It obviously didn’t take very long. It was ugly at first. Then I made it nice looking. And it works.

A friend saw it, asked for a copy, and now a couple of other people use some version of it. None of us would have paid for this. No product manager would have greenlit it. It exists because I wanted it to exist. I wasn’t the only one but it wouldn’t have mattered if I was.

I’ve also made a CRM that works the way I work, a serendipity engine, a hype decay tracker, a draft graveyard and heaps more things.

I’ve started calling this folk software.

The songs of folk music emerged from communities rather than studios, passed around and adapted, never focus-grouped. Folk music didn’t disappear when recorded music arrived. It just stopped being the only option. For a while, if you wanted music, you either made it yourself or knew someone who could.

Software has been in its “recorded music” era for decades. If you wanted a tool, you either bought what the industry offered or you didn’t have it. The threshold for creation was high enough that almost everything had to be commercially justified. Will enough people pay? Can we get budget for this?

That threshold just collapsed.

Tools like Claude Code, Cursor, Replit mean the calculus is now simply: do I want this? The answer can be yes for an audience of one. And sometimes that audience will turn out to be more than just you.

There’s an idea in urban planning called desire paths : those unofficial trails worn into grass by people walking where they actually want to go, ignoring the paved routes (you’ve probably seen the JPG meme).

Eventually, smart planners pave over the desire paths instead of fighting them.

The best folk software works the same way. It’s not imagined from first principles or designed for an abstract user. It’s discovered through living. You needed something, you made it, you realised others walk the same path.

This is the opposite of product-market fit. It’s product-life fit that turns out to be shared.

Folk software has a particular character:

It’s small. Often a single file, a script, a workflow. It doesn’t want to be a platform.

It’s specific. It solves this problem, not problems in general. The constraints are what makes it useful.

It spreads through sharing, not marketing. Someone posts it, someone forks it, someone adapts it for their own situation. The lineage matters more than the brand.

And crucially: the creator is a user. Not a product manager, not a founder chasing a market. Someone who needed the thing and made the thing. The gap between those two roles has closed to zero.

Woodworkers have a concept called the jig drawer. A jig is a custom tool you make to help with a specific repeated task, holding a piece at the right angle, guiding a cut, spacing things evenly. Every woodworker’s jig drawer looks different because everyone’s work is different.

We’re all building jig drawers now. What will happen when millions of people have jig drawers full of small, specific, useful tools they made themselves?

I think we’re about to find out.

If you haven’t made your first folk tool yet, start with something small. Something you actually need. Don’t worry about whether anyone else wants it. That part tends to sort itself out. Why are you still here? Go make.

Dave.

Fast forward to today, and we're helping ASX-listed companies create highly impactful LLM-powered systems that handle vast amounts of valuable interactions.

The technology has transformed the way ambitious organisations are now creating new experiences for customers, driving efficiency and growth internally and augmenting the value that their people and data offer.

In working across a diverse bunch of AI projects over the last few years, we’ve developed some principles that, although not locked in stone, are robustly discussed every day in our work.

We've found that applying product development principles to internal AI initiatives—which forms a significant portion of our work—is just as valuable as using them for customer-facing AI applications.

So whether you're leading AI strategy from the C-suite or building it in the trenches, we think these eight principles will help you ship AI products that users actually adopt and businesses actually benefit from:

1. User Jobs-to-be-Done, Not AI Jobs-to-be-Done

Whereas traditional thinking always had us start with user needs and building features to meet them, in working with AI we still start with user needs but need to figure out where AI actually adds value.

The biggest mistake we see people make? Falling in love with what the AI can (or might be able to) do instead of what users need it to do.

Your users don't care that your LLM can write poetry, count the Rs in strawberry and explain quantum physics. They care about getting their specific job done faster, better, or cheaper or even more creatively. Before you write a single prompt, map out the core user journey and identify the specific friction points where AI can genuinely help.

Additionally, when it comes to AI workflow design, we constantly need to be questioning why things are currently done in a certain way. In 2025, there stacks of ways to reimagine customer experiences and organisational workflows with a view to what’s now possible.. that simply wasn’t yesterday.

2. AI Product Owners: business and technology

The most successful AI implementations we've seen at Move 37 have one thing in common: dedicated business and technology product owners who can make decisions in real-time.

Not just stakeholders or reviewers but actual decision-makers with some form of skin in the game. Just to be clear, these people come in many different forms - CEO of medium-sized businesses, CTOs, CMOs, Product Innovation Leads etc.

Here's why this matters more than ever with AI products:

a. AI decisions happen at the intersection of business logic and technical possibility.

When your LLM suggests three different approaches to handling a customer complaint, you need someone who understands both the business implications ("Will this response align with our brand voice and customer retention goals?") and the technical constraints ("Can we reliably detect when to use this response pattern, and what's our fallback if the confidence score is low?").

b. The judgment calls are constant and nuanced:

• Should the AI be more helpful or more cautious in this scenario?

• When should we escalate to humans vs. push for automation?

• How do we balance personalisation with privacy?

• What level of AI confidence threshold makes business sense?

• How do we handle edge cases that represent 1% of interactions but 50% of user frustration?

AI product development is a team sport that requires both business alignment and technical discernment, at the same time.

3. Deterministic Outcomes from Non-Deterministic Systems

One of the biggest challenges in working with business and technology teams is developing a shared understanding of the characteristic of the AI in play in any given system and job-to-be-done. We literally never rely solely on an LLM’s training data to deliver value in a system. Our collaborative Move 37/client teams work hard at context engineering, and all that that entails (managing local data, memory, user personalisation and state and heaps more). In traditional software engineering, the aim was always to build predictable systems with consistent outputs. But with an LLM and the surrounding engineering, the opportunity is far greater.

So something we’ve learned when working with generative AI, is that we need to:

Embrace variability while engineering reliability.

LLMs are probabilistic by nature so they'll give you different outputs for the same input. This drives traditional product people crazy, but it's a feature, not a bug.

The key is building systems that deliver consistent user outcomes even when the underlying AI responses vary.

So consider your new reliability framework:

• Input standardisation: Control what goes into the model

• Output validation: Check what comes out before it reaches users

• Graceful degradation: Have fallbacks when AI fails

• Human-in-the-loop triggers: Know when to escalate to humans

Initially, think of it like managing a brilliant but inexperienced intern. You give them clear instructions, check their work, and have backup plans.

4. Progressive Prompt Architecture (Your New MVP Strategy)

Forget trying to build the AI assistant that does everything. Start with one specific task, nail the prompt and context engineering, then gradually expand. We call this "progressive prompt architecture"—kinda like progressive web apps, but for AI capabilities.

A potential expansion path can look like this:

1. Single-shot reliability: get one specific task working really well

2. Context expansion: Add relevant context, memory and other

3. Multi-turn interactions: Enable back-and-forth interactions where needed

4. Workflow integration: Connect multiple AI capabilities and external systems

5. Autonomous operation: Reduce human oversight where appropriate

5. Evaluation-Driven Development

There’s not a day (or even hour) that Pan, our co-founder/CTO doesn’t mention evals. In traditional product development we’d largely focus on user behaviour and business metrics to assess the successful implementation of an app or platform.

But with AI, we need to measure AI performance before it impacts these same user metrics.

Traditional A/B testing isn't enough for AI products. You need evaluation pipelines that catch AI failures before they become user failures. This means building evaluation frameworks that test:

• Correctness: Is the AI giving accurate information?

• Relevance: Is the response actually helpful for this user's context?

• Safety: Could this output cause harm or violate policies?

• Consistency: Would the AI handle similar inputs similarly?

• Latency: Is it fast enough for the user experience?

Pro tip: be careful about overfitting in evaluation. If you’re achieving 100% accuracy it’s likely that your product is deeply broken or you are tracking the wrong metrics.

6. Context is Your Product Moat

I get the feeling most people are getting across this these days but it’s still worth saying out loud - your unfair advantage is the context you provide to your models.

Every company has access to the same foundation models. Your competitive moat isn't the LLM itself, it's the proprietary context, data, and domain knowledge you feed into it.

There are a lot of different aspects to context but some common lenses include:

• User context: Who is this person and what's their history?

• Situational context: What's happening right now? What is job to be done?

• Domain context: What are the rules, constraints, and best practices?

• Organisational context: How does this fit into broader company processes?

The companies winning with AI are engineering better context. Invest in your context pipeline like you would invest in your core product features.

7. Human-AI Collaboration Patterns (Not Replacement)

There’s a reason we have the phrase “Human in the loop” embroidered on our Move 37 bomber jackets. It’s because since day one we understood that our role as product creators was to amplify human capabilities with AI superpowers via augmentation.

The most successful AI products we've built don't replace humans, they make humans superhuman. Instead of thinking "how can AI do this job," think "how can AI help humans do this job better."

Some proven collaboration patterns:

• AI drafts, human refines: AI creates first version, human polishes

• Human steers, AI executes: Human provides direction, AI handles execution

• AI suggests, human decides: AI provides options, human makes final call

• AI monitors, human intervenes: AI handles routine cases, escalates edge cases

Your users probably don't want to be replaced by AI. They want to be empowered by it. At a high level, design for augmentation, not automation (even some specific tasks become automated).

8. Cost-Performance Product Trade-offs

AI products introduce a new dimension to the traditional quality-speed-cost triangle. Every AI interaction has a marginal cost, and model performance often comes with exponential price increases. We always need to be keeping an eye on how to balance model capability, latency, and cost per interaction.

Your new optimisation framework:

• Model tiering: Use smaller models for simple tasks, larger models for complex ones

• Caching strategies: Don't re-compute what you've already computed

• Prompt optimisation: efficient prompt engineering = lower costs and faster responses

• Result recycling: Similar inputs should reuse previous outputs when appropriate

For example, we built a content authoring system that uses a small, fast model for obvious cases (clearly appropriate or clearly inappropriate content) and only escalates ambiguous cases to a more expensive, sophisticated model. This reduced costs by 80% while maintaining accuracy.

In Summary, the meta-principle: Build for Humans, Scale with AI

After eight years and a bunch of AI products for ourselves and client partners, here's what we've learned: the best AI products feel magical to users and boring to engineers.

Users shouldn't have to think about prompts, tokens, or model limitations. They should just get their job done better than before. Our product and engineering teams, meanwhile, should have robust systems, clear evaluation metrics, and predictable cost structures.

Now that hardly any organisations need to train their own models, the companies that will capture advantage are the ones that apply these kinds of well-considered product development principles to their AI initiatives.

What's your experience been? How are you thinking about building AI products? What principles would you add to this list?

Move 37 helps ambitious organisations build AI products and tools that drive impact. If you're wrestling with any of these challenges, reach out, we'd love to share what we've learned.

Snapshot : We made a song about building AI agents. Because of course we did.

We’re deep into building a bunch of different agents for a range of clients. Literally hours after OpenAI dropped GPT 4.1, our team is assessing and refining every step of every pipeline to ensure we’re on top of the latest instruction, evaluation and performance considerations.

But meanwhile, we threw together a musical track that incorporates a bunch of different advice for product people and agent builders everywhere. This isn’t a generative AI content demo (although Eleven Labs does read our words out), just a packaging up best practice ideas from our experience over the past couple of years.

Go have a listen :

Welcome to AI Field Notes by Move 37, an AI product studio working with ambitious organisations. Find out more about our work here.

In strategy, some of the most valuable insights often come from experiments at the margins. While a lot of discussions about AI focus on access to the latest releases from the major global labs, I've been testing something different: what happens, and how does it feel, when you decouple AI from connectivity entirely and take it on the road. And then try to talk to it.

This is has been my experimental setup for the past few weeks : A powerful Mac housed in a rugged Pelican case secured to the floor and charging from the supplementary battery that sits behind the passenger seat of my 2017 Volkswagen Multivan. The Mac is running Ollama, loaded with various open-source language models and a pretty hokey voice assistant in a pipeline that pulls it all together.

I don’t code (I obviously “vibe code”) so the setup is straightforward : the MacBook connects to the van's audio system via Bluetooth, with basic speech-to-text and text-to-speech interfaces handling voice interaction.

This isn't about practicality (the voice interfaces remain suboptimal, read “a bit shitty”) but it’s not hard to project forward a month or two to see where this will be.

In recent times a lot of us working in the space have become increasingly uncomfortable with the concentration of compute and capability in a tiny bunch of generalised models and the companies that have created them. We’ve done countless workshops in in the past few years where we’ve hypothesised that nation states (including Australia - see Kangaroo LLM), large corporations and other organisations might develop their own foundational models. This was more of a hopeful prediction than a techno-futurist one. Turns out that economics of that kind of thing are far beyond all but three to four companies in the world, for now.

But then recently DeepSeek's release of their R1 model demonstrated that open-source models are approaching performance parity with proprietary systems like GPT-4. This suggests we might finally see that locally-deployed AI can deliver capabilities previously assumed to require cloud infrastructure. This isn’t the place to debate how DeepSeek put it together, whether they’re a hedge fund shorting Nvidia or if the inability of the deployed DeepSeek Assistant app (different to the model) in the app stores should or shouldn’t talk to about Tiananmen Square. The reality is that for me, it was a hopeful sign that the landscape could be more competitive than we’d previously thought (I just remembered that time an Australian VC asked Pan and I “what’s the point of Open Source anything?”).

So with all this new #vanlife capability now functional, I’ve obviously been imagining remote wilderness situations where I might need to rapidly get up to acquire new knowledge (note: apologies to Move 37 team members who had to listen to me test this in the office by saying “teach me how to fish”)

Infrastructure failures following natural disasters and (for a touch of the fantastical) zombie apocalypse scenarios that test the system's ability, provide both practical guidance and creative problem-solving. These thought experiments pointed me towards what is perhaps a significant advantage of edge AI: resilience.

In critical applications from healthcare to disaster response to remote operations, systems that can deliver intelligence without connectivity potentially demonstrate an increased reliability. Obviously this would need to be intelligence that is baked into localised models, datasets and knowledge bases (offline, remember) - but it’s not difficult to see how that can all come together.

The current cloud-centric AI paradigm has created a pretty fragile dependency chain. When connectivity fails, intelligence disappears exactly when it might be most needed. Of course my local computer needs power too, so the jury’s still out on which paradigm has the best in-built ability to cope with different failure modes.

My Van LLM experiment suggests we're entering the early stages of a new distributed AI phase driven by three converging factors:

* Hardware advances: Devices like the M4 MacBook Pro (or very cool new Framework PCs) deliver computational power that would have been unimaginable in consumer hardware five years ago.

* Model efficiency: Open-source LLMs are becoming simultaneously more capable and less resource-intensive.

* Interface improvement: While still imperfect, speech interfaces are approaching usability thresholds for most applications and most situations (That said, the background noise from the zombies trying to get through your air vents is likely to mess with Coqui-ai or pyttsx3)

For businesses, all of this suggests a near future where critical AI capabilities could be deployed without creating new dependencies or exposing sensitive data to third parties. The compliance and security implications warrant real serious strategic consideration.

My Volkswagen Multivan LLM experiment started as a provocation—a way to test assumptions about what's possible with today's technology - and what’s not. Yet.

The little computer behind my passenger seat, capable of sophisticated reasoning without connectivity, feels like an early indicator of a coming type of intelligence that becomes increasingly distributed and sovereign—creating new possibilities for resilience, privacy, and independence from centralised control and policies.

Van LLM : Stick it in your big car.

Remember the first time you used Google? That moment when you realised the world's information was suddenly at your fingertips? We're living through a similarly transformative period with artificial intelligence, and at Move 37 in 2024, we've had a front-row seat to seeing this happen.

Since 2017, we've been immersed in the world of AI development, but 2024 has been a pivotal year for how businesses and individuals interact with this technology. Working alongside some of Australia's most innovative organisations, we've seen not just how AI has evolved technically, but more importantly, how it’s changing the way people work, think, and tackle challenges.

Building organisational AI brains

One of our most exciting areas of work this year has been developing what we call "organisational brains" - AI-powered systems that transform how companies understand and utilise their institutional knowledge.

Think of it as having a brilliant colleague who's read every document your company has ever produced, understands every process, and can explain complex ideas to anyone, in their language.

We've seen firsthand how these systems can bridge communication gaps between different departments. When a CFO and a CMO need to understand the same initiative, the AI can adapt its explanation and areas of focus to match each executive's perspective and priorities. It's like having a universal translator for organisational knowledge.

AI orchestration : music to our ears

A lot of us have worked in, or are currently working in, companies where digital transformation projects seem to consume years, masses of human effort and millions of dollars.

What’s been super encouraging this year is that we’re seeing that AI can serve as a sophisticated interface layer that makes key aspects of this kind of work more efficient and much less painful. Rather than completely rebuilding legacy systems, we can now use carefully engineered AI interfaces to make existing infrastructure more accessible and useful.

In practice, this means a customer service representative can interact with multiple backend systems through a single, intuitive AI interface. It means HR platforms can become more responsive and human-centric. It means product catalogs can become dynamic, intelligent resources rather than static databases.

The AI Agents are here

This year marked our first major deployments of AI agents - autonomous systems that can navigate complex workflows, make reasoned decisions, and take action.

For one ASX-listed company, we developed an agent that's revolutionising how they handle routine operations, helping making decisions that previously required hours of logging into different systems, trawling through data or searching through knowledge bases.

Go to where the work is

One crucial lesson from our work this year: chat interfaces aren't always the answer. We've developed AI solutions across multiple platforms - from sophisticated web interfaces to Slack bots, Google plugins, and command-line tools.

The key is meeting users where, and how they work, ideally not making them adapt to another platform. We’re desperate to have our AI applications and tooling used, not just deployed. We’re learning a lot about collaborating with internal teams to make sure that our products don’t just work, but are in fact a delight to use every day.

Learning, Adapting, Improving

In 2024, we’ve also continued to apply a lot of focus to the Move 37 methodology and practice itself. Experimentation and iteration have always been at the heart of building effective AI. We start small, test ideas in real-world conditions, and refine based on what we learn.

It’s how we make sure these systems stay useful over time.

A big part of this process is creating evaluations and benchmarks that fit the specific needs of the business problems we’re solving. We recognised the importance of this early on, and it’s become an essential part of our AI engineering methodology — practical, focused, and geared toward real progress. This approach also tackles safety and reliability, making sure our AI isn’t just effective but dependable and secure.

Looking Ahead

As we barrel towards 2025, we're seeing parallels to the early days of the internet. Just as we once had to explain what a website was before discussing its potential impact, we have often find ourselves providing AI primers before diving into specific opportunities. The understanding gap is vast out there - some organisations are deeply versed in AI capabilities, while others are just beginning to grasp its potential.

But we're seeing it happen - that instant when someone realises AI isn't just about chatbots or image generation, but about fundamentally transforming how we work with information and make decisions.

And while 2024 was mostly about individual worker’s AI education and exploration, it’s now strikingly clear that 2025 will be the year of accelerated enterprise adoption. We’ve really found it inspiring to see companies and teams move from individual prompting to workflow design to building incredibly innovative AI applications. Some are uncovering whole new IP and growth opportunities on the back of these efforts. We’ve felt privileged that organisations have trusted us to help architect and deliver on these initiatives and build out their visions and roadmaps for the coming years.

About Move 37

We build AI that works. No buzzwords necessary — just intelligent systems that solve real problems and create new opportunities. We start with your organisation’s needs and end with tools people actually want to use. Think of us as your AI studio: we design, we build, we maintain and we make sure it lasts, and gets even more useful over time.

Wondering what we might build together? Let’s talk.

And incredibly, as they say, timing is everything. In June 2017 the seminal paper "Attention is all you need" sent shockwaves through the AI research community and the Transformer model (the "T" in GPT) was born. This completely changed the way we and others thought about the the emergent opportunities in and around human and machine collaboration.

So for the next few years our team researched, designed, developed and deployed AI products. And boy, did we learn a lot. There's a tangible difference between the demos and experiments you see online, and creating highly usable, stable, consistently high performing Generative AI product. And in this realm, learning by doing is the only way.

When the ChatGPT moment came in late 2022, everything changed. Although the model itself was probably six months old, this launch introduced LLMs to a huge range of people for whom AI was largely seen in science fiction movies and read about in books.

And in the subsequent weeks and months, our phones started to ring and our inboxes binged. A lot. A lot of people who'd been watching AI's progress with interest from afar were now keen to understand how these new capabilities would impact them, their teams and their organisations as a whole.

At this point, we made a big decision. Inspired by the conversations, ideas and opportunities that were suddenly circling around us, we pivoted Move 37 to a whole new operating model. We recognised that the experience and expertise we'd acquired from being hands-on with this tech for years, now held incredible value for enterprise clients. It's impossible to truly acquire the kind of applied knowledge we had without doing the hard yards.

We are now applying our a vision and foresight, expertise, methodologies and capabilities to a range of emerging use cases. We believe most companies are going to need to explore AI product, application, tools and workflows. And they're going to need a trusted partner to inspire, inform and guide them.

Move 37 is an AI Product Studio for the ambitious. If that last bit sounds like you, we can't wait to see we might build together.

Dave

In our experience across a range of projects, products and sectors, we've found some common principles that are useful to keep front of mind as you approach planning and scoping Generative AI initiatives:‍

Dive in, with a business-driven mindset.

Organisations should take a dual approach to experimentation. One, focused on "low-hanging fruit" opportunities using consumable models and applications to realise quick returns. The other, focused on reinvention of business using models that are customised with the organisation's data. A business-driven mindset is key to define, and successfully deliver on, the business case.

Take a people-first approach.‍

Focus on people as much as on technology, ramping up talent investments to address both creating AI and using AI. This means developing technical competencies like AI engineering and enterprise architecture and training people across the organisation to work effectively with AI-infused processes. Given the topical nature of AI, we tend to find a very wide range of knowledge levels and opinions when it comes to integration of AI into the workplace. It's super important to bring people along for the journey, educating and inspiring as you go.‍

Get your proprietary data ready.

Foundation models need vast amounts of curated data to learn and that makes solving the data challenge an urgent priority for every business. Take a strategic and disciplined approach to acquiring, refining, safeguarding and deploying data. Ensure the organisation has a modern enterprise data platform built on cloud with a trusted, reusable set of data products.

We've learned a lot by helping teams and organisations work through exciting opportunities in the space. Get in touch if you'd like to hear more.

Dave

Since 2018, foundation models have been surging with capability and, more recently, popularity. Enterprise is now ready to talk advantage of this by building 'organisational brains' on top of them. When we talk to clients about these kinds of projects, there's a certain flavour of conversation that emerges: "isn't this what the intranet was mean to be?". Pretty much, but how there are more compelling interfaces, more powerful and accurate discovery and retrieval techniques and emerging LLM capabilities.

This is about the opportunity to create an intelligent system that ingests all kinds of unstructured data and is increasingly able to deliver accurate, meaningful, and actionable insights to help with decision-making.

Here is an emerging set of considerations, opportunities and capabilities we're working on in this space:

‍Deployment of LLMs and AI alongside internal systems and processes using a secure, optimised data foundation – retaining full control over what the AI has access to. We choose from commercial or open-source LLMs.

Operational Agents & Actions : we're currently engaged in work where we're creating AI agents powered and orchestrated by LLMs that can understand, recommend, and trigger business processes and actions.

‍Human-in-the-Loop : at Move 37 we consistently implemented the use of AI and LLMs with interfaces for decision making, feedback, and safe hand-off functions among agents, other models, and operators.

‍AI Control & Guardrails : applying AI guardrails to define what actions LLMs and AI can take and where to include a human-in-the-loop is super important. Additionally, as with any technical infrastructure build or integration, being aware of existing organisational policies, like access rights, is essential in implementing complimentary AI systems.

Monitor and evaluate all AI activity. Audit every prompt, output, explanation, recommendation, and action taken to enable compliance with the law and applicable regulations.

We're working on a diverse range of cutting-edge projects incorporating these approaches and technologies. What will you build?

Dave