It’s been an interesting two months.
On 1 September, I set up Knowledge Tree & Butterfly (KT&B), my consultancy aimed at helping analytical knowledge workers make much better use of agentic AI in their work.
And now, on 1 November, KT&B is being ‘shuttered’, frozen, mothballed. Why? Because I’ve been offered a very generous and exciting position as Head of Data at Elevate Great, a research charity focused on addressing social inequalities and inequities through a collaborative, evidence-focused lens.
The Head of Data role is full-time. Subject to written agreement, I’d potentially be willing and able to work with people and organisations with my Butterfly-founder’s hat on. But it’s not something I’ll be pushing or promoting for the foreseeable future. There’s more than enough to do with the Data hat on instead.
Selling the founder
The two months helming KT&B has been exciting, risky,1 and rewarding in every way imaginable, except financially. For a one person business the main thing that needs to be sold is the founder, and what it is that’s distinct and valuable that they can offer to potential clients. I’m a highly skilled and adaptable knowledge worker, highly adept at quantitative analysis and statistical inference, able to translate between verbal and informal descriptions of data and problems, and algebraic formalised representations of the same within a robust framework of statistical inference. If my business were focused on selling my abilities as statistician, health economist and data scientist in the way these trades have been plied for the last couple of decades, I would plausibly have been more conventionally successful at getting work commissioned over KT&B’s period of operation.
But the business was founded on the expectation that these roles, and other analytical knowledge work roles, and other knowledge work roles, will stop being tenable over the coming months and years. Most of the expertise and judgement developed in working such roles will be as valuable as it ever was, if not more. But how this expertise and judgement gets developed and expressed, and what additional forms of expertise analytical knowledge needs to get leavened with, will fundamentally change.
From two doors to supergroups
I’ve pivoted and pivoted and pivoted my framing, leading to many changes over the last eight weeks in my website and landing page. Originally the primary framing was what became known as ‘The Two Doors’: the idea that there are i) analytical folk who are not technical enough (my main door/target audience) and ii) technical folk who are not analytical enough (my secondary door). My modest proposal: adding technique to analysis, or analysis to technique, will supercharge both camps. But this framing suffered two fatal flaws: Firstly, that few people recognise themselves in the categories the framing depends on: no one’s employed as an ‘analytical knowledge worker’, even if they meet my definition. Secondly, the framing emphasises and starts with an assertion of deficit that few skilled and capable people would be willing to assume of themselves: that for someone who’s a trained statistician or health economist (say), and/or has chased accreditation in their highly regulated and professionalised field (such as the actuarial sciences) and has years of qualifications and CPD, there’s still something big and relevant that they’re missing. (An understanding of git-based version control, for example.) So, the initial framing was both opaque and accusatory: not an effective hook for immediate engagement leading to potential business.
A better framing: you might not yet have the full package, traversing analytical and technical and substantive knowledge, but if you can work together, in a psychologically safe environment with others whose expertise complement your own, then you can now achieve much more together than as isolated, siloed individuals. I called this framing Scientific Supergroups.2 This framing, which came from Richmond Davies, flips the deficit accusation implicit in the Two Doors into one of strengths. This pivot, towards empowerment and collaboration, has been a throughline in most of what I’ve produced since. It is, on reflection, a much better framing.
Acute and chronic pain
As well as a steady number of conversations with potential clients and advisors who have all been generous with their time and expertise, I was also accepted into a Techscaler Catalyst programme for new founders. From here I learned the importance of pain in self promotion. A business needs to offer solutions, solutions to problems that potential customers readily and immediately identify themselves as experiencing. A pain, in this sense, is a problem that customers have recognised, and that remains salient and foregrounded in their attentions and concerns.
The key problem for me: salient and foregrounded brings attention to acute pain, not chronic pain, and fundamentally what KT&B has been about offering are salves to chronic pain, not acute. A pain that never resolves itself, a continual hum or throb, loses its salience, stops being a signal to act, and becomes simply a grumble or discomfort. When pain becomes chronic, the call to action stops being heard; the life (or worklife) becomes impaired and degraded, but the sufferer shoulders this burden with fatalistic stoicism, a mustn’t grumble mentality, learned helplessness. The chronic pain sufferer has internalised a lesson in futility: that the chronic pain must just be lived with - that (say) analytical reports that historically took four months to produce and another two to fix errors in - will always be with us, and that trying to reimagine different ways of doing the work - in which these timelines could be vastly compressed and errors caught much earlier - will only lead to distraction and disappointment. So, another way my messaging wasn’t landing was in the scale of its audacity: new ways of doing analytical knowledge work which, in effect, mean the chronic pain of the old ways become choices rather than necessities.
What actually worked
Further supporting excess audacity (mismatch between what people believe is possible and what’s now possible) as a primary cause of lack of engagement is the one thing that showed itself, empirically, to actually work: offering myself as something like an informal ‘Anthropic Trainer’, showing how to get (much) more from Claude Pro subscriptions that the organisations were already willing to pay for and permit to work. In the two case studies I managed to initiate and complete in my two months as founder-director, simply showing how fairly simple changes in the way one interacts with Claude can lead to massive changes in productivity and effectiveness, was all it took for the scales to fall, even if in both cases the use-cases related more to administrative activities required for roles more generally, rather than to ‘analytical knowledge work’ more specifically.
Some top tips
And as a reward, for those who’ve read this far, here are some top tips for doing this:
- Pay for at least a Claude Pro subscription, being willing to pay for a Claude Max x5 subscription (now simply the Max subscription) if you find yourself stopped frequently (as in at least weekly) during the workday by 5 hour usage limits.
- Start with a highly capable model - currently Opus 5.5 - and then decide whether something a bit cheaper might be able to do the job too. You need the capability at the start, because the problems are most complicated and poorly defined at the start.
- Share the tooling and files necessary to do the job with the models you expect to do the job. They can only read the report if you’ve given it them, can only find the report if you let them search all the folders they might be, can only perform the calculations in Excel if they’re allowed to use Excel, can only find and write emails if they’re allowed to use your email.
- Think about your work in terms of projects demarcating types of work activity, and create folders on your computer for each of these projects. Give the model permissions to read and write within these project folders, as well as to search more widely where necessary.
- This includes allowing the models to build, query and maintain project specific knowledge graphs (collections of interlinked plain text files), skills (clearly defined prompt-and-context bundles, also plain text files), and scripts (computer programs written by the AI to be used by later versions of itself). This is how the model appears to be able to ‘learn’ and keep improving.
- Use version control through git or similar. This both provides insurance against breaking changes in a project becoming fatal, and makes it much easier (through git commit message logs) for new instances of a model to understand how and why a project has developed the way it has.
- Use a secure and appropriate organisational github or similar service, both for having a second place where the projects can live, and for allowing multiple people to use and develop a project, with AI, without stepping on each other’s toes.
- Remember that LLMs live in streamworld, whereas we live in senseworld. To get a glimpse into streamworld, consider how an extremely knowledgeable and fast elite programmer, who’s also very physically disabled and generally prefers only to read and write into a command line terminal using a keyboard alone, might achieve a task or series of tasks. With these constraints some tasks we consider easy instead become hard and vice-versa, but generally, with enough support and understanding, this elite keyboard warrior will still be able to do the task, just in a different way to how you might do the same.
- Be willing to think differently, to learn, and to slow down when it comes to some of the technical concepts and tools that it’s good to learn when it comes to getting the best out of LLM-based AIs as agents. You’ll be more than rewarded via increased throughput in routine tasks that used to take hours of your day getting compressed into minutes.
Gone to seed
Anyway, though the two months in which I helmed KT&B had not yet led to a clearly viable and sustainable business model,3 it’s something I’ve learned from and enjoyed immensely, and will be massively useful in terms of what I can introduce to Elevate Great.
The Knowledge Tree has now gone to seed. In the best way possible!
Footnotes
By conventional assessments, but I’ve not been assessing the present epoch as at all conventional.↩︎
Though one thing I need to be mindful of is my own tendency towards idiosyncratic neologism; terms that actually resonate are those freely used and offered by others in conversation, not those I myself coined.↩︎
And these days I’m not sure if anything can be this medium term.↩︎