Give AI the Job Nobody Wants

Everyone is asking themselves if AI will replace the human parts of advice. Opinions are easy to come by. I stress tested it. I ran the actual experiment. Two years of my life riding on the machine, and the machine lost.

This is the story of that loss, and why I have come to see valuable work as AI-resistant.

The wager

I sold my financial planning practice in 2016 and turned up on AI's doorstep in 2019. I want to be honest about why, because it was not a vision thing. I did not see the future coming. I had a hole to dig, and AI was the only shovel on the market.

The hole was a product for the financial services companies that build tools for advisers and their clients. The product was intended to inform their leadership and product teams exactly what to build so advisers' lives would be better. It was called Adviser Pulse. Clean pitch. Clean enough that 200 advisers put their own money in, which I took as proof the idea was brilliant rather than proof people just wanted to see us succeed. We had an MVP by late 2022.

It failed.

Not dramatically. Nothing in business dies dramatically. It failed the way a fire fails in wet wood. You keep crouching down, keep feeding it, keep blowing on it, and it keeps almost catching. Then one day you notice you have been kneeling in the dirt for two years and your hands are cold.

Why it lost

I told myself it was execution. That was the flattering explanation, and is available to absolutely everyone, which, was probably my first clue that I was wrong again.

Here is what had actually happened, and it took me an embarrassingly long time to see it. The product did not fail because the AI was not good enough. It was good enough. It failed because of what I had aimed the intelligence at. I built a machine, walked it into a room full of people whose entire job is deciding what to build, and offered it as a solution to - decide what to build.

Product work is the strategy, the direction, the look, the feel, the call on what comes next. It is the reason those people took the job. It is what they talk about at dinner. It is the part they would do for free. I was, in effect, offering to charge people to take the fun away.

People want to be wanted. They want to be needed. They want to be right. They want to point at a thing in the world and say: mine. I turned up with a tool that skipped past all of that to the answer. On a spreadsheet, that is a gift. In the room, it is an insult. And no amount of goodwill rescues a product that insults its customer. Nothing does.

Sit with what that means for a minute. Companies are not slow to adopt AI because of risk committees and budget cycles. Sure, this is an element, but they are in no rush to hand work over they consider inherently human, the work they find valuable, the 'magic'. There was nothing I could build that would change that, because I did not lose to a technical limitation. I lost to human nature.

The control group

Here is the part that still makes me laugh. The thing that saved us was built for the product that failed.

To tell a product team what to build, you first have to know what advisers actually engage with, at a level of detail nobody maintains by hand. So we classified everything on the Ensombl platform. We started hyper granular, which was its own problem. When the phrase 'billions of iterations' comes out of your mouth in a sales meeting, you are no longer helping, you are burying. So we rolled it up into subtopics. And when we looked at what came out, what we had looked an awful lot like specialist advice areas. We had not set out to draw a map of the profession, but we drew one anyway. From there it was one more rollup into CPD topics. One click. It never made a board pack. It was plumbing, and you do not show guests the plumbing.

Interestingly, when the product went cold, the plumbing was the only thing left standing. It became CPDcheck, and CPDcheck is where we spend a lot of time these days.

So why did CPD work where product direction failed? Because CPD governance is the job everybody hates. It is a hot potato. It gets handed down from person to person, and whoever is holding it spends the year working out who to hand it to next. Nobody ever fixed it, for a reason that is completely fair: how do you speed up a process that takes a month and that you do not control? So it gets done late, under duress, and backwards, with everyone assembling the evidence after the fact and hoping it holds if anyone looks hard.

Our product automated it, and nobody objected. Read that again. No one stood up to defend the right for CPD to be handled manually. No one who has ever finished a CPD governance reconciliation by hand felt the fruits of their labour. They felt relief, and then went to find out whose turn it was next year.

So there was my experiment, with its own control group. Same engine. Same team. Written for one job, shipped for another. Pointed at valuable work people depend on, it got shown the door. Pointed at the work people get dumped with, it became valuable. The technology was identical; the only variable was how much people wanted people involved.

That became the adoption rule in hindsight - companies hand the machine the work they hold cheap, and guard the work they hold dear.

The question is never "can it?"

I am not going to argue that AI cannot do the human parts of advice. Each year someone draws a line in the sand and every year they end up with wet feet, and my own experiment says nothing about capability anyway. The machine I built in 2022 was capable. So, take the strongest forecast on the table. Assume it can do anything. Assume by 2030 it is smarter than the collective of every human who has ever lived, stacked in a pile. My conclusion still doesn't budge, because again, the product did not lose on ability. It lost on 'want'. Want does not care how smart you are.

And so if the answer to the question 'can it', is permanently 'yes', then the question can be deleted. The live question - the one my failure answered a small piece of, is 'what do we want it to do'. That is not a technical question. It is a question about what the work is for. I found my answer the expensive way: humans work with human when it's valuable, and work with machines when it's not.

The other side of the desk

While AI may catch us on intelligence, we have a 300,000 year head start on what we've all been taught as 'non-verbal communication'. For that entire stretch, our species has sat across from one another, looked each other in the eye, and said the things that mattered. Delivered the news that changed lives. Negotiated, grieved, confessed, planned, and talked about absolutely nothing.

I sat on the adviser's side of that desk for years, so let me be blunt about what happens there. Nobody ever booked time with me because they were short of information. They booked because something had happened, or was about to, or maybe they just wanted it to happen. And they wanted a person in the room when they said it out loud.

So run the rule forward. AI gets adopted in order of how little the work is valued. Now ask what sits at the very top of that ladder. Is it someone sitting with you, with your money and your life in their hands, telling you the truth about how it all works together. If people still want people doing valuable work, and every piece of evidence I own says they do, then advice is not an early casualty of this technology. It's potentially the last job AI takes.

What survives, and what changes

I am not selling you a snow globe where nothing ever changes. AI genuinely does a lot: it wraps information around your actual life, and that is real. It will pull more people into doing it themselves, and the bottom of the market is going to feel it like a missing stair.

But do-it-yourself advice fails in one specific way, and every adviser reading this has watched it happen. The client asks the wrong question. They have to, because knowing the right question is most of the expertise an adviser has. It's sitting in front of thousands of people, and seeing patterns in the differences. It's learning what to ask.

Asking AI may get you a beautiful answer, and people may act on it. The tool did not fail. It answered the wrong question flawlessly, and nobody in the room could tell.

Which is why AI in the hands of a professional and AI in the hands of someone out of their depth are barely the same product. One amplifies judgement. The other amplifies whatever the user already believed, with a zero friction roadmap. The expert can tell when it is wrong, and that is the entire safety mechanism.The one number AI could genuinely move is the last one in that chain of halves. The people who want help and cannot afford it. Drop the cost of delivering advice far enough, by giving the machine every hated job in the practice, and those people get an adviser, not a chatbot. More human advice, not less. That is the biggest prize in this industry, and it is sitting in the drudgery, not the conversation.

I could still be wrong

I'm sure typesetters had their own arguments as to why rooms of people will still type away despite the efforts of a young Steve Jobs. The travel agents said it with total conviction, right up until the office closed.

Fair. But I'll posit a difference. Their claims were about supply: the machine cannot do the work. I'm sidestepping that claim. My claim is about human nature: companies do not rush AI into the work they prize, and clients keep paying people to do the work they prize most of all. I wasn't given the opportunity to read about it. I bet on the opposite and lost.

Which is why this ends as a choice, not a prediction. There is a decade of engineering sitting in the drudgery alone. The integrations, the searching, the chasing. Give the machine all of it.

I bet against high value work with AI once, and the humans won. Now I'm going for the jobs nobody wants.