Episode 5: Future Tech Careers

Lee and James set out to talk about the future of tech careers, and what starts as a practical discussion quickly turns into something more uncomfortable. The skills that defined a generation of technologists are being automated in real time, the gap between AI adopters and everyone else is compounding fast, and neither of them can quite square their optimism about new opportunities with their concern about the people being left behind. James brings war stories from networking events and client work. Lee brings a growing unease about workforce displacement and a conviction that the career ladder most people climbed no longer leads anywhere recognisable.

Lee frames the central problem early. There’s an awkward curve to AI adoption: you get exposed to a large language model, experience a ten to fifteen percent productivity boost, and think it’s incredible. Then you hit a wall because the products built on top of those models haven’t matured fast enough. The models understand natural language almost perfectly - atrocious spelling, missing spaces, ambiguous phrasing - and still return something useful. But the experience varies wildly depending on which product you’re using. Claude Code feels like a different generation compared to tools that are still chaining prompts in a linear sequence. Lee now deliberately switches between products not to compare models but to compare how well each product lets him use the model. The value, he argues, has shifted from the intelligence layer to the experience layer. And the speed at which that experience layer is evolving - weeks and months rather than the years conventional software takes - makes it almost impossible for newcomers to keep up. He tried setting up Claude Code in VS Code for his kids and it took twenty-five minutes of subscriptions, authorisations, and permissions. That’s the barrier, and it’s not a trivial one.

James picks this up with stories from a business event he’s just returned from. Two camps of sceptics stood out. The first were using AI as a straight Google replacement - throwing broad, context-free questions at it and dismissing the results when they came back vague or incorrect. The second, a legal firm, had staff reviewing documents with ChatGPT and finding the output superficially adequate but missing the nuanced catches that years of experience would surface. James reframes both as user error, not AI failure. You wouldn’t hand a fresh graduate a document and expect partner-level analysis without a brief, and the same applies to a language model. Without a rule set, without domain specifics, without the right context, you get textbook answers. The free tier makes this worse because it runs on older models with smaller context windows, creating a first impression that undersells the technology dramatically. For the people already investing time and money, this is actually an advantage - the gap between adopters and sceptics keeps compounding.

The conversation sharpens when Lee raises the product-versus-agent distinction. He sees companies wrapping what should be a single focused agent in an entire SaaS platform - adding audio, video, OAuth, pricing tiers - not because users need the complexity but because you can’t charge fifty pounds a month for a well-configured agent. Everyone is land-grabbing, building features to justify subscription models when the actual value could be delivered by something much simpler. James concedes the point but argues the platform layer still matters for non-technical users. People understand websites with buttons. They understand SaaS. If it saves them hours for fifty quid a month, they genuinely do not care what’s under the hood. He draws a parallel from his contracting career: business owners never cared whether code was written in .NET or Python. They wanted the problem solved. The same principle applies to AI products - solve the problem simply and people will pay for it.

They land on something both find genuinely uncomfortable: the speed mismatch between what AI enables and what organisations can absorb. Lee describes working on multiple projects simultaneously and deliberately stretching a delivery to two months because he knows the client isn’t ready to receive it in a week. James has lived the same thing from the delivery side. Brought in as the person who rescues stalled projects, he’s watched his timelines compress from months to weeks to days, and now from days to hours. The problem isn’t building fast enough; it’s that clients question the bill when the work arrives too quickly. He tells a story about sitting with a client for thirty minutes doing logo work, iterating in real time until the client loved the result, and then being asked why they should pay the quoted rate for half an hour of effort. The answer was always that it took twenty years to develop the skill that made thirty minutes possible. But James acknowledges that defence has a shelf life - in two to three years, when anyone can prompt an agent to produce comparable work, the experience premium erodes and the proof-of-concept industry as a whole faces an existential compression.

The episode takes its sharpest turn around workforce displacement. Lee brings up Block, the payments company formerly known as Square, cutting roughly forty percent of its staff and seeing its stock price bump in response. The market rewarded the layoff. What troubles him isn’t that companies are adapting - it’s that the people being let go are being jettisoned, not repositioned. James has seen the better version of this, where companies he worked with retrained and redeployed staff into new roles when automation replaced their old ones. But he admits that was the exception. The productivity gains for people who embrace AI aren’t incremental - Lee puts them at five to six hundred percent - and the disparity between adopters and non-adopters is only going to grow. Lee pushes the idea of digital doppelgangers: sub-agents trained on an individual’s knowledge that go to work on their behalf. You’re not fired; your agent is. You take it home, retrain it, improve it, and send it back out. James likes the concept but flags the generational problem - pre-AI workers have skills to encode into these agents, but post-AI generations may never develop them independently.

They close with an exchange that’s equal parts honest and unsettled. James floats a half-serious theory that the AI revolution might be the thing that finally makes communism work in practice - less work, more quality of life for everyone. Lee points to the European model, where colleagues vanish from mid-July until September and the economy carries on. Neither pretends to have an answer to the work-life balance question that AI is forcing open. Lee notes that even as daily power users they still hit maddening limitations - deleted files, unexplained changes, tools that get the capital of France wrong. If the people at the forefront are still struggling, the bridge to mainstream adoption is longer than anyone wants to admit. They sign off promising next week’s episode will explain how to balance it all, then immediately concede they haven’t balanced anything in their own lives either.

Subscribe to The AI Shift for more honest conversations about AI and the real world.