Most leaders I speak to have the same relationship with AI. They know it matters. They have read the articles. They have probably sat through a conference or two where someone showed them what is possible. And then they went back to work, got busy and handed it to someone technical.

I get it. That was me 18 months ago.

But here is the problem. Delegating AI to your technical team is not leading AI. It is outsourcing it. And the gap between companies where leadership is personally involved in AI and companies where it lives in one department is getting wider every month. We see it in our data. Leadership-level ownership of AI outcomes across the direct selling companies we have assessed scores 0.7 out of 5. That is not low. That is nearly nonexistent.

This article is for the leader who is done reading about AI and ready to actually do something. I am going to walk you through how to build your first AI agent workflow using Claude Cowork, step by step, with proper guardrails built in. No technical background required. No IT ticket needed.

By the end, you will have a repeatable workflow that saves you real time and, more importantly, shifts you from someone who talks about AI to someone who leads it.

First, what is Claude Cowork?

Claude Cowork is an AI agent tool from Anthropic. Unlike a chatbot where you type a question and get an answer, Cowork can perform actual tasks autonomously. It can open websites, navigate between tools, upload and download files, read and create documents, send messages, and chain all of these together into a single workflow.

You give it one instruction. It does the work across multiple tools. You review the output.

That distinction matters. This is not “AI that writes things for you.” This is AI that does work for you, across your actual tools, the way a capable assistant would.

Why this matters for leaders specifically

You could ask someone on your team to build this. And eventually, you should. But there is something that changes when you do it yourself first.

Once you have personally built and run an AI agent workflow, three things shift:

You can evaluate quality. You know what a good AI workflow looks like because you have built one. When your team shows you theirs, you can tell whether it is solid or sloppy.

You can set the standard. You are no longer asking people to do something you have not done yourself. That changes the conversation entirely.

You stop being afraid of it. The biggest barrier to AI leadership is not technical knowledge. It is confidence. The confidence gap closes the moment you do it, not the moment you read about it.

How to pick your first use case

Not everything is right for an AI agent workflow. The best starting point has four qualities:

Repeatable. You do it regularly. Weekly, fortnightly, monthly. If it only happens once, it is not worth automating.

Multi-step. It involves more than one action or tool. Opening something here, copying it there, formatting it, sending it somewhere. The more steps, the more time you save.

Currently manual. You or someone on your team is doing these steps by hand today. Even if it is “only” 30-60 minutes, that compounds.

Low risk to start. For your first workflow, pick something where a mistake is easily caught and corrected. Avoid anything that touches customer data, financial transactions or compliance-sensitive content until you have built your confidence.

Some examples that work well as a first use case:

• Processing and summarising a weekly report or industry update

• Researching a topic and compiling a structured brief

• Taking raw content (a transcript, a set of notes, a long document) and reformatting it into something usable

• Preparing a recurring meeting brief from multiple sources

• Monitoring and cataloguing industry news or competitor activity

If you are struggling to pick one, ask yourself: what is the task I keep putting off because it takes too long relative to how important it feels? That is probably your first use case.

Before you build: set your baseline

This is the step almost everyone skips, and it is the one that matters most.

Fewer than 5% of the AI initiatives we have assessed had any kind of baseline measurement before launch. That means 95% of companies cannot actually prove their AI is working. They assume it is because it feels faster. But feeling is not measuring.

Before you build your workflow, write down three things:

How long does this task take today? Be honest. Time yourself the next time you do it. Include the context-switching, the tab-hopping, the formatting. Not just the “core” work.

What does the output quality look like? Is the current output consistent? Does it depend on who does it? Is it thorough or rushed? You need a reference point.

How often do you actually do it? If this is a weekly task that you only get round to fortnightly because it takes too long, that matters. Frequency is part of the baseline.

Write this down somewhere you will not lose it. You are going to come back to it.

Building the workflow in Claude Cowork

Here is the practical part. I am going to walk you through the process using a real example from my own work.

My use case

I listen to a lot of AI research podcasts. When I hear something that could shape how we advise clients or think about the industry, I need to process it properly: transcribe it, verify the claims and data, interpret it through our perspective, and store it in our knowledge base. If I do this manually, it takes over an hour. If I do not do it, the insight disappears.

So I built a Claude Cowork workflow that does it for me.

The steps

Step 1: Define what you want to happen in plain language. Before you open Cowork, write down the steps as though you were briefing a smart assistant. For me it was: “Go to Spotify and find the latest episode of [podcast]. Download it. Transcribe it. Check whether the statistics and claims in the transcript are credible. Strip out the waffle and adverts. Reformat it with clear source references. Run it through my thought leadership skill to interpret the meaning. Ask me 3-5 questions via Slack to get my perspective. If I approve, commit it to our Notion knowledge base. Then ask me if I want LinkedIn content drafted from it.”

That is the entire brief. Plain English. No code.

Step 2: Connect your tools. Claude Cowork can work across multiple tools through integrations. For my workflow, it connects to Spotify (via browser), a transcription service, Notion, Slack and web search for the credibility checking. You connect these once and they are available for any future workflow.

Start simple. If your first use case only needs a browser and one other tool, that is fine. You can add complexity later.

Step 3: Build in your human review points. This is critical. The workflow should not run end to end without you seeing the output before it goes anywhere that matters. In my workflow, there are two human checkpoints: the agent asks me questions via Slack before committing to the knowledge base, and my social media manager reviews any LinkedIn content before it goes anywhere public.

Decide now where your checkpoints are. A good rule: anywhere the output leaves your personal workspace (gets shared, published, stored in a team system), a human should review it first.

Step 4: Run it and watch. The first time, watch the whole thing. Do not walk away. You want to see where it works well, where it stumbles, and where the output needs adjusting. This is learning, not just doing.

Step 5: Refine. Your first run will not be perfect. That is fine. Adjust the prompt. Tighten the instructions. Add a step you forgot. Remove one that was not needed. Most workflows take 2-3 iterations before they run cleanly.

After the first run: measure the difference

This is where your baseline comes back. After you have run the workflow a few times (at least 3 runs to get past the learning curve), measure again.

Time: How long does the task take now, including your review time? Compare it to your baseline.

Quality: Is the output better, worse, or comparable to what you (or your team) were producing manually? Be honest. AI that is faster but produces worse output is not an improvement.

Consistency: Does it run the same way every time, regardless of whether you are tired, busy or distracted? This is one of the biggest advantages of AI workflows that rarely gets talked about. Consistency often matters more than speed.

Frequency: Are you now doing this task as often as you should? If your weekly report was really happening fortnightly because it took too long, and now it actually happens weekly, that is a meaningful business outcome.

If the numbers are better, you have your evidence. If they are not, you have learned something equally valuable about where AI does and does not fit in your work.

A checklist you can actually follow

Here is the whole process distilled into a checklist. Use it as a reference each time you build a new workflow.

Before you build:

• Pick a use case that is repeatable, multi-step, currently manual and low risk

• Measure your baseline: how long, what quality, how often

• Write down the steps in plain language as though briefing an assistant

• Identify where human review checkpoints should go

When you build:

• Connect the tools you need in Claude Cowork

• Enter your prompt (the plain language brief)

• Run it and watch the full execution

• Note what worked, what did not, and what needs adjusting

• Refine the prompt and run again (expect 2-3 iterations)

After you build:

• Run the workflow at least 3 times before measuring

• Compare time, quality, consistency and frequency to your baseline

• Decide: scale it, adjust it, or scrap it (all three are valid outcomes)

• Document what you built so someone else on your team can learn from it

Before you scale:

• If this touches company data or systems, loop in IT/compliance with your evidence

• Show them what you built, what it does, and what the guardrails are

• Let the results make the case, not the technology

What happens next

You have built one workflow. You have measured it. You have evidence that it works.

Now two things happen.

First, you start seeing these patterns everywhere. Every repeatable task in your week starts to look like a candidate. That is normal. Resist the urge to automate everything at once. Pick the next highest-value use case and repeat the process.

Second, and this is the part that actually matters, you walk into your next leadership meeting as someone who has done this, not someone who has read about it. You can talk about AI from experience, not theory. You can evaluate your team’s AI work because you know what good looks like. You can push for the right investments because you understand what is possible and what is not.

That shift, from observer to practitioner, is the single most valuable thing a leader can do right now.

One workflow. That is all it takes to start.