If you only check one highlight this week: how I used Granola, Claude and ChatGPT to create the first section about agents.

☀️Almost summer! I'm going to take a break for a few weeks as I'll be traveling with my family. See you in July.

Thanks for reading!

-François

What I learned about AI Agents from developers building them

You may have heard the tired phrase “2025 is the year of agents”.

The expression got our attention. It helped agent startups raise a boatload of $ and incumbents (e.g. Salesforce and Box) shape new powerful narratives.

But that hype doesn't help us tell what type of agents our teams should build (or buy). There are big differences between a simple task-focused agent and the ambition of advanced autonomous agents that take on entire roles.

What's really possible? What's just hyped slideware and predictions?

I attended the talks from developers actually building agents at the AI Engineer World Fair last week to learn more. Here's what they're really saying:

Bottom line:

• No agent is replacing entire jobs. That will take a while

• They're automating workflows and tasks, not roles

What agents do well today

Think of agents as smart interns who work 24/7 but need detailed instructions.

They're effective at:

• Data entry and processing (emails, analytics, etc.)

• Research and content aggregation

• Migrations

i.e. any task that follows predictable patterns.

Three factors make agents work:

• High impact when they succeed

• Low risk when they fail

• High probability of success

Current limitations

Agents can't yet handle well:

• Ambiguous requirements

• Tasks requiring business judgment

• Strategic decisions

• Nuanced customer interactions

Yes, reasoning models like o3 can think strategically. But you're still in the driver's seat making final decisions.

Most importantly: they need human oversight. Successful teams use "human-in-the-loop": agents propose, humans approve.

What this means for your marketing team

Use agents for:

• Competitive intelligence

• Lead enrichment and qualification

• Research

• Social media monitoring

• Campaign performance reporting

• and similar activities

These “Ambient agents" work in the background, triggered by events like competitor mentions or customer interactions.

Skip agents for:

• Brand-sensitive communications

• Customer-facing interactions (unless you tell customers they are dealing with one)

• Strategic decisions

• Anything requiring business judgment

How to start:

Ignore the hype to automate entire roles, start small:

1. Pick one manual workflow that eats time

2. Test with clear metrics + human review

3. Build "micro-agents" (20 steps max)

4. Make sure you can easily reverse their actions

This will evolve fast, like everything in AI land… but start here.

By the way, as I wrote last week, coding agents are much further ahead. Why?

1. they have access to code bases: they contain a lot of info, past decisions and code is well structures

2. code can be tested and reverting to a prior version is easy

🌭 Sausage making bonus: how I created the section above

Here are the workflows and tools I used to write the section above (I know some people automate much better, but I like staying in control and proceeding step by step):

1. Turned on Granola (AI note taker) to transcribe all sessions I attended at the conference

2. Organized sessions in various Granola folders (e.g. “building agents - hype and reality”)

3. Dictated my prompt (using Wispr Flow), in the Granola chat of that folder: ”Write a 300-word essay to help B2B CMOs understand what types of agents developers think they can build based on the following talks XYZ (+ some stylistic details)“.

4. Selected the Claude Sonnet 4 model to run that in Granola. I could have used ChatGPT o3.

5. Streamlined the copy and made it match my style by copying it into a Claude Project that I set up with instructions and a style guide*. I then simply hit enter and it streamlined the copy automatically without any prompt.

6. Edited the result

7. Generated the image above with OpenAI o3 by just dropping the text above into a new chat and telling it to “Generate an image using my usual style guide” (I had told it to save that style in its memory). It one-shotted it (i.e. got it right in one take), but I couldn’t help but prompt “C’mon, make me look younger”.

PS: very open to your tips for further streamlining this!

*To do that, I created PDF versions of 10 of my recent issues, attached them to a Claude Sonnet chat and asked it to create a very detailed voice style guide that I saved into another PDF and attached to that Claude project.

My selection of tips, news and workflows

🔍 How to increase AI visibility on ChatGPT, and Google AI

Noah Horner shares a quick recap from James Cadwallader’s work at Profound, which claims to have analyzed “30m citations/sources provided by ChatGPT, Google and Perplexity when answering a standardized set of commercial questions”:

🛠️ Storylane’s CMO shares his 2025 Program Budget and rationales

Madhav Bandhari shares how he’s spending his $3m marketing budget this year and the context behind that. It's refreshing to see a CMO opening the kimono like that. Remember that it's only one data point.

💡A Marketing Hackathon Playbook to accelerate AI usage

Carving space and allowing colleagues to experiment with new AI tools is well known to accelerate AI adoption.

Here’s a playbook (from MKT1 and Clay) to organize marketing hackathons that drive AI usage that will power new use cases, boost productivity, and help generate more random acts of marketing, i.e. the type that makes your brand stand out.

Includes:

• Why Clay run these hackathons

• How Clay hackathons work

• A step-by-step playbook to run marketing hackathons

• Examples of what MKT1 built during their own hackathon + tools they use

Want more? Learn how Twilio ran hackathons for enterprise customers in this lightning talk (their best pipeline generation tactic).

🛠️ Can you tell whether a chatbot wrote that? Only chatbots use em dashes so much

This Oxford-based linguist and computer scientist explains what gives chatbot copy away: em dashes!

FYI:
- hyphen

– en dash (option + dash)

— em dash (shift + option + dash)

PS: Another way to tell is perfect consistent formatting of each paragraph.

Final Words

Thanks for sharing these highlights with busy marketing execs around you.🙏 

Someone forwarded you this email? You can subscribe here.

François | LinkedIn 

I'm a CMO, advisor, and "CMO Wingman". Yes, that's a thing :-). Ask my clients: in this AI era, CMOs need a strategic proactive advisor more than ever. I’m former CMO at Twilio, Augment Code, Apollo GraphQL, Decibel, Udacity and Head of Marketing for LinkedIn Talent Solutions.