more useful than watching outputs alone.
The surprising point for me was that more thinking can help. Longer chains of thought can make models easier to supervise, not harder. In some cases you can use a smaller model, let it think more, and end up with something that is easier to monitor.
I've often thought of monitorability as impractical at scale, but this nudged me to think it might become one of the most practical safety tools we have, especially in high-stakes settings like education.
The chart in the paper highlights the difference in observability when monitoring:
- just the output
- chain of thought messages only
- both sets of messages
Monitoring all outputs produces the best results.
https://openai.com/index/evaluating-chain-of-thought-monitorability/
The work that matters is slow, careful, and human
This year I've spent a lot of time arranging IP and licensing deals so CurricuLLM can lawfully use curriculum content. It's reminded me how slow, careful, and human this work actually is. It involves conversations, trust building, and respect for the people who created the material teachers rely on every day.
This is why policy cycles feel like they're under pressure right now. The Productivity Commission's suggestion of a three-year review before any broad AI copyright exemption makes sense on paper, but three years is a very long time in the AI world. The pace of change is moving faster than most policy cycles, and that creates real tension for builders trying to do the right thing while still moving forward.
If you're working at the intersection of AI, licensing, or public policy, it's worth reading.
The thread tying all this together
Here's what I'm taking from the week.
AI for schools will succeed or fail based on whether it respects reality. Reality is email. Reality is partial information. Reality is trust and relationships. Reality is slow licensing work. Reality is teachers doing incredible work inside messy constraints. Reality is that safety can't be judged by outputs alone.
The exciting part is that we can already see the pattern for what works:
- Don't force a perfect workflow. Support the workflow people already use.
- Don't pretend humans are out of the loop. Design for human judgement.
- Don't treat safety like an afterthought. Build monitorability and transparency in from the start.
- Don't shortcut trust. Especially in education, trust is the product.
If the next wave of AI tools can do that, we won't just get faster work. We'll get better work. And in schools, that's the only outcome that matters.