It could not even reach the internet, and I was hooked anyway by. Wim Cox (Netherlands)

 

Wim Cox from the Netherlands has been working with AI since 2023. Over the coming year he sets out what he built, what went wrong, and what he now refuses to let the machine do. First instalment: a text message from a colleague’s son.

A colleague sent me a message one evening in the spring of 2023. It contained a piece of writing by his son, who had typed in a handful of keywords and received this back. Astonishment on both sides of the conversation. How on earth does that work. My colleague shrugged, put his phone away and, as far as I know, has never given it another thought. That same evening I went and found the website of ChatGPT, or Chat, as I came to call it. For the first six months I said Chat GTP, consistently and wrongly. That has improved, which is roughly the extent of my formal qualifications for writing this series.

I did not see the possibilities or the implications. I saw a toy, and I began by playing with it.

Playing

The first thing I did was paste in an article from my newspaper and ask for it back in the style of one of our national tabloids. The result was very funny and, worse, entirely recognisable. Then the same article as a teenager would write it. Then as a man of my own generation would write it, which was less funny, because it was accurate.

Somewhere in that evening’s entertainment the useful thought arrived. If Chat can imitate a register on request, then I can specify the register I want. Not as a joke. For a letter, for a memorandum, for a client who needs something explained without being made to feel stupid. Everything in this series that came afterwards is that single realisation applied to work that mattered.

Leiden

Not long after, my wife came home in genuine distress. She had been given a day’s notice to organise an outing for her class. Most readers of this magazine know my wife, which spares me having to explain why her judgement on the result counted for considerably more than mine.

I sat down beside her and typed out what she needed. An outing in the city of Leiden for a group of sixteen-year-olds from a gymnasium — a school type for the quickest learners, if one wishes to avoid the word intelligent. It had to contain a treasure hunt. It had to finish in the Zweetkamertje at the university, the small room where doctoral candidates wait for the verdict and where the graduates have signed the walls for centuries.

A draft programme appeared in seconds. Being who she is, she found several things wrong with it inside a minute. We pointed those out, and a corrected version came back immediately.

It was not the first draft that impressed me. It was the loop. Objection, correction, new version, no sulking, no explanation of why the original had in fact been perfectly good. I had spent a career leading people through revisions and I recognised at once that this particular part of the job had just become considerably lighter. That tasted like more.

The first piece of real work

I had a new client at the time, a sole trader who wanted to move into a B.V. Thoroughly decent people, hard-working, with no grounding whatever in tax or administration, and no pretence to the contrary.

I asked for two things from one set of instructions. First, an explanation of the process written for a reader with no financial training, in plain language, short sentences, nothing condescending. Second, a step plan for my own use.

The explanation was good. The step plan contained errors. It also contained several things I had not thought of.

That combination is the whole lesson of the first year, and I would put it to the sceptics in this readership as follows. Everybody was saying at the time that Chat was useless because it made mistakes. That was true. It made mistakes. So did the staff I used to employ, and I did not conclude from that that staff were useless. I checked their work, I corrected it, and I gave it back. What was new was the other half. This one also supplements. It raises what else is in play. Much of what it raised I knew perfectly well, but I did not have it to hand on a Tuesday afternoon with three other files open.

That step plan has been amended, corrected and refined many times since, and I would not now start such a conversion without it.

Translation, and a small trick

Translation turned out to be the next thing. I had been using DeepL, which remains an excellent tool and which I still recommend. The difference was that Chat brought the context with it, rather than the sentence alone.

English and German I can judge for myself. Japanese I cannot, at all. So I worked out a crude control: draft the Japanese, put the result through DeepL back into Dutch, and read carefully where it had gone astray. Not elegant, and it caught a great deal. It also taught me more than I expected about the way Japanese business correspondence is constructed, which is a benefit I had not bargained for and which I would not have obtained from a dictionary.

I mention this because it is the first example of a habit that runs through everything since. Where I cannot judge the output directly, I build a second route to the same place and compare the two.

One project, one client, one tone

As the tools developed, then came projects. A standing instruction per client, held separately, applied automatically.

Educated reader, likes a full explanation. Short, no lecturing, get to the solution. Never open with the conclusion, this one wants the reasoning first. For one English colleague the standing instruction is that every message opens with a summary and the questions set out plainly, with the elaboration only after that. For the same man, on other days, the instruction is a maximum of three words: will do, no can do, talk later.

Readers of this magazine are invited to work out who.

That sounds like a small convenience, and for a fortnight I took it for one. It is the difference between a tool you have to re-brief every morning and one that knows how you work with each of the people you work with. That is the point at which I stopped playing.

What this series is

This is a log, not a vision. I am not going to tell you that AI will transform the profession, partly because you have read that already and partly because I do not know. What I can tell you is what I have done with it in my own practice and at my own cost, and where it went wrong.

Over the coming months: what I had to anonymise before I dared upload anything, and how that problem was eventually solved. Why I moved from Chat to Claude, and what a badly cited statutory provision costs you. The error values sitting in red in a spreadsheet that neither of us had looked at. Why production and review now happen in two separate places, and why segregation of duties did not lapse the moment the worker stopped being human. What all of this does to the fee. And, at the end, the question of who will train the next generation once the machine does the work by which one used to learn the trade.

Nothing that follows is advice on what to buy. It is an account of what one practitioner learned by doing it.

A note on the current state of the art, since we have come this far. These days the machine even gives you a running commentary on what it is doing while it works. Somewhere during the preparation of this article that commentary read, in so many words, drafting in a sardonic Dutch tone of voice. I have been described less accurately by people who have known me for twenty years.

One last thing before you turn the page.

So what do you think: did AI help me with this?