We’ve come a scarcely believable distance with AI. There are mic-drops coming on the product side, but it’s the internal story I want to tell.
We haven’t rolled out the capability company-wide yet but our SMT are spending loads of time working with AI. I don’t mean asking them to render pictures (although trolling Pete Farmer has reached a whole new level of potency), or even ‘chatting’ as most people will with an LLM (although Comet and Perplexity have changed the game for search and browsing), I mean getting shit done.
From reconciling Gigabytes of CDRs in order to identify why a peer invoice doesn’t match our accrual, to answering the mail that used to eat my mornings before the first green tea, this is proper work done properly, not a parlour trick.
Email is the mundane end of it, but telling. Triage was the start: the AI and I worked my inbox together for a few days, and every decision we made got written down as a standing rule – this sender is trash on sight, this invoice forwards to accounts and archives itself, this newsletter gets labelled and left for me to actually read. Deterministic rules became native Gmail filters; judgement calls got documented. But triage is table stakes. The bigger win is handling. Dozens of mails a day just need delegating on, and now they are. The monthly reports that needed a human to open the attachment and key numbers into another system by hand? Handled. The regular jobs that took me hours and could never be conventionally automated because they involved clicking around a browser? Handled too, because it clicks around the browser. When it meets something it hasn’t seen before it asks me once, one concrete question, and writes that answer down as well. What’s left in my inbox is only what actually needs me, which turns out to be a fraction of what lands there.
Another for the list, yesterday: month-end. The ritual of grabbing valuations of our treasury and posting journals to our accounting system, via arithmetical gymnastics that give me a headache every single month. This month I didn’t do it; the AI did, with me watching. And because every step was documented as we went, it does it from now on. Not quite the conversion moment (we’ll come to that), but definitely OMG.
That loop is the real discovery: do the task together once, document every decision as it’s made, and next time it just happens. Doing, documenting, learning. Toyota turned continuous improvement into a religion and called it kaizen. This is kaizen that remembers. We’ve gone further and built what I’ve come to think of as a Simwood OS: a folder of plain markdown files holding who I am, how I write, what the business is, and every standing rule we’ve agreed. The heart of it is a wiki that the AI builds and maintains itself from everything we do together, ingesting my daily notes (and doing any tasks therein), filing what it learns, and auditing itself for contradictions and rot on a schedule. That wiki is the main thing, because it means I can talk to it like a super-close EA with absolute context: it knows the people, the projects, the history and my preferences, because it wrote them all down as we went. I often return to my desk now and just ask “what do I need to know?”, and get exactly that. It’s like having a Gen X worker who remembers everything I tell them, calls me on the contradictions, and still remembers when I’ve forgotten. Recurring jobs get codified as skills, so “pull the CDRs for these 20 accounts for the last financial year [because the auditors have nothing better to do]” is now a sentence rather than an afternoon. The documentation isn’t a by-product of the work; it is the work.
But the moment that actually converted me, the holy-fuck moment, was watching it write 12 blog posts and 3 product docs in appropriate voice, from technical documentation that had itself been written by Charles working with his AI. AI distilling AI, each of us in our own voice, weeks of writing done before lunch. My voice is codified too BTW: the style guide even has a register it calls “full Simon”, which this post was written in. I still edit, invariably because the result isn’t Simon enough, which I choose to find reassuring.
Looking at this another way: I’ve long maintained a second brain. Every time Pete forgets what FTR is (read: every time Pete needs to know what FTR is), it gets consulted as well as hundreds of other times every day. It was transformational, but it was unmaintained and fixed in time. The Simwood OS now is the second brain. All those years of notes are imported, just part of the knowledge and context that gets used, and upgraded, at every turn. Strictly, today, it’s more Simon OS than Simwood OS: my role, my knowledge, my rules. But imagine it shared: Charles’ AI drawing on my knowledge, level 3 support querying both of ours, guardrails deciding whose word wins on what. That’s where this is heading.
Contrast that with Standard Operating Procedures, which exist so a process can survive contact with a new human: write the steps down, stick them in a wiki nobody reads, and hope. SOPs were unknown here until the clipboard people arrived; the clipboard people are thankfully long gone, but their legacy – not all bad, if I’m honest – remains. Those documents are still used, but some are incomplete and others out of date, which is the trouble with the genre: they start rotting the moment they’re written, because the person doing the job quietly evolves the job and nobody updates the document. And that’s the best case. In the worst case, SOPs are written by clipboard people who don’t actually know how to do the job, and you get the blind leading the blind based on what they’ve never done. Either way, an SOP is a lossy description of how work really gets done, written for someone who skims it once on day one. But when the operator is an AI, the procedure is the execution. The document doesn’t describe the process, it runs it, and because it’s corrected every time reality disagrees, it can’t rot. That’s not a procedure, that’s memory that executes. SOPs, it turns out, are for saps.
There was a stat in Business Leader this week that caught my eye: according to the president of NetApp, large companies typically use only 30% of the data they store. The other 70% just sits there. His advice was to bring AI into your data rather than dragging your data to some grand AI project, and his observation was that smaller firms are closer to their own technology and data, so the playing field doesn’t just level, it sometimes tilts in the small guy’s favour. Quite. We’re engineer-led, and the marginal cost of pointing intelligence at decades of data is now a few minutes of teaching. But here’s the thing: we were already unambiguously a decade or more ahead of the dino-carriers, because we’ve always moved faster than them and run at much lower cost. AI doesn’t create that advantage, it amplifies both halves of it. Recent posts on this blog came out of this machine, as will the mic-drops I teased at the top, delivered in days rather than quarters. A dinosaur can’t buy that outcome; it has to unlearn its way there through three layers of procurement, a data governance board and, at the other end of the diligence and competence scale, whatever’s left after the Ferrari payments. The gap widens every time the machine writes another rule down.
A word on deployment. We’re doing this top-down, deliberately: technology like this changes the game company-wide, so the SMT are learning it first-hand rather than commissioning a strategy deck about it. But as Spiderman said, “with great power comes great responsibility”. The last thing we want is to repeat a Klarna, or any of the litany of big-company AI train-wrecks: fire the people, put a machine where your best judgement used to be, apologise, quietly rehire. Our AI faces inwards, amplifying people who know what good looks like. The judgement stays human, and so does the accountability. And before the pearl-clutching starts: this all runs under the same controls as everything else we do. Our data stays ours, nothing customer-facing is left to a machine, and nobody’s CDRs are training anyone’s model. If a dino-carrier proudly tells you they don’t let AI anywhere near their business, believe them – and picture what that looks like in five years.
Pete wrote recently about the four AI strategies of our time. This is ours in microcosm: not bolting a chatbot onto the same old dysfunction and calling it transformation, but rebuilding how work happens from the inside, one documented decision at a time. None of this is written to gloat, by the way: our customers are mostly exactly the smaller, nimbler firms this playing field now leans towards, and we would far rather arm you with it than crow about it. Enshitification is one path; teaching your business to remember is another. Simwood OS never forgets what we taught it yesterday. What did you teach yours?
