Create, Understand, Act: A Simple Framework for What to Hand to AI

How the three buckets found their names

Our stated position is “Human-Led, AI-Supported,” and it sounds tidy now, but it came out of a messier place: trying things, keeping what worked, quietly dropping what didn’t. What stuck was a pattern. Almost everything we handed to AI fell into one of three buckets long before we had names for them:

• Things we wanted it to make from scratch,
• Things we wanted it to make sense of, and
• Things we wanted it to just go and do.

We call those Create, Understand, and Act now. But the labels matter less than the habit behind them, which is asking every time which bucket a task actually sits in before deciding how much to hand over.

Bucket 1: Create

This is the biggest bucket, and the easiest to trust. Say we’re building SEO content for a long-running client. We hand Claude the brand guidelines, the existing content, and the keyword research, and ask for topic ideas and rough outlines. Writers take it from there. They add judgment, tone, and everything a model can’t know about a client’s business, and nothing publishes without a review.

That single change pulled a content timeline from two weeks down to three days, and the client’s tone remains intact. Visuals work the same way: Canva’s Magic Media gets concepts and storyboards moving, and designers finish what the model starts.

Create is where AI earns its keep fastest, because a rough first draft costs minutes, not a relationship. It’s probably where most businesses’ AI use actually sits, too. Not exotic agent work. Just a faster route to a workable first pass.

Bucket 2: Understand

This is the bucket that used to eat the most hours. Getting campaign data out of half a dozen platforms and into something a client could actually use meant a full day of copy-paste before anyone reached the interesting part. Now Claude and ChatGPT pull it together, find the pattern in the noise, and draft a first read on what’s driving a result.

What doesn’t change is who signs off. A model can tell you what happened in the numbers, but it can’t own what we tell a client to do next, and it shouldn’t. That’s exactly why Understand stops at “here’s what the data shows” and not “here’s the advice.” The advice stays with a strategist, every time.

Bucket 3: Act

This is the newest bucket and the smallest, and the one we’re most deliberate about growing. We let Meta Advantage+ and Google’s AI ad features handle real-time budget shifts, creative testing, and placement optimisation, inside limits we set and review. They don’t run unattended, with no one watching the account.

Anthropic, whose models we run the agency on, describes the principle well in its guidance for people building with agents: good oversight isn’t approving every single action, it’s being in a position to step in when it actually matters. That’s the standard we hold our own AI-run workflows to. We extend an agent’s rope only after it’s earned it on a narrower version of the job. And we’ve started borrowing a pattern Anthropic uses internally: one process does the work, a second checks it, and only the exceptions land in someone’s queue.

If “agent” is still a fuzzy word, our plain-English guide to agents, MCP and Skills unpacks what one actually is and what it can reach. For the fuller, tool-by-tool version of these three workflows, see how we’re using AI in our marketing.

The decision checklist

The bucket tells you what a task is. It doesn’t tell you how much to trust AI with it. For that, we run through four questions before handing anything over.

Answer those honestly and most tasks sort themselves:

Once a task is sorted into a mode, the next question is who checks the output and when. That’s a different document, and we’ve written it: the AI guardrail playbook covers the review gates, the pre-publish checklist and the data rules that sit underneath all of this.

The line we don’t cross

Notice that the checklist always leaves something with a person. That’s deliberate, and it’s less a constraint than the whole point. People trust a tool faster when they can see where its authority ends, which is a big reason a recent enterprise rollout we ran reached 72% of 4,000 staff using AI weekly within 30 days.

The research tends to agree: work built on judgment and relationships is growing in value, not shrinking, even as the repeatable half of most jobs keeps getting faster. We protect that half on purpose, not by accident.