Steve Hind is co-founder and CEO of Lorikeet, an Australian AI customer support company that have raised over AU$100 million and expanded to the US and UK. Before that he was a product lead at Stripe and Watershed.

He defines being AI native as looking at every problem as an opportunity to define work and hand it to an agent, so the amount one person can get done stops being bounded by hours in the day.

Measured against that, Lorikeet has seen substantial growth across 2026 and virtually no headcount change. Steve walks through how: why it was easier for engineers than for operations teams, why he now believes it's easier to change agents than people, what he puts in the knowledge layer and what he refuses to, and why he thinks most companies will end up with only engineers, sellers and owners.

What you'll take away:
› AI native, defined: "the amount one person can get done or supervise is effectively unbounded" because it's bounded by their ability to package work for agents, not by hours.
› Engineers adjusted fast because the job stayed the same; they climbed "a new rung on the abstraction ladder". Operations teams face a bigger jump: the job itself became building systems.
› "It is easier to change agents than people." His Stripe-era "Good Updates" doc is now a skill that drafts the update and pushes it to the person to edit and send.
› Don't ask people to automate their own job. Ask what's valuable that they never get to. "Instead of automating one job, I might be able to automate five jobs."
› Grinding as a service: a PM once watched a month of sales calls over a weekend and found a feature coming up on 40% of them. Agents now do that on demand.
› Put the raw terminal in non-engineers' hands. Claude Code beat the wrapper products at Lorikeet, even for people who had never opened a terminal in January.
› What still fails: AI writing full of rhetorical flourish, and agents that report ten crises a day when one is real. "It has the energy of a very junior chief of staff."
› Only two roles are certainly durable: engineer and seller. Maybe a third: owner. Hire for builders, and use work trials rather than interviews to find them.

Links
› Steve Hind: https://www.linkedin.com/in/shind/
› Lorikeet: https://www.lorikeetcx.ai/
› Alasdair Bell: https://www.linkedin.com/in/alasdairbell/
› Fractional Head of AI: https://www.alasdairbell.com/

Hosted by Alasdair Bell, fractional head of AI for startups.

Chapters
00:00 Cold open
02:39 Meet Steve Hind
03:06 What AI native means at Lorikeet
04:34 Two kinds of adjustment pain
06:18 Engineers adapt, ops rethinks
08:03 Easier to change agents than people
10:23 One builder per ops team
11:50 Pair builders with experts
13:06 Blessed patterns for agents
14:57 The knowledge layer: build or buy
17:02 Battle cards built at runtime
18:12 Don't automate your own job
19:59 Grinding as a service
22:59 What AI still gets wrong
25:23 Why AI can't write
26:22 Claude Code for non-engineers
28:48 Noncompliance is a system problem
29:56 The last 20% of adopters
31:57 Can versus should
33:59 Engineer, seller, owner
34:54 Hiring for builders
36:30 Work trials over interviews
37:41 Work in ten years

Matt Geleta is Director of AI Operations and Enablement at Harrison AI, one of Australia’s leading AI companies, its radiology model passed the UK’s toughest exam and beat every frontier lab, it’s in 40 NHS trusts, and it recently raised a US $112 million Series C.

You’d assume a company like that gets a free pass to becoming AI native. It doesn’t. The pace of change means even Harrison is racing to keep up, so Matt built a dedicated team whose entire job is making the company AI native. He walks me through the whole model, the agent they had to build because nothing off the shelf was good enough, and why he thinks AI should make us better humans, not just more productive ones.

What you’ll take away:

• An AI company isn’t automatically AI native. The technology moves faster than any organisation, so Harrison built a dedicated team to close the gap.

• The model: one full-stack, generalist team across four pillars — infrastructure, knowledge, tooling, enablement — plus AI champions embedded in every function.

• The biggest unlock was context. Pensive, their in-Slack agent, reads emails, Confluence and Slack through a knowledge graph and vector database that compounds over time. It writes tenders and automates business cases.

• Optimise for speed, not cost. “There’s only so much cost that you could ever take out, but there’s really no limit to how fast the company could move.”

• AI can make you a better human, not just a more productive one. Matt’s personal agent reviews everything he writes and says, then flags his EQ and knowledge gaps every week.

• Juniors: hiring has slowed for grads across the market, but younger people often upskill on the tools faster than anyone.

• Chapters

00:00 Cold open
03:56 Meet Matt Geleta
05:47 The title he won’t put on LinkedIn
07:36 Why commit fully to AI now
10:29 The real cost of waiting
11:33 Harrison’s principles for AI
15:16 Bringing every person along
17:31 Even an AI company had to change
18:36 Superhuman, not cost-saving
22:55 Building the AI ops team
26:49 A full-stack generalist squad
31:07 OKRs and owning the work
34:27 Do we still need juniors?
39:46 Pensive, the in-Slack agent
45:07 Matt’s personal second brain
50:53 The weekly coaching loop
54:04 AI making us better humans
55:23 What hasn’t worked
57:21 Fighting work slop
1:01:00 Why today looks anarchic
1:05:08 A bright, unknowable future

Links

› Matt Geleta: https://www.linkedin.com/in/matthewgeleta/
› Harrison AI: https://harrison.ai/
› Matt’s website:

Paradigm

Conversations with the world's deepest thinkers in philosophy, science, and technology.

By Matthew Geleta

› Turing Works: https://www.turingworks.ai/

Hosted by Alasdair Bell, fractional head of AI for startups and partner at AI consultancy Turing Works.

Josh runs Ballpoint, a UK growth agency managing £60M of brand revenue with a team of 12 — and 90% of its work is now AI-first: AI does the task, a human reviews it.

He also tells me why he thinks the agency model is going to die, and what he’s building to replace it. We unpick exactly how he got there: the two-day automation sprint that failed, the January Claude Code month that didn’t, the memo and off-site that reset the culture, the team members who chose a different future, and the agent named after a Mad Men character that writes better meeting notes than any transcription tool.

What you’ll take away:

• AI-native means AI does the task and you review it. Assisting is the stage before and most teams are stuck there.

• The adoption playbook: one memo making the change non-optional, one off-site forcing everyone into the command line, build sessions every other Friday.

• Context is the difference between a demo and a tool. Cosgrove, Ballpoint’s account-manager agent, works because it reads Notion, Slack, and meeting history — and was tuned against a scoring system until it hit 95%.

• Build the data layer first. Models and tools on top swap in and out.

• Junior roles as we knew them are disappearing. Josh’s bet: apprenticeship-style learning by hand before you’re allowed to delegate to AI.

• For grads: build something real with Claude Code and put it in a public repo. CVs are dead.

Custom Agents take Notion’s AI and give it tools and triggers. It’s powerful because what are we humans but a brain + tools? Notion have given us a way to make AI workers.

The possibilities are the problem. You can trigger your agent on Slack, access your CRM, send an email, document a bug, book a meeting, write cat poetry…. Endless possibilities.

And that’s Notion’s blank page problem. Users sign up, are greeted with a blank page, don’t know what to write, and give up.

Creativity needs constraints. Here’s yours: Focus on the problem.

Notion Custom Agents are good at plenty of problems, so start with a quick win. Getting tasks into a database from Slack.

I’ve got dozens of databases: tasks, projects, decisions, issues. Getting a new item into the right one means opening Notion, finding it, and fiddling with the fields. A minute of faff, which when busy, is enough friction to break your flow.

Let’s fix that in <10 minutes.

I’ve recorded myself building this agent step by step. Follow along and you will get two things:

1️⃣ A handy agent that adds tasks to Notion from Slack

2️⃣ The skill of building agents

And once you have a hammer agent, everything starts to look like a nail automation. Making the next build much easier.

Still stuck? You can always hire Turing Works to do it for you 😊

And yet OpenAI’s Atlas feels like a new paradigm in browsing the internet.

Paradigm is an abused word. When something changes a little, it gets used a lot.

But Atlas is the biggest change in how I browse the internet since I escaped AOL’s walled garden onto the World Wide Web.

The old way we browse is search first. Atlas is AI first. In simple terms, the search bar becomes the AI chat bar.

Subscribe now

I’m finding information faster. Instead of the first page of Google, an enshittified mess of ads and SEO, you get a summary of what you need to know.

When you go deeper? Search used to end once you clicked into a site. With Atlas, you keep chatting. ChatGPT sits in the sidebar, ready for follow-up questions.

There are even glimmers of autonomous browsing through Agent Mode. Though for now it feels like watching your parents use the internet complete with ponderous thought process, slow cursor movements, and occasional surrender.

My only criticism is the app icon... It’s dark and loses the “explore the web” design of other browsers. So when Chrome was still open I found myself teleported back by keyboard muscle memory.

Curious if others feel the same shift?

Subscribe now

The comment was due to Jessy Wu posted that she wasn’t letting her team at her communication agency use AI for ghostwriting, it went against my whole belief system. How could technology make things worse? Surely we can leverage AI to help us do better work?

One of Jessy’s objections stuck with me. She claimed that using AI, rather than the process of writing and rewriting “atrophies the muscle”, making us worse wordsmiths.

But we don’t use AI just for writing, my main use is to help develop my thinking. So what if using AI is making me dumb?

One month later I think it is.

However I believe AI can make us smarter too.

Subscribe now

Use it or lose it

It turns out this proverb is scientifically backed.

The 2020 study “Habitual use of GPS negatively impacts spatial memory during self-guided navigation”, showed that using GPS makes us worse at navigating over time by weakening our spatial memory.

To add insult to injury, the study found no significant relationship between a person's self-reported sense of direction and their actual navigation skill, meaning we may still think we are good at navigating even after our ability has lessened.

A second (weaker) study “AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking”, makes the same point but more directly relates it to the impact of using AI. Their surveys showed using AI more make people’s critical thinking ability worse. Concerningly they also found that younger people “cognitive offloaded” more than their elders.

So using AI makes us dumb?

Only if we let it.

The case for optimism

Coaches make us better.

They could do the work for us, because they are experts at the activities, but they don’t - that would be pointless. Instead, they watch what we’re doing and help us work towards a better version of it.

In the Kraft et al (2008) meta study it was found that coaching is more effective than performance incentives and professional development workshops. The impact of coaching was equivalent to several years of experience. So it clearly works.

This is where I see the difference between AI making us dumb vs making us smart. We need to use it in a way that coaches us, rather than does the work for us.

Using AI like a coach

So how do we use AI like a coach, rather than it just doing the work for us?

It comes down to two things: how we use the tool, and the tool itself.

1. How we use it

A coach never does the work for us. So it’s important to do the work ourselves first, rather than just asking for the answer in ChatGPT.

That means running through our own mental models, then using prompts to get feedback, broaden our thinking, and apply rigour. Think of it like working with a thought partner or manager— not a replacement for your own effort.

2. The tools themselves

The second part is about the tools themselves. The character of pretty much all AI models up to now is that of a helpful assistant (apart from Grok which chose chaos).

If you look at the system prompt for Claude, it mentions “help” or “helpful” 21 times, but doesn’t reference “coach”.

They have built a question (or homework) answering machine

Claude should give concise responses to very simple questions, but provide thorough responses to complex and open-ended questions.

That’s not the only way AI could be built.

“Study mode” is a promising OpenAI innovation to help students learn.

I’m excited to see more experimentation with manager or coach agents, which could challenge us, not just do the work for us.

A warning from social media

We can’t leave this to chance. Social media had utopian ideals which were lost in the push for advertiser profitability.

• OG Facebook was about connecting with friends.

• Current Facebook, is about maximising screen time using variable-reward newsfeeds filled with socially useless short video to help advertisers reach customers.

Engagement and ad revenue is excellent - Meta is a two trillion dollar company.

But we are lonelier than when it was founded in 2004. Facebook has failed on its mission to “bring the world closer together”.

Source: Our Epidemic of Loneliness and Isolation, U.S. Surgeon General, 2003

I consider the enshitification of Facebook a result of it being a service paid for by advertisers. Ultimately the company goal is to increase user engagement, not social connection.

But we are paying for AI. So maybe we can determine our fate.

In summary

I remain a techno-optimist, albeit a cautious one. AI can help us achieve so much more. But we need to be careful in both how we use it, and how we design it, so that we don’t lose our ability to think for ourselves.

Because we know where that leads.

The closest we have to a universally hated file format is also the PDF.

What it gives in consistent formatting across devices. It takes away with an inconsistent ability to work with the content. And don’t get me started on editing.

So when we were approached by an investor to help them get data out of PDF board meeting decks I didn’t know if we should help or run.

In this article I’ll run through the challenge they were facing, how we fixed it, and share the prompt and N8N workflow so you can make your own PDF extraction automations.

Subscribe now

The problem

Our client was spending hours before board meetings checking KPIs given in the board pack. That meant:

1. Find the KPI (say ARR)

2. Looking for what it was reported as the previous month and comparing

They would do this for tens of KPIs which could take hours and would inevitably involve human error as we are unreliable meat computers.

And even then the analysis was written on a PDF. Not even the best RAG search engine would be able to return those results.

The solution

Our workflow took the PDF from Google Drive, extracted the KPIs using AI, and then put the data in a spreadsheet for analysis.

This was quicker, more accurate, and put the data in a format more amenable to analysis.

The workflow

1. Trigger: New file in Google Drive folder

2. PDF AI Extraction: Gemini 2.5 Pro for most accurate result (explained below)

3. JSON parsing: Getting data structured for next steps

4. Add to spreadsheet: Template shared below

You can download the JSON template here

Prompt & Model

It’s set-up with 20 common KPIs from board decks. However it’s easy to change them to whatever you need.

I’d suggest asking AI to make edits to which KPIs are required so you don’t have to fiddle with the JSON. You will also need to change the validation example in the JSON parser, just paste it into a LLM and ask AI to change the KPIs there too.

You can download the prompt here

How to get to 99% accuracy

It’s all about the model chosen and tuning it for your file:

1. Gemini 2.5 pro is the most accurate AI for reading PDFs

Benchmarks from the Scale leaderboards back up what I discovered through painful trial and error, Gemini 2.5 is the best for visual understanding.

It’s not the cheapest or quickest, but it sure does read PDFs well.

2. Read its thoughts to tune the prompt

2.5 Pro is a reasoning model, so if it gives you the wrong answer, you can read its reasoning to find out why. Interestingly it usually had a good reason for giving the wrong answer, like data duplication or inaccurate charts (yes it looks at those too).

Tuning the prompt for each document allows you to tackle these document specific issues and get to 99% accuracy.

Spreadsheet Template

Here’s the template of the spreadsheet required to run this workflow. Create a copy of the file and link to it in the sheets node.

About Me

I’m Alasdair Bell, founder of Turing Works. We help companies automate boring workflows, like digging KPIs out of PDFs so you don’t have to.

📧 alasdair@turingworks.ai

🌐 https://www.turingworks.ai/

• Email newsletters - you can now add your newsletters to your reading list by tagging them. They get downloaded, summarised and add to your list daily ready to be read at a more convenient time.

• Notion 3.0 - the Notion AI template includes more AI features. You can also ask it to research books on Goodreads (but can’t trigger the AI to search the web when the page is added like with n8n).

Pocket has run out of things to save, primarily, itself.

So since Pocket shut down, I’ve built my own reading list app that saves and summarises blogs, YouTube, PDFs, and even books.

I treated Pocket as my junk drawer for blog articles. Anything Present Me didn’t have time for got passed to Future Me, a person with abundant time and attention.

This kept Present Me focused on the present. Future Me inherited the problem and dealt with it in the same way as the junk drawer - by ignoring it.

My Pocket replacement does even more than the original:

• 🌍 Captures everything: Saves articles on mobile or browser with a click

• 📚 Multi-media: Works for YouTube, PDFs, blogs and even books

• 📝 Summarises: Quotes, key points, categorisation, reviews and overview

I’m sharing two different versions depending on your technical ability:

• 🗄️ A Notion template — Two minute set-up, captures everything, summarises, but can’t digest YouTube, books, or PDFs

• 🤖 N8N workflow + Notion — 15 minute set-up (need a couple of API accounts), summarises anything, better reading list experience than Pocket.

Subscribe now

Here’s an overview of how to set-up your own version

Option 1: Simple Notion + AI Database

• Spice level: 🍋🌿 Lemon and herb

• Requirements: Notion + AI add-on

• Template: Copy to your Notion workspace

Walkthrough of features and setup

Summary of workflow

You clip an article with the Notion Web Clipper (or share to the Notion database when on mobile), and this adds it to your reading list. Future You will find the article waiting, including the full-text you can read offline — a feature Pocket made you pay for.

Each item comes pre-digested using Notion AI:

• One-liner: the gist in a sentence

• Paragraph summary: a quick read if you’ve got a minute

• Key points & quotes: distilled insights, but with the author’s words intact. I prefer pulling out quotes rather than letting AI paraphrase everything. It keeps the meaning sharper, and the language closer to the source.

Once you’ve got a few articles in your list, you can switch between different views:

• Feed view: skim the opening of each article like a newsreader

• Table view: organise by topic, date, or priority

Set-up takes two minutes. All you need is Notion and the AI add-on enabled. After that, just duplicate the template, choose where to store it, and start saving.

If you want the “lemon and herb” version of a Pocket replacement, this is it.

Option 2: Spicy Notion + n8n Auto Summary Workflow

• Spice level: 🌶️ Peri-peri hot

• Requirements: Notion (no AI) + N8N + API keys (OpenAI, YouTube Transcripts)

• Templates:

  • Notion database template

  • N8N template

Walkthrough of features and setup

Summary of workflow

This is the full-spice version, a full Pocket replacement that goes beyond articles. With N8N augmenting a Notion database, you can capture, summarise, and organise:

• Blogs: Full-text capture, summarised automatically

• YouTube: Transcript pulled in, key points extracted

• PDFs: Text extracted and digested into summaries + quotes

• Books: Reviews and ratings fetched from Goodreads, distilled into highlights

• NEW Gmail: Stores & summarises email newsletters labelled in your inbox

Like the Notion template, everything gets broken down pyramid-style: one-liner → paragraph → key points with quotes.

How it works

You clip content into Notion with the Web Clipper (or share from mobile to Notion). That’s the trigger. From there, N8N takes over:

1. Detects the media type (blog, PDF, YouTube, book)

2. Runs the right workflow (extract text, fetch transcript, pull metadata, etc.)

3. Sends it to AI for summarisation + formatting

4. Adds it back to your Notion reading list

The configuration is straightforward:

• Upload the provided N8N workflow JSON and duplicate the Notion template

• Follow instructions in N8N template to

  • Add Notion database to Notion trigger node

  • Connect your API accounts (OpenAI, YouTube Transcripts)

  • Adjust the summarisation prompt if you prefer different outputs (e.g. ratings, categories, longer digests)

Once it’s live, you’ll have a single database that handles everything from PDFs to YouTube lectures — all stripped down to digestible notes.

Leave a comment

Share

When I graduated, I was one of the few who didn’t join an MBB graduate scheme. I wanted to build things, so I joined start-ups instead.

Yet here I am, launching a consultancy, and I’m loving it.

Exploring AI has led to projects implementing AI. I’ve seen the technology reduce workload, improve quality and increase capability. But I also see a gap between what’s possible and what companies are doing. Turing Works will bridge that gap.

What does Turing Works do?

We tackle AI adoption in two ways:

1. Training organisations to become AI-first: embedding gen-AI into the everyday tool-stack, up-skilling teams to prompt, experiment and build.

2. Building AI agents and automations: reduce manual steps, so quotes appear in minutes, data analysis just happens, and reports write themselves. Work gets done faster, cheaper and more accurately, freeing up teams to do more.

Who are we?

We’re two former Uber operations managers turned founders and MBB consultants, happiest in the thick of complex operations, messy workflows and hands-on implementations.

What kinds of companies?

Our experience lends itself to SMBs with complex operations (wholesale, construction, manufacturing). Where teams need off-the-shelf fixes rather than custom-made tools they can’t maintain.

That’s still a hypothesis, so tell us your toughest workflow or AI hurdle and let’s test it together.

Get in touch!

Below is my summary of the best techniques from Google, Anthropic and OpenAI. It’s one page long.

Prompt Improver Prompt

Here’s the prompt I use to improve my prompts. Just put your prompt between the <prompt></prompt> tags.

I find it works well to point me to techniques which I’ve overlooked in the guide. Remember to attach or link (LLM allowing) the guidelines you want to reference:

• Google Prompt Engineering Guide PDF

• Anthropic Prompt Engineering Guide

• OpenAI Prompt Engineering Guide

# Role

You are an expert prompt engineer who likes to help other people improve their prompts so they are more effective

# Task

- You will be given a prompt below which the user would like help improving
- You are to provide advice based on the Google Prompt Engineering tips attached
- When providing your advice please include quotes from the guide attached and reference which page it comes from. This helps the user find more information about the
- Finish your advice with an improved version of their prompt with your advice included


<example>

# Role

Including a role will help the LLM understand what kind of output is most helpful. This is explained in p21 of the guide "Defining a role perspective for an Al model gives it a blueprint of the tone, style, and focused expertise you're looking for to improve the quality, relevance, and effectiveness of your output."

Improved prompt suggestion

...

</example>


<prompt>

[Insert prompt you want improving here]

</prompt>

But then I got carried away. The guide’s now massive (and growing), because agents touch nearly every part of AI implementation.

So here are six key lessons so far:

1. 💼 Working with chatbots is like managing an intern on their first day. They don’t have the tools, information or business understanding to get work done - so we hesitate to give them any.

2. 🔨 Agents give LLMs the tools, workflows, and data they need to fully own tasks, not just assist.

3. 🧑‍💻 Technical ability is no longer a barrier to automation. I’m a no-code kinda guy but I’ve had no problems with some-code tools like N8N as ChatGPT offers the worlds best tech support.

4. ✏️ Start simple. A well written prompt + data is a great start and often enough. Complex systems hide issues, so only add advanced workflows when absolutely necessary.

5. 🔁 The most successful agents so far (coding and customer support), both have straightforward validation steps (compiling successfully or customer satisfied). Checking work and feeding back suggestions are even more important in agent systems as LLMs prioritising helping to the point of making things up.

6. 🤖 The potential for agents goes beyond task automation. LLMs may have human-like intelligence, but they’re also tireless machines. Why not research that subject? Reply to those emails? Delegate that task?

Still reading? Love long documents about robots doing your job?

OK, here’s the 👉👉 full guide 👈👈

"Although these transformer models were trained to predict the next word in a text, they do a surprisingly good job at other language tasks"... "it may be that transformer models and their relatives are learning latent representations that capture the same basic ideas as grammars and semantic information, or it may be that something entirely different is happening within these enormous models; we simply don't know."

Artificial Intelligence, A modern Approach, Stuart Russell and Peter Norvig

The most popular textbook on AI, used by more than 1,500 universities, can’t explain why large language models work so well.

These models were designed to predict the next word in a sentence. Then we fed them more data and trained them with faster computers. Now, they’re passing medical licensing exams.

That insight blew my mind.

The unexpected intelligence of large language models has changed how I’m approaching AI implementation. Instead of purely using first principles, like an engineer calculating the load on a bridge, I must be more experimental. Given we don’t fully understand how these models are so capable, it’s better to hypothesise their abilities and test to see if I’m right.

I was going to write “building with LLMs is more of an art than a science.” But that’s wrong. What I’ve described—making hypotheses, testing, refining—is literally the scientific method. So it’s more accurate to say, “building with LLMs is more of a science than engineering.”

That’s why it’s time for me to close the textbooks and start experimenting.

“The world was to me a secret which I desired to divine.”

— Mary Wollstonecraft Shelley, Frankenstein

What’s next:

• Given the rapid pace of AI development and fire hose of new research, I’m going to narrow my focus to make progress mastering AI implementation. I’ll tackle one area at a time (e.g. prompting, agents, image generation).

• I’ll focus the investigation of each theme around a project which will allow me to experiment and learn.

• I’ll share learnings on Substack, but also keep my AI Implementation guide updated as a reference wiki.

Alas, solving problems with AI isn’t as simple as asking ChatGPT for the answer. Sure, it might be right — or it might be confidently wrong, or just behave weirdly.

I need a first-principles understanding: being able to break down complex problem into their fundamental elements. But first-principles thinking only works if you understand the fundamentals - How AI works.

That’s why I’ve written part 1 of my AI Implementation Guide — “First Principles.” It explains the AI basics necessary for business implementation:

• 🧩 The core components of large language models and image generation

• 🤖 How these models work and how they’re trained

• ⚠️ Uses, pitfalls, and why it can behave strangely

(And surprisingly, the amount of magic involved is non-zero 🪄🎩)

If you're interested in implementing AI, fascinated by fundamentals, or just want to see token generation explained with ASCII art. This guide is for you.

This guide has been a huge learning experience for me, and I’d love to improve it further with your feedback. Please share any questions, critiques, or comments 🧠.

👉 Check it out here 👈

I would have hated working for Amazon.

Applying to join Amazon’s marketplace team felt like progress. With experience at four marketplaces and three e-commerce startups, I could have used my expertise in 'the world’s biggest startup.'

But it would have paralyzed my career and been a terrible experience.

Instead, I’m betting on AI and dedicating the next phase of my career to pursuing opportunities in this space.

This article explains how I made that decision. Maybe you’re interested in my career? Maybe you’re considering this shift? Maybe you’re my mum? You will all enjoy this article.

It’s a story of boy meets opportunity:

• How I evaluate opportunity. An equation of what’s still possible, multiplied by scale and skills.

• Why there’s no opportunity for me in marketplace startups. There’s less growth and so the focus shifts to efficiency which requires specialisation but I’m a generalist.

• The opportunity for Al(asdair) in AI. We are living through a technological leap enabling new products and services.

How I Evaluate Opportunity

“Luck Is What Happens When Preparation Meets Opportunity” – Seneca

Or, more completely:

Defining the terms:

I aim to maximise the value extracted from the opportunities I pursue. The value I care about is:

• Money is necessary and easily quantifiable. It’s an essential outcome but not everything — if I wanted to be rich and miserable, I would have become a banker.

• Fulfilment is harder to quantify but more valuable. It comes from living my values and doing work that matters to me. I care about making things, learning, and having fun.

The inputs to this function are:

• Opportunity I see as the gap between what people want vs what they have. People will pay you if you can help them bridge this gap. (Concept from Almanack of Naval Ravikant )

• Preparation refers to my relevant skills and experience which allow me to make the most of the opportunity.

• Market size gives the scale of the opportunity. Note the time function of market size matters more as this is an investment. My ability to grow financial and career capital depends on the difference between my entry and exit points.

Why there’s no opportunity for me in marketplace startups

• Opportunity: Low - the best ideas have scaled

• Market: Large, but low growth - Amazon has achieved scale and only grew 11% last year

• Preparation: Medium - I have broad experienced, but not the specialisation required at scale

  ---

My first job in a start-up was at onefinestay, a vacation rental marketplace founded in 2009. I was their second employee and worked in ‘operations’, a role best described as whatever we needed to do to make the business work. On my first day I organised the warehouse into a process, later I implemented a CRM system. We hired generalists as we didn’t know what role, responsibility and skills would be most important.

The late naughties were a great time to start a two sided marketplaces. Etsy was founded in 2005, Airbnb in 2007, and Uber in 2009.

The opportunity arose as internet penetration reached a level where a two sided marketplaces could achieve liquidity and scale quickly. Smartphones were now smart enough to let people use their service on on app on their phone.

But the wave has passed. If it works as a marketplace it’s probably been tried. The ones that work have scaled and are locked in a stable competitive market which would be difficult to disrupt head on.

Professional opportunity hunters (VCs - feel free to use this title) know this. Funding for marketplaces was only 3.6% of Australian venture capital in 2024, that’s 17th out of 25 categories.

Which brings me back to Amazon. A crushing moment in the interview process made me realise the issue. I asked the hiring manager what the main challenges were?

| “the same priorities we’ve had for the past few years"

I can’t imagine the drudgery of working the same problem every day. I like new problems and launching new products (six so far). This role would not have been fulfilling.

It also illustrates a shift in priorities as a marketplace scales. Instead of the foundational problem solving required to make a concept work. The focus is on improving efficiency at scale.

And the skills required to do this are better defined and less changing, so it favours specialist. Not me. Thank goodness Amazon recognised this and didn’t offer me the role.

Why I’m betting on AI

• Opportunity: large - the rapid development in AI capability has outpaced implementation

• Market: potentially massive, high growth

• Preparation: required - I have the necessary early stage skills, but need expert AI knowledge

The opportunity in AI is massive as rapid development in capability has outpaced implementation.

In a year and a half performance of AI has improved 1.5x and cost has decreased 99%

But AI implementation is lagging in most industries. Only developers have seen widespread adoption with AI tools reducing how long it takes them to produce code by 55%.

I’ve only seen basic use of AI in start-ups, mostly using ChatGPT as a work assistant. As a champion for AI in product research, I taught ways to use LLMs to improve our work - but now I realise that my approach only scratched the surface. So I don’t think most companies are getting the most out of AI.

The market for AI is massive. McKinsey predict “generative AI’s impact on productivity could add trillions of dollars in value to the global economy.”

It’s growing rapidly. OpenAI has grown revenue 248% in the past year. And the professional opportunity hunters know it. 82% of YC's latest startups are AI-focused. Global funding to AI companies increased 80% year over year from $55.6 billion in 2023 - a third of all VC investment.

Rapid growth means I would be able to grow with the company. Ownership is much more interesting in a high growth sector. And there are more opportunities to take on responsibility when a company is doing well.

But I need to work on my preparation. This can be considered a combination of role skills and subject knowledge.

• Role skills: Good - I have a broad, early stage start-up skills. I’ve validated new businesses, built products, solved scaling blockers and managed teams. In short I can validate, launch and scale new products.

• Subject knowledge: Required - In order to solve problems from first principles you need to know what the first principles are. To achieve this I’m going nerd deep into AI until I know more about AI implementation than anyone else in Australia.

What's next for AI and Al(asdair)

I’ve quit all current interview processes and given myself three months to transition by:

• Learning from first principles how AI works and to understand implementation better than anyone else in Australia.

• Practice implementing AI to consolidate learnings and demonstrate my credibility.

• Find AI opportunities by understanding the challenges of operators trying to implement AI.

I’m excited. Wish me luck!

But it didn’t work out. Why? As is start-up tradition I’ve written a failure blog to explain.

You did what?

Yeah, lawns. We helped people with lawns (est. half of all Australians) get great results. The key insight is that no-one knows what they should do with their lawn. And so with our tailored lawn subscriptions based on climate, soil, and satellite imagery we could send exactly what customers’ grass needs, when they need it.

And did it work?

Yeah, well — sort of. I mean the fertiliser worked. And it should have done, we got it developed by literally the world’s leading expert on site specific fertiliser recommendation. This stuff was science in a bottle.

Sort of?

Yeah. A great fertiliser does not a brilliant start-up make.

So what didn’t work out?

The ultimate cause of mortality was the #1 start-up killer. We ran out of cash.

But we tried to raise funds. Oh how we tried.

• 91 investors contacted

• 31 first meetings with investors

Why couldn’t we raise more money?

I’m glad you asked. Here’s a summary of the main feedback themes from investors:

Small market

VCs don’t get out of bed for revenues of under $100m in 5 years. Lower outcomes don’t “meet their fund dynamics” as each investment needs the potential to return the fund. This was compounded by Yarda having a retail business model which is harder to expand internationally (and thus expand your market) than something less physical like SASS.

My initial market estimate was around $1bn. Despite my market sizing skills this figure was never entirely believed and was a sticking point.

The trump card in this (every?) situation is traction. Ours was middling, 200 subscribers with a customer acquisition cost of $200. Not enough to make the opportunity compelling.

Learning: If you are going to start a venture scale business choose a venture scale market.

Not defensible

Direct to consumer went through a boom in the 2010s with companies like Casper, AllBirds and Dollar Shave Club scaling rapidly. Unfortunately the 2020s have been less kind to the model with one solitary $1bn+ exit. The D2C darlings were unable to maintain their early advantage as new competitors entered the market and established companies went online.

It’s just too darn easy to start a retail business. Which makes running one hard. It’s easier than ever to source a product from Alibaba and launch a Shopify. So it’s harder than ever to maintain your advantage if you don’t have some kind of defensibility to your product (tech, exclusive contracts, supply chain etc.).

This all results in investor confidence in DTC being at an all time low.

For an excellent analysis of the problems with DTC I recommend this article by Packy McCormick, The Hard Thing About Easy Things.

Learning: Invest in tech, not marketing.

Founder product fit / customer understanding

Would I die for this start-up? Do I live and breath lawn? Honestly I chose this business as I loved the model, rather than lawns being my passion. I don’t even have a lawn.

It’s easy to dismiss the importance of this. If businesses were only focused on founders interests we would have an overabundance of start-ups making productivity tools, web3 projects and improving food delivery (oh, wait).

But there’s a chasm between a product and an incredible product. And the insights which get you there are based on a deep understanding of your problem. And achieving that level of understanding is easier when you have a deep connection to the problem.

Learning: The problem must be your obsession

Ohhh nasty. But what did you get right?

The most positive feedback was for about the team. Phil (CMO) and Blake (Founder’s Associate) were highlighted as the right people to have along for the ride. I’m incredibly grateful to have worked with these two and for them to dedicate themselves to Yarda.

Why were you raising money in the first place?

Great point Alasdair, I’m glad you asked.

There’s another route to success with DTC — bootstrapping. This allows you to,

grow slowly, get profitable before all of your credit cards are maxed out. If you’re targeting small niches that you know how to reach without giving all of your margin to Google or Facebook, you can build a really nice business.

And that’s a really great position to be in. Your options are open on how to grow from there. You don’t have revenue targets which can only be achieved by gambling your capital on Facebook adverts.

But Yarda was a venture backed business. Flash Ventures backed us from the very start and were supportive throughout. I wouldn’t change anything about my investor. But they wouldn’t back a company promising slow growth and modest outcomes. So that course was set at the start.

💸 As an aside there has been a welcome trend of DTC start-up crowdfunding to fuel their growth (see Ouvira, Zero, Heaps Normal). Consumer investors don’t have a fund to return and so are happier with modest returns. And those passionate about the product are more likely to forgive it for being hard to defend. As a result some excellent Aussie DTC brands have grown with help from their customers. With more time and traction this would have been a great route for Yarda.

So why didn’t you pivot?

We weren’t fast enough. If I could open a portal to day 1 Founder Alasdair and shout something before it closed it would be “FAIL FASTER”.

Knowing what I know now I could work out that the market wasn’t attractive enough in a few months rather than a year. That would have left most of the money in the bank ready to give a crack at a related start-up using the insight from the first.

Part of that is better learning the skill of how to test a market. I like to think of this a reframing of launch. Instead of working towards a launch (product in hand of customers). You are working towards proving the product (website hits, sign-ups etc). As that means you can be making valuable contributions earlier in the process, rather than giving yourself the excuse of waiting until launch.

But the larger part is embracing ambiguity. Building a company from 0 to 1 is so hard for so many reasons. It’s the most exposed to failure I have been in my life. And that’s scary.

Deferring launch defers failure. And that feels good.

But the path to success is embracing the things which can kill you. Because only by resolving those can you progress.

Ultimately not enough people bought our product. And we didn’t know that until too late. By leaning into this challenge. Getting in-front of customers early. Learning, iterating, pivoting. We would have been more likely win. Waiting on “launch” only sheltered me from the storm. And gave me less time to escape it.