Stop Deploying AI Tools. Start Helping People Build Their Own.

VisiCalc was not built by IBM. It was built by Dan Bricklin because he needed a better way to do his Harvard MBA homework. He was sitting in a lecture, watching a professor erase and recalculate an entire spreadsheet on a blackboard, and thought: my computer could do this. So he built the thing himself. That piece of software became the reason businesses bought personal computers in the first place – not because IT departments decided to deploy spreadsheets, but because one person built a tool that matched how he needed to think.

We are at exactly the same inflection point with AI. And most organisations are getting it wrong.

The question everyone is asking

Every enterprise I talk to is asking the same question: “Which AI tool should we deploy?”

They evaluate vendors. Run pilots. Write procurement documents. Deploy Copilot or Gemini or Claude to everyone, identically. Then they wonder why adoption stalls.

The MIT NANDA report found that the vast majority of generative AI pilots produce no measurable P&L impact.¹ Deloitte’s 2026 State of AI in Enterprise puts numbers around the gap: 97% of executives say they benefit from AI, but only 29% see significant organisational ROI.² And S&P Global reported that 42% of companies abandoned AI projects entirely in 2025, up from 17% the year before.³

The tools work. ChatGPT, Claude, Copilot – they demonstrably produce value for individuals who use them well. The failure is not in the technology. It is in the deployment model.

What if the problem is not the tool? What if the problem is the question?

How people work (which is not identically)

Every standardised tool deployment makes an implicit assumption: that people work the same way. They do not.

Cognitive diversity research shows that people differ along multiple dimensions simultaneously – visual versus verbal processing, structured versus exploratory problem-solving, deep focus versus context-switching, sequential versus parallel thinking. These are continuous spectrums, not categories, and every individual sits at a unique intersection of all of them.⁴ Cognitively diverse teams outperform homogeneous ones, enhancing innovation by 20% and reducing risk identification failures by up to 30%.⁵

But here is the tension: cognitive diversity only produces those benefits when people can work in ways that match their cognitive patterns. When tools enforce uniformity, the diversity premium disappears into disengagement.⁶

Organisations want the benefits of cognitive diversity while deploying tools that enforce cognitive uniformity. That is the contradiction at the heart of the 95% failure rate.

The neurodivergent signal

The strongest evidence for personalised AI comes from neurodivergent workers, who represent an amplified version of what is true for everyone: standardised tools often do not match how your brain works.

A UK Department for Business and Trade study found that neurodivergent workers were 25% more satisfied with AI assistants than neurotypical respondents, and more likely to recommend the tools to colleagues.⁷ People with ADHD, autism, and dyslexia describe AI as a bridge – not doing their thinking for them, but translating between how they think and how workplaces expect them to communicate.⁸

An EY survey of Microsoft Copilot users found that 88% of neurodivergent respondents reported higher productivity with AI assistants.⁹ It is worth noting that this was a self-selected sample from an existing Copilot user base in a Microsoft-partnered study – so the figure likely overstates the general case. But the qualitative evidence is consistent and compelling across multiple independent studies: AI that adapts to how someone thinks produces dramatically better outcomes than AI that expects everyone to adapt to it.

And this is the key insight: what is true for neurodivergent workers in concentrated form is true for all of us in diluted form. We all have cognitive preferences. We all work differently. We all have gaps between how we think and what standard tools expect.

Tools shape thinking

There is a line often attributed to Marshall McLuhan – “We shape our tools and thereafter our tools shape us” – though it was written by his colleague John Culkin in a 1967 article about McLuhan’s work.¹⁰ The distinction matters less than the insight, which takes on new weight in the age of AI. A December 2025 piece in Psychology Today argues that AI is the first medium that joins our thinking rather than shaping it from outside – “a participant in thought itself.”¹¹

Follow the logic. If a standardised AI tool shapes everyone’s thinking in the same way, you get standardised thinking. You lose the cognitive diversity that produces the innovation premium. If a personalised AI system reflects and reinforces an individual’s cognitive patterns, you get amplified diversity. Each person’s unique way of thinking becomes more powerful, not more homogeneous.

When I built my own AI team – a coordinator agent that dispatches specialist agents for research, writing, knowledge management, infrastructure, and quality review – I was not just building a productivity tool. I was building an externalisation of how my brain works. The dispatch model reflects how I delegate. The specialist agents reflect how I categorise expertise. The daily digest reflects my need for structured morning orientation. The knowledge base reflects how I connect ideas across domains.

Another person would build a completely different system. And that is the entire point.

Why personal systems work: the psychology

There are three psychological mechanisms that explain why building your own AI system works better than being handed one.

The IKEA effect. Norton, Mochon, and Ariely demonstrated in their 2012 Harvard Business School study that people who invest labour in building something value it significantly more than an identical pre-built product.¹² A tool you built is part of you. A tool you were given is just a tool. But the research includes a crucial boundary condition: labour leads to love only when it leads to successful completion. If someone tries to build and fails, the effect reverses. This means we cannot just hand people raw components and say “go.” We need scaffolding – starter templates, clear progression, communities that help people succeed.

Self-Determination Theory. Edward Deci and Richard Ryan identified three innate psychological needs that drive human motivation: autonomy, competence, and relatedness.¹³ Building your own AI system satisfies all three. You choose the agents, define their roles, and decide what they work on – that is autonomy. You understand how the system works because you built it, and you can extend it – that is competence. You share what you have built, learn from what others built, and belong to a community of builders rather than a user base of consumers – that is relatedness.

Being assigned a corporate AI tool, by contrast, satisfies none of them. You did not choose it. You do not understand its internals. And “user adoption” is not the same as belonging.

The endowment effect. Objects viewed as part of the “extended self” are valued more highly, and this effect increases with duration of ownership.¹⁴ A personal AI system that grows with you over months becomes genuinely yours in a way that a corporate deployment never does. Every configuration decision, every agent definition, every knowledge base entry makes the system more valuable to you – and more costly to abandon.

The personal computing parallel

We have seen this pattern before. In the 1970s, computing was centralised. Mainframes served entire organisations. Users submitted jobs and waited for results.

Then the personal computer arrived. And VisiCalc was not the only example. The explosion of personal computing happened not because everyone used the same machine the same way, but because individuals could configure their own tools, write their own software, and build solutions for their specific problems.¹⁵

The companies that thrived in the PC era were not the ones that standardised everyone on one machine. They were the ones that created environments where people could build their own solutions.

AI is making the same transition right now. The question is whether organisations will recognise it in time, or whether they will spend another five years trying to find the one tool that works for everyone – the AI equivalent of insisting that everyone use the mainframe.

The practical path

This is where theory meets reality. Not everyone starts by building a fourteen-agent team. The progression looks like this:

1. Start with chat. Use Claude or ChatGPT as a conversation partner. Build intuition for what AI can do and where it falls short. This is the equivalent of turning on your first PC and opening a word processor.

2. Add skills. Claude Skills – reusable instruction packages – let you teach the AI your preferences, context, and workflows without writing a line of code. This is the most accessible entry point for personalisation.¹⁶

3. Build a CLAUDE.md. This is a written representation of how you work – your context, your preferences, your role, your thinking patterns. It is the foundation document that makes every AI interaction more useful because the system understands who you are.

4. Create your first specialist agent. A researcher. A writer. An analyst. One agent that does one thing well, configured to match how you think about that domain.

5. Build the system. Multiple agents, a coordinator, a knowledge base, daily rhythms, quality gates. This is where you arrive at your own AI team – a system that compounds in value every day because it learns your context and reflects your thinking.

The pattern from AI-assisted development is instructive: week one uses community templates, weeks two through four involve observing what you change repeatedly, and by week five you are creating custom templates for your own patterns.¹⁷ The system is never “done” – you keep refining it as your work and thinking evolve.

What about governance?

This is the question I expect from every enterprise leader reading this, and it is a fair one. If everyone builds their own AI system, how do you ensure data security, compliance, and quality?

The answer is: govern the framework, not the implementations.

Set clear boundaries around data handling, security requirements, and compliance standards. Define what agents can and cannot access. Establish quality gates and review patterns. Then let people build freely within those boundaries.

This is not a new model. It is how every successful platform works. AWS does not tell customers what to build – it provides guardrails, compliance frameworks, and shared responsibility models, then lets people build what they need. The same principle applies to personal AI systems within an organisation.

Deloitte’s 2026 research found that leading firms “build and share reusable GPTs and API-powered assistants” – with BBVA regularly using more than 4,000 custom GPTs.² That is not anarchy. That is a governed framework with four thousand personalised implementations.

A note on access

I want to be honest about something. Building a personal AI system requires access to the tools, time to experiment, and a degree of comfort with technology that not everyone has today. The digital divide is real, and it does not disappear just because the entry points are getting lower.

This is precisely why the scaffolding matters – starter templates for different working styles, getting-started guides that produce a working system in one session, and communities of practice where people help each other. The goal is not to create another technology that benefits the already-advantaged. It is to make the building accessible enough that anyone who wants to can start.

The progression from chat to skills to CLAUDE.md is deliberately designed with low barriers. You do not need to write code. You do not need a computer science degree. You need curiosity about how you work, and a willingness to experiment.

The question that matters

Every enterprise is asking “which AI tool should we deploy?” The better questions are:

• How do we help each person build their own AI system?
• How do we teach people to build and evolve their own AI infrastructure?
• How do we create an environment where personal AI systems thrive within governed boundaries?
• What is the compounded value of five hundred people, each with a system tuned to their thinking?

Deploying a tool is a procurement decision any competitor can replicate in ninety days. Building a culture of personal AI systems is a methodology, a community, accumulated knowledge, and a philosophy. It takes time to build. It compounds. And it cannot be purchased from a vendor.

Kim Carson, writing for the Long Now Foundation, put it well: “Purpose is not a download. It is a discovery.”¹⁸ I think building your own AI system is one of the most practical forms that discovery can take in the world we are heading into. Not being told what tool to use, but building the system that reflects what you uniquely bring to the work.

Stop asking which tool to deploy. Start helping people build their own.