AI is everywhere – but are you actually using it?

Last week, two of us spent two days in deep AI work time.

AI is now the most frequent conversation topic of anyone in business I know now. One of our copywriters Mark said to me recently that “every conversation eventually turns to AI”. It’s become the modern Godwin’s Law.

There is a mixture of excitement and anxiety.

The AI influencer community is already in full swing accelerating that anxiety. If you believe every AI influencer then you might think there are entire companies now fully operated with just 25 AI agents and one person. You’re probably worried you’ll be left behind.

Anxiety in times like this is good. One of the best business books you’ll ever read is Only the Paranoid Survive by Andy Grove. The period we’re going through now is one of existential change, and if you bury your head in the sand, you probably will be left behind. Realising that this is an issue is important.

But at the same time, once you dig in to most of what the AI world is currently selling in today-terms, I’m often left wondering how many people actually use all of this day-to-day?

And what is there today that can be useful, recognising that ultimately every part of a system and workflow today will likely feel old in six months time.

I want to share very openly how we’re experimenting at the moment and our approach to it. I hope this is less a ‘here’s stuff you can borrow to be better at growth’ and more ‘here’s how we could experiment in our organisation' specifically.

Things to cover

1. How we ran our recent AI hackathon/weekender

2. Key lessons learnt

3. What we built

4. What comes next

How we ran the AI Weekender

“We’re a highly lean agency, as are most businesses I know these days, and so most of the team are at 100% capacity all of the time. That means, there isn’t by default idle time for experimentation.”
– The 43 tools that make the modern growth AI & tech stack at Ballpoint

Most teams I know now are highly lean. Revenue per employee as a metric is back. And with AI companies now reaching tens and hundreds of millions with only a handful of employees, that will only continue.

But one of the issues with operational efficiency is that it doesn’t allow room for experimentation and play.

New systems and technology require experimentation and play. And there’s only so much snippets of time in evenings and weekends you can do to properly explore.

This was the backdrop to why Fiona, our head of ops and creative studio lead, and I went and did an our own Weekender. Except it wasn’t over the weekend, and we didn’t go clubbing. We spent two days of blocked out, deep work time, off site working only on AI.

Here’s how we did it

• Booked a workspace and hotel away from London to create different environment setting

• Blocked out calendars entirely, did handovers as if we were on holiday

• Came up with a long list of ideas ahead of the two days of what we wanted to build

For us this meant two days in Guildford where we worked solidly in deep work mode on AI ideas for two days.

What do we really mean by AI?

In this context important to point out the parameters.

1. The use of AI within automation frameworks to help us automate things we couldn’t have done previously

2. Exploration specifically of AI tools and how we could bring those into our workflows

How we worked throughout the day

We had a Notion board with our long list of ideas, and we then gave each an expected value.

We measured: (1) potential forecast contribution margin over a year the initiative could bring to the business and (2) probability we could finish it in two days

The CM forecasts looked at:

1. Can this reduce existing human time to the business?
   For example, can we replace 90 minutes of human time per week with 5-10 minutes of human time by automating a repetitive process? And therefore we can calculate 75 minutes of salary x 52 weeks of the year to give us a CM figure.
   

2. Can this accelerate our ability to do our jobs better and therefore increase client spend, which would then increase our revenue as well?
   Harder to forecast, but to be simplistic, let’s assume 1 ad in 20 we create allows us to push client spend by 15% per month, does this activity increase that ratio to be 1 in 18 and therefore offer a 10% increase in potential revenue to the business? Apply our gross margin to that revenue increase multiply it by number of clients it would impact, and hey bingo we’ve got a CM figure.
   

3. Does this activity reduce the likelihood of churn through better servicing of the client?
   Another hard to forecast one, but we have a set of principles we believe are important to client retention. Can we improve our ability to do that through an activity? And to do this we looked at our churn rate across the year, forecast a cost of churn in revenue terms, and again applied the CM calculation to it.
   

4. Does this improve our marketing of our agency?
   Very hard to quantify as measuring individual pieces of activity within the agency’s marketing is difficult. Yes we can look at substack reads/engagements, or LinkedIn likes, or follower increases, but ultimately, we don’t have a good answer to specifically how much that ties back yet.

And then we did a probability forecast of likelihood to complete within two days.

---

Key lessons learnt from our AI experimentation

1. Computational thinking

First and foremost, this felt closest mentally to coding work that I’ve done in a long time.

This is process of approaching a problem in a way a computer scientist would do. It requires thinking through the ideas of like data structures, inputs and outputs, loops, and abstraction.

I’ve done CS50 in the past, and am fairly good at SQL, but I never really got into a proper language. I found it incredibly useful however, to understand things like defining variables, or iterating across loops.

For anyone who’s never done any coding or computer science stuff, in plain English this means:

• Breaking apart your day-to-day actions into the tiniest and most specific steps. For example, you might say:

  • press the button on the top-left of the coffee machine that has a has a circle with a dashed line through the top of it.

  • Wait for two minutes until the machine stops making any noises

  • Put your hand on the handle of the wooden block above the coffee machine, grab it, and move it to the right.

  • Pick up the bag that sits on the lowest shelf (position 0), on the far left of the shelf. Check that the bag says “coffee” on it.

  • and so on

• When presented with an either or action, using IF statements to set conditions and directions. For example:

  • IF the coffee machine has no coffee in it THEN add coffee to the back ELSE continue to next step

• Thinking about everything as data. I.e. every action or object or anything needs to be a data structure. And to use them properly, you’ll need to define them a lot. For example:

  • Every object in the cupboard is an object in a data array with different parameters

  • Every mug on the shelf is a different object

• When presented with a list of tasks you want to repeat for a list of things, ‘iterating’ with a ‘for loop’, i.e.:

  • For every object in the cupboard:

    • Check that it says coffee on it

I always find the first few easy to comprehend, but for the last one, it’s amazing how frequently you have to think like that when it comes so naturally as humans.

🌟 Key lessons: be prepared to think computationally and it would probably help to go and do lesson one of Harvard’s CS50.

2. What to automate

One of the first things we tested was seeing how we ChatGPT could automate analysis of different adverts.

It was one of the four hour deep pieces of work that worked based on a series of triggers or prompts. Ask a question about an ad in Slack? Great it gets an AI response reviewing the ad. Have an ad automatically hit a certain spend threshold, great get it to analyse.

It was technically beautiful.

Except the actual raw prompt at the end wasn’t something we ever do manually. We use AI a lot but we don’t use it that much in analysis and find humans far better analysts of visuals at the moment. So why did I think automating a route would be better.

🌟 Key lessons: only automate a process you have already tested

3. You do need the time

I context switch 9 or 10 times per day. Most work I do fits into 15, 30, or 60 minute chunks of time. If it doesn’t fit into that I have to make time for it.

With AI, it takes 60 minutes just to get your head back into a problem.

Inevitably, it could take another 60 minutes until you figure out how to solve it.

Then you need another hour to test it, let it break, and go and rebuild it.

And then you’re at a point that it’s done.

And so, you really do need to set aside big chunks of time. And given how many meetings and usual day-to-day work people have, I imagine that’s impossible to properly do without building that time in.

🌟 Key lessons: you can’t do AI properly in 60 minutes. It needs days.

4. It doesn’t naturally suit my mindset

I am very much a “done is better than perfect” person who is “vision, not detail.” I’ve hired exceptional people who are the opposite to that, but this level of work does require going very deep and stuff works or it doesn’t.

I still think AI sits with leadership to properly learn and implement, at least for us, but I’d murder a technical co-founder right now.

🌟 Key lessons: it suits perfectionist, detailed-orientated people

5. I now know n8n & how APIs work better

I always knew sort of how APIs worked before but now I get them a lot better. I’d still like to do a bit of a crash course on them rather than learn by doing. And want to explore Postman a bit so I get really comfortable – maybe even via Terminal. But overall my knowledge is 0 to 1.

🌟 Key lessons: n8n & APIs are great

---

What we built

“This is all well and good Josh, but what did you build?” I hear you cry.

Ok. Here’s what we achieved.

Auto-generate simple statics using created backgrounds on nano-banana

GenAI comes down to great prompting. Great prompting is a skill that a handful of people have on the team but not every one else.

Workflow

1. Notion card where you drop a product link and a description of a scene

2. n8n takes that and sends it to ChatGPT

3. ChatGPT takes your input prompt and generates 10 variants of that prompt.

4. Each of those prompt get sent to nanobanana

5. Results come back to a thread on Slack and go to a Google Drive

Analyse recent creative winners

This was the ‘fail’ because we don’t use AI to do our visual analysis usually, but we got it working with a few inputs.

Workflow

1. Either (A) field updated on Notion, or (B) specific adname requested in an internal-client channel

2. Data is joined with other data from Notion related client insights

3. Claude prompt combines visual + insights and asks questions to analyse it

4. Result gets returned to Slack

Call actions shared to internal client Slacks

This never seems to work 100% but it does cover 90%.

Workflow

1. New Fathom call transcript

2. Transcript gets uploaded to Claude with a prompt about being marketer/exec assistant

3. Asks it to pull out all of the actions discussed on either side

4. Share those actions to client slack afterwards

High/Low Spend Alerts

We did get the non-AI version of this working. Still to build the next layer.

Workflow

1. Check Meta every hour across all clients to pull spend + conversions, then calculate CPA

2. Share the CPA update at a few key points to internal client Slacks

3. Check against a “danger” and “great” CPA range and if it’s good then either WhatsApp account owner / DM the account owner.

The bit here that’s not AI-ed is step 3, I’d love this to be relative to overall performance “recently” rather than set CPAs like this. If you’ve got ideas on how to build, let me know.

---

What comes next

We’ve got 20 unfinished AI initiatives and dreamt up plenty more. I’ve also been tinkering away more at weekends and evening. It’s addictive.

We will also likely bring on someone to support with AI as a focus at least on a freelance basis while we build in additional workflows. But ultimately it’s still something I’m investing a lot of time in.

How we implement this amongst the team, tbc. Lunch & Learns are definitely one thing, but ultimately it’ll probably be something we do at the next offsite and dedicate a four hour chunk to so people can understand the true power of it.

I still haven’t done any vibe coding, but that is coming next.

---

🔗 When you’re ready, here’s how Ballpoint can help you

→ Profitably grow paid social spend from £30k/m → £300k/m
→ Create full funnel, jobs to be done-focused creative: Meta, TikTok, YouTube
→ Improve your conversion rate with landing pages and fully managed CRO
→ Maximise LTV through strategic retention and CRM - not just sending out your emails

Email me – or visit Ballpoint to find out more.

---

❤️‍🔥 Subscribe to our Substack to learn how to grow yourself
… because agencies aren’t for everyone, but our mission is to help all exciting challenger brands succeed and so we give away learnings, advice, how-tos, and reflections on the industry every week here in Early Stage Growth.