People sometimes assume I came to artificial intelligence recently, drawn in by the same headlines that have everyone else asking questions. The truth is that AI and I have been in the same room, in one form or another, since before most of today's practitioners were born. I want to walk you through that journey here, not to impress you with dates, but because I think the shape of it says something useful about how these shifts actually play out, long before the rest of the world notices them.
Between 1980 and 1984, while working toward my Computer Science degree, I was already immersed in what AI looked like in its formative years, long before it resembled anything recognisable today. My coursework included AI Lisp, one of the primary languages of symbolic AI research at the time, and YACC, the parser generator central to how compilers were built and understood in that era, alongside a broader foundation in COBOL, C, Pascal and Assembler. Neural networks were barely part of the conversation then, dismissed by much of the field after Minsky and Papert's critique of the perceptron, so the AI I learned was built on logic, rules and symbolic manipulation rather than the statistical methods that would come to dominate decades later. That grounding in AI's first, harder era gave me an intuition for the discipline that few of my contemporaries carried into their careers.
The Branch of the Future
My first real world use case encounter with applied AI was in 1988, when Abbey National was converting from a building society into a public company and wanted a pilot branch that looked nothing like a bank. The idea was radical for its time: remove the security glass, remove the high counters, and let customers sit across an open desk from a member of staff the way they might sit across from a financial advisor today. Nobody had built a branch like that before, and the reason nobody had was simple. Once you remove the physical barriers, you remove the thing that made cash handling safe.
We (part of the core design team) solved that problem with engineering, routing physical cash through sealed under-counter robotics and a network of pneumatic tubes that moved money to a secured vault without it ever sitting exposed on the floor. But the part of that project that shaped the rest of my career was the software running underneath it. We built the advisory logic on Crystal, one of the first commercially viable expert system shells, and used it to give branch staff real-time, rules-based guidance on mortgages, savings products and eligibility, at the exact moment they were sitting with a customer. It worked because the system did not try to replace the person across the desk. It gave them better judgment to work with.
It was, in effect, an AI copilot for a bank teller, a decade before anyone used that phrase.
Teaching a Machine to Judge Creditworthiness
Not long after, I carried that same Crystal-based approach into British Credit Trust, a car loan subsidiary of Bank of Ireland, where the problem was different but structurally similar. Loan officers were making credit decisions under time pressure, with incomplete information, and with a real cost attached to getting it wrong in either direction. We built a rule-based scoring engine that could rate and approve a customer's creditworthiness close to instantly, encoding years of underwriting judgment into a system that could apply it consistently at scale. It was one of the earliest examples I know of AI making a real lending decision rather than simply flagging a file for a human to review later.
That distinction, between a system that assists a decision and a system that makes one, is a line I have thought about in every project since.
From Banking to the Streets
By the mid-1990s I had moved into a very different kind of intelligence work, helping UK and Irish police forces bring technology to bear on crime pattern analysis, associate networks and suspect identification. This was a world away from banking, and yet the underlying challenge was the same one I had been solving all along: enormous amounts of unstructured information, in this case free-text reports, voice recordings and video, that needed to be turned into something an investigator could actually act on. We integrated technologies including Autonomy and Memex alongside custom rules management systems, working across several projects in the UK and Ireland to bring genuine intelligence-led capability to forces that had, until then, relied almost entirely on manual cross-referencing.
That work led to my being invited to advise on the UK's National Strategy for Police Information Systems, and separately on Scotland's equivalent SPIS strategy, both in the mid-1990s. Advising two national strategies at once, on a technology that barely had a name yet, is the kind of thing that only happens when you are early enough to a field that the people setting policy still need practitioners to explain what is actually possible.
Tracing the Untraceable
That policing work eventually led to one of the projects I am proudest of: Ireland's TRACE platform, built in partnership with Esri, designed to help investigators trace missing persons and unsolved crimes through a combination of intelligence-based suspect matching, forensic evidence correlation and geospatial crime pattern analysis. TRACE took the geographic dot-on-a-map approach that most policing tools were still using and turned it into something closer to a living intelligence layer, capable of surfacing hotspots, linking suspects and associates across cases, and giving commanders a genuine tactical view of where resources needed to go next. It is still, decades later, one of the clearest examples I know of geospatial data and rules-based intelligence combining to solve a problem neither could solve alone.
Betting on Machines in the Market
By the late 2000s, my attention had shifted toward markets. Between 2009 and 2011, I built and ran automated Forex trading bots, systems designed to read currency market signals and execute trades without a human sitting at the screen making each call in real time. This was well before algorithmic trading was a mainstream retail conversation, and it taught me a lesson that has stayed with me ever since. Building and refining those bots through the volatility of that period was, in hindsight, the closest thing I had to a graduate education in what it actually takes to trust a machine with real financial risk, and how much discipline it takes to keep improving that trust rather than assuming it once it has been earned.
That education paid off directly in 2017, when I moved into building AI and machine-learning-based algorithmic trading platforms and exchange infrastructure, at a point when both crypto markets and mainstream trading desks were racing to bring genuine machine intelligence into execution rather than just analysis. The systems I built by then were a different generation entirely from the Forex bots of 2009, learning from live market behaviour rather than following only fixed rules, but the underlying instinct was the one I had carried since 1988: find the point where a human is making a repeatable judgment under pressure, and build something that can make that judgment faster, more consistently, and without ever getting tired.
An automated system will execute your logic flawlessly, including all the flaws you did not know your logic had.
I am not telling you this history to suggest that my path should be yours. Almost none of it looked deliberate from the inside. What I want you to notice instead is the pattern underneath it: every one of these projects existed at the edge of what people believed a machine could responsibly be trusted to do, and every one of them succeeded not because the technology was flawless, but because someone was willing to sit with it long enough to make it trustworthy in a specific, narrow, high-stakes context. That is exactly the work this book is asking you to do with the AI tools sitting in front of you right now.
The scale is different. The instinct required is exactly the same.