Source: https://jubileejoyzirebwa.com/notes/the-expensive-part-of-an-idea/
Canonical: https://jjzirebwa.substack.com/p/the-expensive-part-of-an-idea

[← All notes](https://jubileejoyzirebwa.com/notes/)

# The Expensive Part of an Idea

Who gets to be ambitious when the first conversation with expertise becomes cheap?

[JJ Zirebwa](https://jubileejoyzirebwa.com/about/)5 October 2026 9 min read

19 November 2026

Written in October 2026, before I'd named the parts. What I call NEXUS here is now [Entrust](https://jubileejoyzirebwa.com/entrust/): the first business NEXUS has built, and the platform it runs on. NEXUS is the institution behind both.

![Charles Babbage’s 1840 plan for an Analytical Engine](https://jubileejoyzirebwa.com/images/notes/the-expensive-part-of-an-idea/babbage-analytical-engine-plan.webp)

*Charles Babbage’s 1840 plan for an Analytical Engine. A machine for calculation is a very different thing from a social system for deciding what to calculate, who gets trusted, and what happens next. [Image source](https://commons.wikimedia.org/wiki/File:Babbage_Analytical_Engine_Plan_1840_CHM.agr.jpg)*

At 11.47pm, a blank chat window can make a person briefly dangerous to their own sleep.

You type in a thought that’s been annoying you for months, or just entered your head. Perhaps a problem at work? Perhaps an idea you have been too embarrassed to say out loud because it sounds vague, overambitious or like it needs someone in a suit to approve for the right to simply THINK about it. Ten minutes later you have a market map, a policy summary, some competitors, a rough technical architecture and a list of questions that make the original thought look like it showed up wearing one shoe.

Something important has happened there. The first conversation with expertise has become much cheaper. Probabilistic/mathematical *insert more AI and machine learning jargon*-mediated expertise, but expertise regardless.

This has been described as a productivity story, which is true in the same way that a library is a building story. There is a bigger change in who gets to entertain an ambition long enough for it to become embarrassing, specific and possibly even beneficial.

For a long time, ambitious inquiry came with entrance fees. You needed access to books, a university, a supervisor, a patient friend, a professional vocabulary, spare evenings, some money or super confidence that lets a person walk into an unfamiliar room and ask a question with their whole chest. Plenty of talented people had one or two of these. Fewer had enough of them at once.

Now somebody with a serious curiosity can interrogate a field in a language that makes sense to them. A nurse can ask what a proposed workflow would mean on a bad Tuesday. A scientist can trace a commercial question without pretending that a spreadsheet is beneath them. A graduate can find the broad shape of a regulatory pathway before meeting the person who will eventually explain why the broad shape is where the trouble starts.

That is a sign to an opening. Permission hasn’t been granted yet.

The [ILO’s 2025 update](https://www.ilo.org/publications/generative-ai-and-jobs-refined-global-index-occupational-exposure) estimates that one in four workers globally sits in an occupation with some exposure to generative AI. The report expects transformation to be the dominant effect for most exposed work. Do not underestimate the uncertainty packed in that sentence. Exposure is uneven across countries and occupations. Clerical work is still highly exposed, while some professional and technical work has moved further into range. A single chatbot interface does not produce any sort of labour market.

Even the way people use these systems is uneven. [Anthropic’s June 2026 Economic Index](https://www.anthropic.com/research/economic-index-june-2026-report) finds differences in reported and expected capability across geography, experience and the degree to which users hand work over or work with the model. This might just be platform evidence, but I find it to be a sobering antidote to the fantasy that a tool shows up and distributes its benefits evenly.

But let’s take a step back, there is an interesting result in the data. In a customer-support setting, a generative-AI assistant raised productivity by about 14% on average. The gains were much larger for newer and lower-skilled workers. [Brynjolfsson, Li, and Raymond’s study](https://www.nber.org/papers/w31161) is narrow, as it should be. It cannot tell us what will happen to every profession, wage or graduate scheme. It does suggest something strange and hopeful.

The system had absorbed fragments of what experienced agents had learned across many previous calls. A newer worker could borrow pieces of that accumulated practice. The model did not turn them into a veteran overnight. It did make some of the veteran’s hard-won pattern recognition available before the typical years of learning by getting it wrong in public.

That complicates all the excitement. We may be able to make the beginning of expertise conversational. We still need a way through everything after that beginning.

I have spent enough time around health innovation to know that a persuasive slide deck can lose a fight with evidence, procurement, governance or a clinician’s entirely understandable suspicion. In health, a plausible analysis is the introduction. Somebody still has to understand the data well enough to be contradicted by it. Somebody still has to explain the intervention to people whose workday it may complicate. Someone else still has to carry the responsibility when a neat diagram meets an actual patient pathway and produces seven additional clicks for a nurse who was already busy.

This is not a complaint about administration. I used to find it faintly offensive that the world required so much co-ordination after an idea seemed so obvious to me. Then I spent more time around systems that could injure, waste money or exhaust people and the administrative middle became the point. Reality is allowed to be difficult because it is often doing a job.

The expensive part of an idea is therefore not always the idea. It is the period in which the idea becomes answerable to other people.

Oddly, a historical concept I kept gravitating towards was apprenticeship. I had filed it in the same mental drawer as fish n chips, afternoon tea and other random things I found in England that don’t seem to concern me. Patrick Wallis’s study of premodern England makes it harder to look at that casually. Apprenticeship was a major route into non-agricultural work. Many seventeenth-century London arrangements ended before their formal term. The contract had to distribute the cost of training and the risk that the relationship, the work or the person would prove a bad fit. [Wallis’s account](https://ideas.repec.org/a/cup/jechis/v68y2008i03p832-861_00.html) revealed an arrangement for learning through work and for sharing an uncertain bet on what that learning could become.

The old system deserves no nostalgic restoration. It carried hierarchy, exclusion and the familiar trick of calling someone else’s limited options an educational opportunity. Still, it held one important idea. Developing capability was treated as a social and economic process. Somebody invested time. Someone made room. Someone tolerated the early bad work because the later work could negate it. The relationship could also end. That was important.

Modern professional life has some versions of this. Modernly popularised apprenticeships themselves, graduate schemes, good managers, funded doctoral work, patient employers, family support and the occasional colleague who will explain the unspoken rule without making you feel like a small administrative error. They are unevenly distributed. They have also become strangely hard to name when we talk about innovation. We praise the founder who “took the leap”, as if the leap occurred in a vacuum and not somewhere between savings, social permission, immigration status, contacts, a forgiving partner and an uncle willing to pretend he understands seed rounds.

AI could make another route possible. An experienced professional may have a real domain question, a career they do not want to abandon and enough money to fund a disciplined inquiry. A developing operator may have the time, appetite and intelligence to do crucial research and co-ordinate difficult work, while needing exposure to judgement that a chatbot cannot grant. A specialist may enter when the work reaches a point where over-confidence in that operator would simply be dangerous.

There is no requirement for this to culminate in a company. It may create one. It may improve something inside an existing organisation. It may produce a clear and well-earned decision to stop. I find the last outcome oddly attractive. We have made quitting look brave and stopping look sad. In many cases, stopping after a proper investigation is an act of capital preservation and self-respect. It is also how you clear space for the next question.

This is a speculative institutional idea I intend to give a real shot through [NEXUS](https://jubileejoyzirebwa.com/nexus/). Its difficult questions are the whole point. Who pays? Who supervises? Who gets credit? What work is safe for a junior operator to own? When is the right answer to bring in somebody more experienced? Who takes responsibility when an AI-assisted piece of research looks convincing and is wrong?

The strongest objection arrives quickly. Apprenticeship can become a polite name for cheap labour and AI can make the disguise more efficient. A senior professional buys an enthusiastic graduate’s time, asks them to produce ten brilliant documents a week, calls it development and then tells a panel about creating opportunity. Nobody needs that institution. We have enough of it already.

Any modern version worth building would need pay, serious supervision, clear attribution, a route out and exposure to decisions whose consequences a person can understand. It would need to allow the operator to become more capable. Busyness alone is a poor substitute. It would need an explicit point at which the work goes to a scientist, lawyer, clinician, engineer or other specialist whose judgement cannot be simulated by enthusiasm and a long prompt.

[Acemoglu and Restrepo’s account of automation and new tasks](https://www.aeaweb.org/articles?id=10.1257/jep.33.2.3) gives the argument its necessary cold shower. Technology can displace work. New tasks do not appear because history enjoys a balanced ending. Firms, institutions, incentives and demand create them. The same is true of routes into ambitious work. A model can generate a logo, a business plan and a mildly upsetting amount of enthusiasm. It cannot decide that a person deserves a paid chance to learn in public.

The UK government’s [AI Opportunities Action Plan response](https://www.gov.uk/government/publications/ai-opportunities-action-plan-government-response) speaks confidently about adoption, productivity, growth and improved services. I want all of those things. The geographical awkwardness remains. Someone in Cambridge, Harare, or a village with intermittent connectivity may receive a similarly fluent answer from a model. Their access to a customer, a funder, a laboratory, a mentor, a reliable payment system or a room where they can fail without catastrophe remains very different. The conversation has travelled further than the institution.

At first, I thought the story here was that AI may democratise entrepreneurship. I still think it can democratise beginnings. What we perhaps need to build is the middle. A paid, supervised, accountable period between “I have a thought” and “I have to bet my whole life on it.”

That middle might be small. It might begin with one professional, one operator, one real problem and a very honest budget. It could become a new kind of apprenticeship if it learns the right lessons from the old one and refuses the ugly ones. It could also become nothing, which would be fine if the people involved leave with more judgement and no one has been used as scenery for another person’s ambition.

The machine will keep making first conversations cheaper. It may make them absurdly cheap. Good. I want the next conversation to be worth something too. The one in which someone says, “Here is a real problem. Come and help me find out whether you can carry it.”

That is the expensive part.

## Sources and notes

- ILO, *[Generative AI and Jobs: A Refined Global Index of Occupational Exposure](https://www.ilo.org/publications/generative-ai-and-jobs-refined-global-index-occupational-exposure)*, 2025
- Brynjolfsson, Li and Raymond, *[Generative AI at Work](https://www.nber.org/papers/w31161)*
- Wallis, *[Apprenticeship and Training in Premodern England](https://ideas.repec.org/a/cup/jechis/v68y2008i03p832-861_00.html)*, 2008
- Acemoglu and Restrepo, *[Automation and New Tasks](https://www.aeaweb.org/articles?id=10.1257/jep.33.2.3)*, 2019
- [Anthropic Economic Index, June 2026](https://www.anthropic.com/research/economic-index-june-2026-report)
- UK Government, *[AI Opportunities Action Plan: government response](https://www.gov.uk/government/publications/ai-opportunities-action-plan-government-response)*, 2025

I'm JJ. I studied biomedical science, then spent a year working out which health technologies were worth backing.

Follow [RSS feed](https://jubileejoyzirebwa.com/rss.xml) [LinkedIn](https://www.linkedin.com/in/jubileejoy-zirebwa)

[![Substack cover for The Expensive Part of an Idea, over Charles Babbage’s Analytical Engine plan.](https://jubileejoyzirebwa.com/images/notes/the-expensive-part-of-an-idea/substack-cover-640.webp)Originally published on Substack ↗ 7 October 2026](https://jjzirebwa.substack.com/p/the-expensive-part-of-an-idea)

[Image source](https://commons.wikimedia.org/wiki/File:Babbage_Analytical_Engine_Plan_1840_CHM.agr.jpg) · [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) · Photograph by ArnoldReinhold. Resized and compressed.

## Mentioned in

- [The Omnitrix Problem](https://jubileejoyzirebwa.com/notes/the-omnitrix-problem/)
- [Work with me](https://jubileejoyzirebwa.com/hire/)

Related work · [Health Innovation East](https://jubileejoyzirebwa.com/work/#health-innovation-east)
