Most people use AI as a faster typewriter. I run it as working systems, each with written rules, real data and a person at the approval points. That’s the difference between quick, and quick but wrong.
Anyone can get a paragraph out of a chatbot. Getting reliable work out of AI, week after week, takes the same discipline as any other production process.
A fresh prompt typed into a chat box each time.
Reusable systems with a written procedure, so the same job is done the same way twice.
Whatever the model happens to know.
My own knowledge base: services, benchmarks, past jobs and your brand rules, loaded before any work starts.
The first answer gets used.
A loop of concept, production and critique, so weak options are killed before a person sees them.
Output is trusted because it reads well.
Defined approval points, and a record of what was decided and why.
The same mistake turns up again next week.
Feedback becomes a written rule, tested against past work before it’s adopted.
AI lives in a browser tab, apart from the real tools.
AI is connected to the analytics, content and data tools where the work actually happens.
“AI makes it cheap.”
AI shortens production. Judgement, direction and checking take as long as they ever did.
These are working tools in my own practice. They’re also the pattern I use when building AI workflows for clients.
A brief goes in and a scoped, costed proposal comes out for me to review. It works from a services catalogue, effort benchmarks and rules for reading a brief. When a job finishes, the real hours go back in, so the next estimate is better.
It first builds a working design system for a brand. A brief then runs through concept, render and critique against that system and a library of proven ad formats, and I approve or reject the finished candidates.
What I know about pages for paid traffic, written down as rules the AI must follow: one purpose per page, a message that matches the ad, and tracking that makes the click attributable. Pages are drafted and audited against it.
Campaign pages, tracking code, data pipelines and queries are written with AI coding tools working inside the real project, then tested. This website was built that way, from a written brief and the brand’s rules.
My own application for marketing mix modelling, built on Google’s open-source Meridian. It guides a team through mapping and checking their data, trains the model in the cloud, and shows diagnostics, results and budget scenarios in a browser.
Scripts written around an agency’s own workflow. The largest balances a national Google Ads account by itself: it keeps spend on pace and moves budget towards the campaigns with the best return on the next dollar. Before it raises a bid, it checks whether the bid or the quality of the ad and landing page is holding the campaign back. Others sort thousands of keywords into groups and produce bulk ad builds ready to import.
Competitor pages, search results and AI answers collected and compared automatically, so audits rest on current evidence.
The hours AI removes are typing, formatting and first drafts. What’s left is deciding what’s worth doing, directing it and checking it, which is the part you were paying an experienced person for anyway.
A person approves it
Nothing goes to you, your client or the public without me reading and signing it off.
Built and checked by different hands
What “done” means is written down before work starts. One AI agent builds, a separate one reviews, and I check the result in the real product before it ships.
Code does the arithmetic
Anything numerical, such as a calculation, a lead score or a budget figure, is done by tested code. AI isn’t asked to do sums.
Facts need a source
Numbers and claims are checked against your data or a named source. AI doesn’t get to fill gaps with something plausible.
Your data is handled deliberately
Before any build, I set out which information goes into which tool. Sensitive data stays out of services that don’t need it.
Honest about what’s unproven
Some of this is new. Where a method hasn’t been tested enough to rely on, I say so and price it that way.
You own what’s built
Systems built for you live in your accounts, with documentation, so they keep working without me.
Two free tools built the way this page describes, and the service where it all comes together.
Free tool
Describe a marketing, advertising or tracking problem and get an outline of the work that would fix it.
Read moreFree tool
Turn a short campaign brief into creative directions, built from a reference library of real campaigns.
Read moreB2B
Plan, pages, creative, media, lead capture and CRM, built as one connected system.
Read moreProduction is quicker, and quotes reflect that. The hours that remain are the ones that decide whether the work is any good: briefing, direction, review and testing. Those haven’t shrunk at the same rate, and I don’t pretend they have.
Generic content does badly whoever wrote it. The fix is the same: start from specific knowledge of your business, write to your voice and edit with care.
Mostly Claude for reasoning, writing and code, coding tools such as Cursor, and image and video generators where a campaign needs them. The tools change often. The method matters more.
Yes. Building a knowledge base, written procedures and approval points for your own team is some of the most useful AI work I do.
Client material is used only for that client’s work. I’ll tell you which tools are involved and what they receive before anything starts.
Tell me the task you’d most like off someone’s plate. I’ll tell you whether AI can do it safely, and what it would take.