By Tristram Dyer | CEO, Webtopia

Last week on a client call a founder asked me how many people work on his account. I said three. Plus fifteen AI agents. He laughed. Then I started telling him what each of the agents actually does. He stopped laughing.

That conversation is the reason I'm writing this. We've been building AI into how Webtopia runs for about eighteen months and we've never properly written down how it works. Founders ask us about it on most calls now. The thing is, the honest answer has very little to do with tools or stacks. It is a way of operating.

So here it is.

Two modes: fast and slow

Everything we do with AI falls into one of two modes. I think about them as Fast and Slow.

Fast AI is help in the moment. A person is at their desk doing the work, and AI takes the grunt work off their plate. The kind of stuff that used to take an hour now takes two minutes. Pulling data, summarising a meeting, drafting a brief, building a quick tool. Whatever the human would have done by hand, but faster.

Slow AI is the opposite. Nobody presses a button. It runs in the background on a schedule. The job is to make sure nothing important slips through. The strategy a media buyer recommends on Monday is informed by what happened across every account over the weekend, not just what they happened to remember.

Net net, Fast AI gets work done. Slow AI makes sure the right work gets noticed.

That distinction matters because most of the conversation about AI in agencies is stuck on Fast. ChatGPT to summarise a deck. Claude to draft a brief. Fine, useful, every team should be doing it. But that's the easy bit. The interesting work is in Slow, and almost no one is talking about it yet.

Fast AI: what it actually looks like

A media buyer at Webtopia opens his laptop in the morning. He's running four accounts. Before he's even on Slack he's already used AI three times.

He's pulled yesterday's performance across his accounts without logging into Meta Ads Manager, because we have a Kaizan connector and a Motion connector that he just asks in natural language. He's read the call notes from a client conversation he wasn't in, because Kaizan summarised it. He's built a quick comparison of two creative concepts, because Claude can read the Motion data and tell him which one is fatiguing.

That's all before 9am.

If something is repetitive, or it starts with "pull me this data" or "build me this", he reaches for AI before he does it by hand. Rule of thumb across the whole team. Everyone has a Claude account. The whole team uses it every day, from the most senior strategist down to the youngest analyst.

The work that flows out of this is hard to overstate. A creative analyst on our team estimated she saves probably four hours a day, give or take, on tasks that used to be manual exports and rebuilds in spreadsheets. Four hours a day across a fifteen person team is sixty hours a day of work that no longer happens by hand.

Look, that doesn't mean we use fewer people. We use the people we have to think harder. A media buyer who isn't manually pulling data has more time to actually look at the data. That's the trade.

Slow AI: the agents working overnight

Here's where it gets interesting. While the team is asleep, fifteen agents are running. Some of them every day. Some weekly. One monthly.

Every weekday at 3:04am, an agent called the Performance Anomaly Detector checks every account for problems. ROAS dropping, CPA spiking, spend out of pace. If it finds something, it DMs me before I'm awake. Half an hour later, the Budget Pacing Monitor does the same thing for spend versus monthly target.

By 8:35am, the Morning Briefing agent has posted a one paragraph performance update into each client's Slack channel. By 9:23am, the Daily Reports agent has dropped a detailed thread underneath it.

At 11am, an agent called John, our Meta Ads Operator, suggests changes for the day. Pausing fatigued creatives. Shifting budget to scaled ads. New ads to test. John never makes a change. He proposes it and waits for a human to approve.

At 5:14pm, Close of Day checks which flags from the morning got actioned and which didn't, and sends a recap. Friday afternoon, the Weekly Rollup ranks the most important items of the week. Sunday at 6pm, an agent called Mirror reviews how the agent team itself performed and tells us what to fix.

That's a lot. The count matters less than the coverage. The strategy work our team does on Monday morning is informed by what happened across twenty two accounts over the last seven days. Not what someone happened to notice. Not what made it into a status update. Everything.

The specialist agents

Some of the agents have a defined role. They work like a team member, except they're available at 4am.

Mia is our Google Ads specialist. She works like a specialist who's early in her career. Right now she supports our senior search team. The plan is for her to become genuinely senior in about three months, making proactive optimisation calls on her own. Today she analyses search terms for any account and any date range, finds negatives and checks every page of the site before recommending them, writes ad copy and extensions from the website, suggests new keywords using Keyword Planner, spots which search terms should be promoted to keywords, and writes fresh shopping titles for low performers to A/B test. She replies in Slack to anyone who tags her, on any account, any time. The team treats her like a team member who's slightly junior but always available.

John is our Meta Ads operator. He reviews suggested ad changes, sanity checks them, and puts them in Slack for someone to approve. He never acts alone. We built him deliberately that way, which I'll come back to in a minute.

The Config Steward keeps the system honest. Every day it looks for feedback we've given the agent team and out of date settings, and drafts a fix for someone to approve. The agents are improving themselves, with human review on every change.

The rule that makes it work

Here's the bit people miss when they hear we run fifteen agents on twenty two accounts.

We don't let AI act alone on anything that touches a live account.

Three rules underneath that.

Rule one: draft first, act on approval. Anything that touches a live ad account, a client Slack channel, or its own settings proposes the change and waits for a human yes. John doesn't pause an ad. He proposes pausing the ad. The Config Steward doesn't change the prompt. It drafts a new prompt for review. The Daily Reports agent doesn't pick what gets called out. It drafts a report and a human posts it.

Rule two: one source of truth. Every target, every budget, every account configuration lives in one place. No stray copies that quietly go stale. Because the moment you have a media buyer's note in one document, a strategist's spreadsheet in another, and the agent's prompt referencing a third, you get drift. The Config Steward exists to prevent drift.

Rule three: everything is tracked. Agent instructions and settings are version controlled. Every change can be reviewed and undone. If the Morning Briefing agent starts producing weaker summaries this week than last, we can see exactly what changed and roll back.

The reality is the rules are more important than the agents. You can build a hundred agents. If they can act on a live account without human review, the day one of them gets a target wrong is the day a client account loses money. The rules are what make this safe to actually use.

Why this matters if you're a founder

Two things, I think.

First, the cost of marketing operations is changing. A team that used to need six people to run twelve accounts can run twenty two with the same team and produce better work. That doesn't mean agencies will get cheaper. It means the spread between agencies that have built this and agencies that haven't will widen. Founders should be asking their agency what they've actually built, not what tools they "use".

There's a difference between an agency that has a ChatGPT account and an agency that runs fifteen agents on a schedule with version controlled prompts and an audit trail. Those are different operations. The first one is what most agencies are doing right now. The second one is rare. Both will call themselves "AI enabled" on a sales call.

Second, the upper bound on what an agency can notice is changing. We built the agent team so that nothing important goes unnoticed across twenty two accounts. The hour saving is a nice second order benefit. When you only have human attention, the things that get raised in the Monday meeting are the things someone happened to remember. With the agent team, the things that get raised are the things that actually matter. That's a different operating model.

I keep coming back to a conversation I had with a founder probably six months ago. He was unhappy with his previous agency, but he couldn't articulate why. We pulled his account and within an hour we had a list of fifteen specific things that should have been raised with him in the last quarter and weren't. None of them required new ad spend to fix. All of them required someone to be paying attention.

That's the bit AI changes. Not the doing. The noticing.

In a sentence

Fast AI gets the work done in the moment. Slow AI makes sure the right work gets noticed. People stay in charge of every decision that matters.

If you're running a founder led DTC brand at $5M to $20M and you've never asked your agency what they've actually built into their operations, ask. The answers will be useful, whatever they are.