A field report from three months in my own lab – a 90-day, R975 ($60) experiment, and what it changed my mind about.
The blog is my lab notebook; this is a field note from it. In three months I ran an experiment on myself. Call it a self-made MBA. No cohort, no case studies, no R1.3M ($80k) tuition. One rule: ship things, measure everything, write it down.
Here’s the syllabus I set:
- Build an app for free – to learn mobile store deployment mechanics end to end. Not the code. The machinery. Developer accounts, review queues, the publishing lifecycle nobody tells you about until it bites.
- Release a music album and promote it aggressively – Instagram, TikTok, LinkedIn, TheKeegzVerse.com. Then read the analytics like a hawk. Content types, scaling, time-of-day posting. What actually moves a number versus what just feels good to post.
- Build a prototype AI agent as a writing partner – and accidentally build a mobile game with monetisation potential along the way. The best products are the ones you trip over.
- Open-source the wild ideas – a blog. Ad-free. No paywall. A deep dive of learnings and ideas, put out into the world to get real business feedback on. I like to think of this as my portfolio meets my thinking.
The constraint: two hours a day. One R325 ($20) AI subscription a month, alternating between Claude, ChatGPT, and Ollama + Qwen running locally. That’s the entire R&D budget. R975 ($60) over the whole quarter.
That constraint is the point. Anyone can do impressive things with infinite money and infinite time. The interesting question is what one person can do with the budget of a streaming bundle.
The tools, and what they actually are
Spend enough time with these things and you stop seeing “AI tools.” You start seeing personalities. You start casting them.
Codex is the sculptor. A great one, for creatives and rapid-prototyping teams. Around R325 ($20), used without precious rules, gets you roughly two to three hours of solid sculpting before limits. Think about what that replaces. A development team runs roughly R150k (~$9.2k) a month. In the right hands – for prototyping, experimentation, first-pass implementation – a R325 ($20) tool now hands a product person capabilities that used to require access to a whole development team. That’s the gap. The creative and product people no longer wait on a build team to figure out the nitty-gritty. The wait collapses from days to minutes.
In its spare time it does graphic design too, if you can hold the same patience through an hour of corrections. But if you’ve been in the business, it would take an hour-long meeting to sit with a graphic designer and tell them what you want. Wait a few days. Then get a result to iterate on. Still better value than an entry-level designer at R500 (~$31) an hour. Now you get a production asset, or an 80%-ready one that you hand to a high-paid graphic designer to polish. Logos, marketing assets, memes. Because in some businesses, a meme turns your brand into a verb faster than a marketing team with guardrails.
Then you point it at editing and writing. But here’s the thing nobody markets: the more context you feed it, the sloppier it gets. I learned that the hard way, by feel, over a quarter.
Claude is the architect. True story. In a prototype shop I don’t need every app built with production-grade tooling for something I’m never going to sell. I’m just fixing smaller problems over time. So I let the sculptor run fast – five times the thresholds – on the smaller functions, the building blocks. And I let the architect think about the two-in-ten products I actually need to productionalise.
Claude is the costly architect that thinks it’s going to upsell me. But by week one I’m already refactoring its output into reusable pieces. Legos. Sure, I can burn through a usage limit in thirty minutes flat with one to three prompts. And here’s the part the efficiency obsessives miss: I don’t care – even if I have to wait nearly five hours to get the rhythm flowing again on this cheapo plan. The rhythm and latency I needed for a flow state was in the creative. The premium to refactor after building isn’t what it used to be in the old days, when we did this with humans.
In business speak – that five-hour wait is time spent networking my product with marketing, finance, legal, sales, maybe an engineering team, to get it scalable and prod-ready. Some burn through massive budgets; I’m looking at how I extract maximum value with my time.
A lack of efficiency, in the right hands, is perfect. It buys me the time to realise I can’t actually market the app I built three weeks ago, because I’m figuring out pricing and marketing now as a one-man team – which is a finding, not a failure. So I redirect the credits I can’t spend on product and developer tasks into writing better content, building marketing collateral, building playbooks to scale a team, and realising I don’t want to sell busy products – those need teams. And that’s where AI is going: what used to take teams now takes one person.
The economics shifted
Here’s the prediction the experiment pointed me to. The build is becoming the cheap part. What stays valuable is the judgment around the build – knowing what to build, where the AI genuinely helps, and where it quietly makes things worse.
Watch what’s happening to the people who only do the build. The disruption is concentrated almost entirely at the entry level – the routine analytical layer that AI now eats – while strategic advisory, relationship management, and oversight keep growing. Harvard Business School’s field experiment with BCG, run across 758 consultants, found AI users finished 12.2% more tasks, 25.1% faster, with 40%+ higher quality – but the gains went to the people steering, not the work itself. The same study’s quiet finding matters more: on the tasks that sat outside what the AI was actually good at, those elite consultants did worse with it than without. Knowing where that line falls is the job.
Same pattern in code. A senior software engineer in Johannesburg averages somewhere around R85–90k (~$5.2k–$5.5k) a month, and the most senior architect roles run into seven figures a year (north of $60k). A lean senior team clears R150k (~$9.2k) a month comfortably.
So ask the question every business will eventually ask: why bring in a vendor for long-term technical building when the math no longer supports it?
The old model was: outsource a dev team for a 3-to-24-month block, eat the agency margin, deal with the admin of standing them up and tearing them down when the economics turn. The new math points somewhere smaller and sturdier: a small in-house team of two to four people, servicing a whole stack of products, highly efficiently.
The trick is to build in fat by design. Not bloat – slack, on purpose. Enough redundancy that the team keeps running when one person leaves. Enough that someone can take real paid time off – without a phone buzzing mid-holiday because nobody else can hold the thread. You staff for resilience, not for raw throughput. You make people valued enough to stay, and free enough to leave without disruption. Both at once. That’s the culture, and it’s the whole game.
Each person runs a sustainable rhythm under it – call it six months of real output, a month of genuine time off, and five months of maintenance and leftover slack: time they’re actively encouraged to use, improving things or spinning up mini R&D silos to experiment and build their own projects on the side – sharpening the high-performance skills you’ll cash in next cycle.
This isn’t really about dev teams, or apps. It’s about the questions a business has to ask itself now that the economics of building have changed underneath it.
Build, buy, or outsource – and the bias baked into the answer
A caveat, because everything above could read like always build. It shouldn’t.
My early take: anyone advising you to build, buy, or outsource who doesn’t understand both the options and your business deeply enough isn’t giving you an answer – they’re selling you their bias. The SaaS rep sells SaaS. The agency sells the build. Everyone’s incentive is to make their thing your answer. The honest version doesn’t have one answer, because it depends entirely on who you are.
Small-to-medium, and technology isn’t your strong suit? Buy the SaaS. The platform leaders have already outsourced the hard, boring, existential stuff – security, data protection, resilience, uptime – into a package you never have to think about. That isn’t a compromise. That’s someone else carrying risk you have no business carrying. Don’t build what you can rent for less than the cost of getting it wrong.
A tech upstart that already has the skills? An enterprise? Different question. Now it’s: at what point does scaling on someone else’s platform start to burn you? And there’s no blanket answer – it’s per tool in the stack. It turns on which skills you actually want to keep in-house, and how much growth you’re forecasting through each tool. The CRM you’ll never outgrow stays bought. The thing quietly becoming your core differentiator, the one you’re scaling tenfold – maybe that’s the one you pull in-house before the platform tax eats your margin.
And it’s worse than biased salespeople now, because the evidence you’d use to check them is gamed too. Search “should I build or buy” and you’ll drown in articles that look like research and aren’t. Most are SEO bait – written to rank, not to be right. A growing slice is AI-generated, paraphrasing each other in a circle until a statistic nobody can trace back to a real study gets repeated enough to feel true. A vendor blog dressed up as a “report.” A percentage three hops from wherever it was invented, with no primary source at the bottom of the chain. The same machine that writes the marketing now writes the “independent analysis” that backs it.
Most people can’t see the seams – and that’s the entire business model. Content engineered for the reader who takes a confident paragraph and a percentage sign at face value. I built this piece the slow way: the sources at the bottom took real digging, because what surfaced first was an SEO farm citing an SEO farm. A couple of the numbers I started with got cut outright – I couldn’t trace them past a single blog, so out they went. That’s not diligence theatre. Telling a randomised controlled trial from a press release with a chart is the skill. Most of the noise is designed so you never learn to.
So perhaps it was never about what someone’s selling you. It’s whether you’ve built the capability to filter the noise fast – to tell bias from advice. Building that filter is the actual skill, and it doesn’t ship in a package.
The real asset is the human context window
Here’s the line the whole thing hangs on: the real value is what a human context window holds over time – and how you structure a team’s culture and perks around protecting it.
I borrowed the phrase from the machines on purpose. A model’s context window resets every session. A person’s doesn’t. A person accumulates. The institutional knowledge, the client relationships, the why we don’t do it that way – that’s the asset that compounds, and it’s the one every org systematically under-protects.
The numbers on losing it are ugly. Widely cited HR research puts the cost of replacing an employee at 50% to 200% of their annual salary – higher still for senior and specialised roles – and most of that isn’t the recruiter fee. It’s the silent drain of knowledge walking out the door: the institutional memory, the half-documented why-we-don’t-do-it-that-way, the relationships, gone. Whole organisations quietly burn time and money every week because the things that matter live only in people’s heads – undocumented, unprotected, one resignation away from gone.
Which is the whole case for fat by design. If a single departure can cost you up to double a salary in vanished knowledge, then a little deliberate redundancy stops looking like waste and starts looking like an insurance premium you’re already paying anyway – just badly, after the fact, in chaos.
So the design problem isn’t “how do I get the work done cheaply.” The sculptor does that now. The design problem is: how do I build a team where the human context window deepens instead of leaking? That’s a culture question and a perk-package question – exactly the kind AI can’t answer for you.
Where this leaves me
Stripped to one line: the economics of building software have changed, and they now point in-house – toward small, resilient teams that compound context instead of renting it by the month. I think that’s the direction of travel. I don’t think it’s finished, and I don’t think I’m fully right yet.
Sources
A note on sourcing: where the story abstracts away the precise study, the credible underlying research is listed here.
- AI and developer productivity – METR, Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity (2025 randomised controlled trial), plus large-scale engineering-delivery research (Google’s DORA program; the Stack Overflow Developer Survey) showing adoption rising while delivery throughput stays flat.
- AI and knowledge work – Harvard Business School & Boston Consulting Group, Navigating the Jagged Technological Frontier (Dell’Acqua et al., 2023) – where AI lifts knowledge work, and where it quietly makes it worse.
- Cost of turnover and lost institutional knowledge – SHRM and Gallup (employee-replacement cost ranges); the Center for American Progress (higher costs for senior and specialised roles).
- South African engineering salaries – OfferZen, State of the Software Developer Nation; Glassdoor; PayScale.
Figures reflect ranges, not guarantees. The opinions on specific tools are mine and lived, not sponsored.