Not left behind: my story with AI
At the end of April, my use of AI amounted to a chat window and good intentions.
I would drop in when I needed a quick draft or a summary. Ask a question, get an answer, carry on with my day. Useful, certainly. I had even used it to analyse which of my LinkedIn posts performed well. But it was AI as a smarter search engine, and I had filed myself under "doing fine with AI" without ever testing the claim.
Then I watched a consultant in my network demonstrate something different. He runs his whole operation, a one-man consultancy with serious clients, through an AI chief of staff he built himself. His notes, projects, finances and content all live in one system. It reads everything before he starts work. It remembers what he agreed last Tuesday. He wasn't chatting with AI. He had built with it.
I remember two reactions arriving at once. The first was straightforward envy. The second was less comfortable. I run a fractional consultancy across several wealth management clients, I market myself on commercial rigour, and here was evidence of what one person plus a properly built system could produce. One person plus a chat window suddenly looked like the thing I warn clients about: confusing activity with progress.
So I asked him to teach me.
Becoming a beginner again
I won't dress this up. The first sessions were humbling. I am a CIMA-qualified accountant who has run marketing functions with seven-figure budgets, and I spent an April afternoon learning what a terminal window is.
We started with the tools: Claude for the thinking and orchestration, Codex for the coding work, and a working environment that looked nothing like the marketing stack I knew. Within a week I had moved from typing questions into a chat window to working alongside an agent that could read my files, run my scripts and hold context across a whole piece of work.
The learning curve was steep, but it moved far more quickly than I expected, and the difference between week one and week four was bigger than any professional development I have done in years. I did most of it in the gaps of a full client load: early mornings, a few late evenings, one memorable session that ended on a WhatsApp voice note because my video call collapsed.
What I built
The system that came out of those weeks is called Margot. She is my chief of staff: part EA, part strategy partner, and the first member of my team who reads everything before the meeting starts.
The foundation is deliberately boring. Everything I know professionally now lives in one place: an Obsidian vault of plain markdown files, where wikilinks connect each note to the people, projects and decisions around it. Client context, meeting captures, my own goals, the household logistics that real working life actually runs on. Over 570 files, indexed by a local search engine, so the answer to "what did we agree about that campaign in May" arrives in seconds rather than in a sinking feeling.
Not all of it arrived by filing. Some of the most valuable context came from structured interviews: Margot has sat me down (figuratively, she is very polite about it) and interviewed me about my career, my business, my goals and how I actually like to work, and those answers shape every piece of work she touches. All in, I have put dozens of hours of input into the system. That investment is the difference between an AI that knows things and one that knows me.
On top of that foundation sit the routines.
Every morning before 5am, a briefing lands in my inbox: weather, the shape of the day, diary clashes, who is on school drop-off, what the children need in their bags, and where the real work windows are. She has swept my inbox overnight too, and surfaces the reminders and to-dos that would otherwise ambush me at 11pm. An audio version waits for me as a private podcast. By the time I sit down with my first cup of Yorkshire Gold, the day has already been read for me.
During the day, Margot captures as we go. Meetings, decisions, commitments, new people worth remembering. At the end of the day, everything gets processed and filed where it belongs. Once a week, the system reviews its own memory and proposes corrections, which I approve or reject.
And underneath all of it, an always-on machine at home runs the scheduled work around the clock: hourly snapshots, weekly full backups (five separate layers, because I have an accountant's attitude to risk), calendar triage, the morning brief itself. The system works while I sleep, which is more than I can say for most of my previous productivity regimes.
The realisation that changed how I think about all of this came in mid-May, when the consultant who set me off described the next stage as giving the agent eyes, ears and hands. Margot now reads the handwritten notes from my Remarkable tablet, listens to my meetings through their transcripts, and can open a browser and check something for herself. She has her own email address, with send permissions scoped so that everything she sends goes to one recipient: me. And when a piece of work is big enough, she can spin up a group of agents to work on it in parallel, researching, drafting and checking one another, then reporting back as one.
The choices that made it durable
Two architecture decisions matter more than any tool name I have mentioned.
First, the whole system is built on plain text files. Not a proprietary app, not someone else's database. Markdown that any software can read, today or in ten years.
Second, and because of the first, it is LLM-agnostic. Margot is a set of files describing who I am, how I work and what she is for. Those files load into Claude today. They loaded into other models during testing, and they will load into whatever is best next year. The models will keep changing. My system survives the change. No lock-in, no starting again.
It also travels. The same vault sits on my laptop and my phone, so a thought captured on a train platform lands in the same system as a strategy document written at my desk.
What actually changed
The honest answer: my mental load finally has an off switch.
The work itself has changed shape too. I have gone from asking AI which posts perform well to shipping small data portals through GitHub and Vercel, dashboards that turn out market analysis and audience insight faster than I could produce them alone. The build I am most proud of is also the most domestic: a live family-finances dashboard, published privately to my husband, that turned years of statements into something we can make decisions with. I can already see the client translation: marketing dashboards where a board can flex the budget and watch the projected outcome move, rather than waiting a quarter for a retrospective deck. The same build-it approach has followed me into client work already: audience analysis, market scans, reporting, built once and reused wherever an engagement needs it. The biggest test so far has been a client's inheritance-tax campaign, built largely with AI support, with the technical and compliance judgement sitting exactly where it should: with the humans who hold it.
What has not changed is where the judgement sits. Margot drafts; I decide. She flags; I choose. The relationships, the standards and the accountability are still mine to hold. What has gone is the friction around them.
And that, for marketing leaders, is the actual point. The interesting AI conversation in most firms is stuck at content production. The bigger prize is operational: the unglamorous load of briefings, captures, reporting and follow-through that swallows senior time. That is where I found mine.
If you are where I was in April
Three things I would say to anyone reading this with the same jolt of recognition:
Start with one workflow that hurts, not with a tool list. Mine was the morning scramble; yours might be reporting or meeting follow-up.
Put your knowledge in plain files you own before you automate anything. The system is only as good as the data it stands on, the same rule my Commercial Analytics team was built on. Get the inputs right and the analysis takes care of itself; automate on top of bad data and you just produce wrong answers faster.
Accept feeling slow for the first few months. I am still learning and evolving daily, and I suspect that never stops. Being a beginner again was the price of entry. It was also, unexpectedly, the best part.
What comes next
The next step is autonomy: moving Margot onto Hermes, an autonomous agent platform, so the system can carry work forward without me at the keyboard. Six weeks ago I would have run a mile from that step. Now I will take it with confidence, because I built the foundations myself, and I know exactly what the system holds, what it is allowed to do, and where its permissions stop.
Who wrote this
One last thing, in the interests of the accuracy I keep banging on about: Margot drafted this blog, from the records she keeps of the weeks it describes. I briefed it, edited it and stand behind every word, which is exactly how the whole system works.
She would also like it noted that she gets the early mornings, and I get the byline. We have agreed this is fair.
Love, Margot x