The Refrigerator Principle

I have a go-to example for explaining AI context windows to people who have never thought about AI context windows, which is approximately one hundred percent of my workshop participants. The example involves a fridge. It is, I maintain, an excellent example. It is also, I have come to understand, a kind of beautiful lie.

You open the fridge. You see leftover rice, half a tin of coconut milk, some wilting spinach that has perhaps three days left on its optimism, two eggs, and a jar of something you bought for a recipe you made once in 2022 and have been moving to the back of every fridge since. You say to an AI: what can I make for dinner?

And here is the beautiful thing: the AI can actually help. Given those specific, concrete ingredients, it can tell you about the rice bowl, the scrambled eggs with spinach, the coconut curry that will finally vindicate the jar. The fridge example works because it’s personal, it’s low-stakes, and it contains everything an AI needs: a clear inventory of available resources and a request. No ambiguity. No authentication requirements. No SharePoint.

The fridge is not connected to SharePoint. This is the problem.

What the Fridge Is Teaching (And What It Isn’t)

The pedagogy behind the fridge example is sound. What I’m trying to instil — across the span of a two-hour workshop in a room where at least three people are quietly wondering if they can claim this as TOIL — is a mindset. Specifically: AI is a capable assistant, possibly a remarkable one, but it requires context the way a recipe requires ingredients. You cannot ask it to help you without telling it what you’re working with. The fridge makes this obvious in a way that “context window” never will, and in a way that doesn’t require anyone to have previously heard the phrase “transformer architecture” or expressed any desire to do so.

What the fridge example does not teach — and what I glide over with the cheerful velocity of someone changing the subject at a dinner party — is that the workplace equivalent of “what’s in your fridge” is a question nobody has ever been asked before. And the answer, if anyone bothered to look, is considerably more complicated than leftover rice and two eggs.

The workplace fridge contains: a Microsoft Teams channel from 2021 where a critical decision about the training budget was made in a message thread that also contains forty-seven responses about the Christmas party, a booking form for the wrong system, and a GIF. A SharePoint site last reorganised by a person who has since left the institution, whose filing logic was self-evident only to them and possibly not even to them. A legacy database that predates the current strategic plan, the previous strategic plan, and possibly the institution’s current logo. An email chain forwarded seventeen times with an attachment at the bottom titled Final_v3_REVISED_FINAL_USE THIS ONE.docx, which is, in fact, not the final version but is the only one anyone has a link to. And somewhere, in someone’s head, the actual answer — the institutional memory that has never been written down because there was never time, and now the person who held it is on long service leave until March.

None of this is accessible to the AI. I have not mentioned this yet. I am building to it with the careful pacing of someone who knows they are about to disappoint a room.

Their kitchen, I am preparing to explain, has a locked pantry, a fridge they can’t open without a ticket to IT, a head chef who hasn’t decided whether cooking is permitted yet, and a stove that was connected to the network by the previous occupant and now requires a password nobody has written down. They have been asked to make dinner. I am very enthusiastic about what they could make.

We will be ordering pizza.

The Information Landscape Nobody Can Describe

The first sign of trouble comes when I ask participants to do the workplace version of the fridge inventory. Just: what information do you regularly work with? Where does it live? What would the AI need to know about your context to actually help you?

Silence. The specific, contemplative silence of people discovering, in real time, that they have never thought about this. Not because they’re being difficult — I want to be clear about this, because the Territory has a particular flavour of weaponised indifference that this is not — but because nobody has ever asked them to audit their information environment. The information environment just is. It has always been. It predates them, in many cases. It will outlast them. It is simply the water they swim in, except the water is SharePoint and some of it has been there since before the current IT provider.

I ask one participant — a thoroughly typical government administrator, because she is — what she’d need the AI to know to help her write a weekly status report. She thinks for a while.

“It would need to know what happened this week,” she says, carefully.

Yes. And where does that information live?

“…In my head.”

And before that?

“In the meeting notes.”

Where are the meeting notes?

A pause of considerable duration. “In a Teams chat. I think. Or it might have been the other Teams. We have two.”

We have arrived at the SharePoint archaeology problem, which is less an archaeological dig and more an unexploded-ordnance situation. Every institution has, embedded in its digital infrastructure, layers of information deposited by people with different intentions, different filing conventions, and radically different tolerances for the concept of a naming system. Somewhere in the sediment is the thing you need. It is adjacent to seventeen things you don’t need, one of which is a complete mystery, and two of which should probably be deleted but nobody wants to be the person who deletes them because what if.

The Teams graveyard deserves its own consideration. Every organisation has Teams channels that were created with great enthusiasm — the Project Steering Committee Q3 channel, the New Staff Onboarding Resources channel, the We Should Probably Do Something About The Website channel — and that now exist primarily as digital monuments to optimism. Important things happened in these channels. Decisions were made. Files were shared with the confidence of people who believed this system would last. And then the channels fell quiet, and the important things settled gently to the bottom like sediment, and above them accumulated three hundred messages about the Christmas party, a poll about whether to move the all-staff from Thursday to Friday, and a message from someone in November 2023 that just says “sorry wrong channel.”

The AI cannot access any of this. It is a text box. A remarkably sophisticated text box, capable of feats that would have seemed genuinely miraculous five years ago, but a text box nonetheless. I have not yet said this out loud. I am saying it now, to you, in the manner of a nature documentary narrator who has been watching a situation develop from a safe distance and feels the audience deserves to know what the wildebeest cannot.

The fridge is full. The fridge has always been full. The question is whether any of what’s in it is still usable, and whether you can actually get the door open, and whether the light comes on when you do.

“Open Claw with Your Org’s Single Sign-On” (The Dream)

The dream usually surfaces around the forty-minute mark, just as participants have grasped that the AI can help with things they bring to it manually. Someone — there is always someone, and I love them unreservedly for it — says the thing.

“Can it just… do things for me?”

This is, technically speaking, a question about AI agents. The participant does not know this. They would not call it that. They are imagining — with impeccable precision and no vocabulary whatsoever — something that connects to their email, reads their calendar, checks their outstanding tasks, pulls from their project management system, accesses their document library, cross-references the relevant policy documents, and then drafts the thing they need drafted while they attend the other meeting that has been scheduled at the same time by someone who does not know about the first meeting and would not have cared if they did.

They are imagining an AI with their organisation’s single sign-on, their permissions, their context, and a general mandate to sort things out. They are imagining, in other words, something that does not yet exist for them — and I am the only person in the room who knows how wide the gap is between this dream and the current reality. My job, at this moment, is to walk them gently back to the chat box without removing from their faces the expression of someone who has just glimpsed, briefly, a better version of work.

The answer to “can it just do things for me?” is: yes. Eventually. Conceptually. In some enterprise deployments, with appropriate governance frameworks, after your IT department has completed the security assessment, after the vendor has finished the compliance review, after the data sovereignty questions have been resolved, after the policy position has been drafted and approved and revised and re-approved and communicated to staff via a SharePoint page that two people will read, at which point the technology will have moved on considerably and we will be having a version of this conversation again, probably in a room with better air conditioning.

For now: text box. Very good text box. Genuinely impressive text box that has, in the past six months, done things I would not have believed possible. Still a text box.

I feel, at this moment, like someone who has been asked to teach a person to drive and has handed them a steering wheel that is not attached to a car. It is an excellent steering wheel. It responds beautifully. It has a very intuitive interface. We are, in this metaphor, doing our best with the equipment available, which is the unofficial motto of every regional training team in Australia and has been since roughly 1988.

The Widening Corridor of What We Cannot Do

There is a list of things AI tools can help participants with, and a list of things they cannot do — cannot do specifically, as in, there is a policy document that says so, or there is an IT restriction that enforces it, or there is a legal grey area that nobody has resolved yet and therefore the default position is “not yet,” which in institutional language means “not this year” and sometimes means “not this decade.”

Both lists are growing. The useful list is growing slowly, in the way that coral grows — visible only if you measure it against something fixed and are patient about what constitutes progress. The restricted list is growing faster, and for entirely understandable reasons. Every new use case someone thinks of surfaces a category of risk nobody had modelled until someone tried it. Someone pastes client information into the public-facing chat interface. Someone uses the AI to draft a communication that should have been reviewed first. Someone discovers, to the general alarm of a Friday afternoon, that the AI has reproduced something it shouldn’t have reproduced, and now there is an incident, and incidents generate investigations, and investigations generate policies, and policies generate restrictions, and restrictions become the list.

The result is what I have started calling, in the privacy of my own skull, the widening corridor. You can use AI for this, and this, and this thing over here in the third column — but not for that, or that, or the cluster of applications your most enthusiastic participant just described with the barely-contained joy of someone who has finally understood what a technology is actually for. I have watched, more than once, the specific sequence of expressions on a participant’s face as they describe the thing they were most looking forward to doing, and I am required, with the measured delivery of someone breaking mildly bad news at a reasonable volume, to tell them that this particular application sits in a policy grey area we are currently navigating. The expression goes: hope, then calculation, then the particular variety of professional disappointment that is too well-mannered to be called devastation but requires a moment to recover from. I have become an expert in the micro-pause that follows. I do not enjoy it.

I drive home along the Laterite Highway at seven in the evening, the sky doing that thing it does up here where it turns approximately seven colours for no reason anyone has satisfactorily explained, and I ask myself the question I have been avoiding for most of the drive: am I running AI workshops for people who are effectively forbidden to use it?

Not quite. The answer is not quite. But the gap between what is permitted and what is genuinely useful is not narrowing at a rate that inspires confidence, and the people who could most benefit — more on this in a moment — are precisely the people for whom the corridor is narrowest.

The Narrowing Window

The people who carry the highest cognitive load in any organisation are rarely the people making decisions about governance frameworks. They are the case managers processing forty files a week with tools that were old when they arrived and have not been updated since the last restructure. The administrative staff who hold institutional memory in their heads because the knowledge management system was never properly populated and there was never going to be time to fix it. The education support workers navigating three databases that do not speak to each other, a referral process that still, in 2026, requires a printed form, and the knowledge that if they get sick, the whole thing stops because there is nobody else who knows the workarounds.

These are the people who would benefit most from a capable AI assistant that could read their context, understand their load, and absorb the forty-five minutes of repetitive information-processing that currently eats the end of every day that was already too long. And these are also, in many cases, the people for whom the governance situation is most complex. Government. Health. Education. The sectors with the most sensitive data, the most comprehensive restrictions, and the most legitimate reasons for both.

The gap between what AI could do for them and what they are currently permitted to do with it is not a technology problem. The technology is, in a meaningful sense, ready. The gap is governance. Policy. Trust. The legacy IT architecture that was never designed to interface with a third-party AI service and would require significant work — work that has been assessed, scoped, estimated, and placed in a queue — before it could be made safe enough to try. These gaps will close. They are closing, somewhere, in some organisations, for some roles. But they close slowly, unevenly, and usually only after an incident that nobody wanted, which generates the next layer of caution before eventually, after sufficient time and iteration, producing something workable.

In the meantime, I am teaching people to fish in a river they are not yet permitted to stand next to. I am giving them a steering wheel and genuine enthusiasm about what they will be able to do with it. The car is coming. I believe the car is coming. But we are, right now, in the section of this particular journey that is primarily steering wheel, and it is my professional responsibility to make the steering wheel as useful as possible while being honest about what it is not attached to.

The Refrigerator, Which Is Still Something

Back to the fridge.

The fridge is not nothing. A text box that can help you draft a difficult email you have been avoiding for three days, think through a problem that has been occupying the back of your mind since Thursday, summarise a forty-page document you’ve copied and pasted, generate three options you hadn’t considered, give your quarterly report a structure when you know exactly what you want to say and have no idea how to begin saying it — that is real. That is genuinely, measurably useful. That is worth a two-hour workshop and the drive to the training room and the slightly too-cold air conditioning and the colleague who spends twenty minutes explaining why they already use ChatGPT and therefore don’t need to be here.

(They need to be here.)

It’s not the dream. The dream is an agent with single sign-on and calendar integration and the quiet competence to have drafted the quarterly report before you’ve finished your first coffee. The dream is the thing that would make the person carrying forty files feel like they’re carrying thirty-eight. The dream is close enough to see from here and far enough away that I remain professionally cautious about committing to an ETA.

For now, we are in the refrigerator. We are taking stock of what we have, learning to use it with some skill and a reasonable amount of confidence, and accepting that the pantry door is a governance question and not a technology question and will open when it opens. The locked pantry is real. The incomplete fridge is real. The head chef who hasn’t decided whether cooking is permitted yet is, I can confirm, extremely real.

But the fridge we have — the ingredients we can bring to it, the things we can make right now, today, without waiting for the policy to be finalised or the IT ticket to be resolved or the vendor to complete the compliance review — that is a start worth having.

It is, all things considered, a very good refrigerator.

Next week, I’ll show you what you can make with it.