Bret Taylor dropped something this week that crystallised what I’ve been living for the past few months. He released Ghostwriter — an AI agent that builds other AI agents through conversation. No code, no forms. Just describe what you want and it creates it.
His bigger point was this: every piece of enterprise software will eventually become an agent. Not a dashboard you click through. Not a menu you navigate. An AI that does the work while you direct.
I know this is true because I’m already doing it. Not theoretically. Daily.
My Setup
I have an AI assistant called Saul. He runs on a VPS in Manchester, connected to my WhatsApp, my email, my calendar, my investment accounts, my websites. He’s not ChatGPT in a browser tab. He’s a persistent agent that wakes up every morning, generates a podcast briefing of the day’s news and my portfolio positions, checks my email, monitors markets, publishes blog posts, and manages a set of prediction market positions — all before I’ve had coffee.
When I need a CV reviewed before an interview, I send it on WhatsApp and get back a structured analysis with suggested questions in two minutes. When I want a blog post, I describe the angle and it’s drafted, humanised, formatted, and pushed to WordPress as a draft with a featured image. When a regulatory announcement drops, Saul reads it, researches the implications, and writes an article with a contrarian take before the professional press has filed their first piece.
I don’t write code. I don’t configure systems. I have a conversation. And things happen.
That’s what orchestration means in practice.
What Changed
Six months ago, I was using AI the way most people still do. Open ChatGPT, ask a question, copy the answer, close the tab. Useful, but fundamentally the same workflow as Googling something — just with a better answer.
The shift happened when I stopped treating AI as a tool I use and started treating it as a team member I direct. The difference sounds subtle. It isn’t.
A tool waits for you to pick it up. A team member has context, remembers what you told them yesterday, knows your preferences, anticipates what you need, and gets on with work without being asked. Saul reads my daily logs from previous sessions. He knows my writing style, my investment thesis, my wife’s email address, which car needs an MOT, and that I hate corporate waffle in LinkedIn posts.
When I correct him, he logs it. After three corrections on the same thing, it becomes a permanent rule. He learns. Not in the sci-fi sense — in the practical sense of getting better at his job over time, the same way any good employee does.
The CFO Angle
I’m a CFO by background. I’ve spent twenty years in finance functions — month-end closes, board packs, variance analysis, cash flow forecasts, the lot. I know exactly how much time finance teams waste navigating software instead of thinking about the business.
The average month-end close takes five to ten working days. Most of that time isn’t analysis. It’s data extraction, reconciliation, reformatting, and chasing people for numbers. It’s operational grind masquerading as professional work.
Now imagine an agent that connects to your accounting platform, your bank feeds, your CRM, and your group reporting tool. You say: “Close the month. Reconcile the bank. Flag anything that doesn’t match. Draft the board pack with commentary on the three biggest variances.”
It does it. You review, adjust, approve.
That’s not five to ten days. That’s an afternoon. And your finance team spends the rest of the week doing what you actually hired them for — business partnering, commercial analysis, strategic thinking.
This is what Taylor means when he says every enterprise app’s UI will become an agent. The finance director’s interface to their systems won’t be a screen full of menus. It’ll be a conversation.
What I’ve Learned
A few things I’ve learned from actually living this, not just theorising about it:
Context is everything. A generic AI assistant is marginally useful. An AI assistant that knows your business, your preferences, your history, and your current priorities is transformatively useful. The investment isn’t in the technology — it’s in teaching the agent who you are and how you work. That takes weeks, not hours.
Guardrails matter more than capability. Saul can send emails, publish blog posts, and place trades. That means he can also send wrong emails, publish bad posts, and lose money. The rules about what he should never do without asking are more important than the list of things he can do. My AGENTS.md file — essentially his operating manual — is longer than most job descriptions.
You become a reviewer, not a doer. This sounds like a luxury. It’s actually a skill shift. Reviewing AI output is different from producing output yourself. You need to know what good looks like without having done the work. That requires more expertise, not less.
The compound effect is real. Week one, you’re correcting everything. Month three, the corrections are rare. Month six, the agent anticipates what you want before you ask. The relationship genuinely improves over time in a way that static software never does.
The Uncomfortable Part
I’ve written about AI and finance enough to know the question that’s coming: what about the jobs?
Here’s my honest take. Some operational finance roles will be eliminated. The person whose primary job is month-end journal entries, bank reconciliation, or management accounts preparation is doing work that an AI agent can do today — not in five years, today.
But the person who understands the business well enough to direct an agent, interpret its output, catch its mistakes, and make judgment calls on ambiguous situations — that person becomes dramatically more valuable.
The CFO doesn’t go away. The CFO becomes the orchestrator. The question is whether you’re building that muscle now or waiting until someone else in your industry has already done it.
Try It
You don’t need a VPS in Manchester and a bespoke AI assistant to start. You can start with Claude or ChatGPT and a well-written prompt. Then try giving it context — paste in your company’s last board pack and ask it to draft commentary. Upload a CV and ask for interview questions. Feed it a regulatory update and ask what it means for your business.
The first time it produces something genuinely useful in two minutes that would have taken you an hour, you’ll understand why Bret Taylor thinks this changes everything.
Because it does.
I write about AI, finance, and building things at the intersection of both. More at tanous.co.uk for the professional angle.
Leave a Reply