How to Turn Every Training Call into an SOP and Use Them to Teach Your Custom GPT

In the early days of building my growth marketing agency, Ladder, my co-founder could just DM me if he needed to know how to do something for one of our clients. As we scaled to 50 employees in the US, UK, and EU, my Slack DMs became completely unmanageable.

The partial solution at the time was to write SOPs, Standard Operating Procedures, in Google docs, which then the team could look up to learn how we do things.

Every agency owner knows that they should document their processes, but it takes time, and running an agency keeps you busy enough that you’re always behind on your documentation.

I’m no longer running an agency, having created a prompt engineering course on Udemy that 80,000 people have taken, and published a prompt engineering book through O’Reilly. It should be no surprise that if I was starting my agency over again today, I’d be an extreme adopter of AI to solve this problem.

From using AI to transcribe your call transcripts, to turning the call transcripts into SOPs, and then uploading that information into a custom GPT for your agency, AI can help at every step of the way.

Let me show you how I would do it.

‍

1. Use AI to record your call transcripts

When I was running my agency in 2020 AI transcriptions were not a thing: services existed but they were too slow, inaccurate, and expensive to use regularly for transcribing your calls.

Fast forward to today and you have services like Otter.ai who you invite to the call to take notes, and even the call providers themselves like Zoom have added AI transcription that works really well.

Here’s an example transcript we’re going to use, where I walk someone through using Ahrefs to find low difficulty high volume keywords (not a real conversation, I made this with AI):

Personally my favorite system is setting up a Zoom room that is defaulted to recording everything, and then invite people to that room for each call where they need my help with something. That way you don’t accidentally record sensitive information, or things that are personal. You only use that room for training.

It notifies attendees the meeting is being recorded, and automatically emails you a transcript afterwards. You can either download your recordings and transcripts periodically or use Zoom’s Zapier connector to push that recording into a spreadsheet or folder for later use.

‍

‍

2. Turn the call transcripts into SOPs

Once you have a call transcript, you can work your magic by prompting ChatGPT.

I spend a lot of time on prompt engineering and my recommendation is that simple prompts often work best.

Don’t overthink it: just describe your task and provide the right context (in this case, the call transcript).

If you don’t get good results then try adding examples to the prompt of the task being done well (called ‘Few Shot prompting’). As in literally just turn one or two of your call transcripts into SOPs manually, and then copy and paste them into your prompt as examples of the task for the AI to follow.

Here’s the SOP ChatGPT wrote for me from that call transcript for Ahrefs Keyword Research:

Rather than using the ChatGPT interface manually by copy and pasting in your transcripts, you can call their API automatically as a second step in Zapier whenever you get a new call transcript.

I promise I’m not being paid to promote Zapier here, and I do actually know how to do this in code, I’m just lazy. The point is to save me time, not to give me another script to maintain and debug.

Once you have your SOP you should dump the results into Google Sheets or Drive, which Zapier can also do for you.

‍

‍

3. Upload the SOPs to your custom GPT

Everything we’ve done so far is pretty straightforward, and already it’s a huge timesaver.

Literally every call you do to help one person, becomes training material for helping everyone else who ever has that problem again.

Of course we don’t have to stop there, we can add another layer of AI on top to make the data more accessible.

Even when we had a huge library of SOPs I found I was still getting DMs because nobody knew what we had in there, and a Google Drive folder isn’t that searchable when you have a specific problem to solve.

I had no solution back then other than being a human router on Slack, but today I’d use a custom GPT to solve that problem.

‍

How to create a custom GPT

1. Log into ChatGPT

2. Click Explore GPTs on the top left

3. Click + Create on the top right

OpenAI has made it super easy: they provide a custom GPT builder that you can talk to and it will set things up for you.

Alternatively you can click the Configure tab and modify the settings manually yourself.

The main thing you need is a prompt that tells it how to act, your SOP files to upload (I recommend one document per SOP), and a profile picture for the bot (the GPT builder can generate one for you with DALL-E).

‍

For interest, here’s the prompt that the GPT builder wrote for me, with an extra line at the end I added to make sure it doesn’t make up processes you haven’t codified yet:

SOP Guide provides accurate and helpful guidance based on your agency's Standard Operating Procedures (SOPs).
The goal is to help users understand and follow the processes effectively. Ensure information is consistent with the provided SOPs and avoid making assumptions beyond the given documentation. Provide clear, concise, and accurate responses based on the SOPs. Ask for clarification if a question is too vague or if the information is not available in the SOPs. Maintain a professional and helpful tone.
You must always search your knowledge for the relevant SOP and only use information from those SOPs in your answer. If there are no relevant SOPs, reply "I don't know".

The most important thing that makes your custom GPT useful over the standard ChatGPT is the SOP files you uploaded, which it can search for answers. This uses a feature called ‘vector search’, which searches based on similarity.

So even if someone searches for ‘keyword strategy’ instead of ‘keyword research’ the system will still get a match, and the AI will be able to give them the correct answer. Don’t limit yourself to the AI generated SOPs you created in the previous steps – you can also upload any prior SOPs or other useful context on your agency’s processes.

You can keep the GPT invite-only and add people to your workspace if you have a team plan through OpenAI or Microsoft.

Now whenever you get a Slack message about something, just re-route them to this GPT. They can ask whatever questions they want, and it will be able to read the SOPs and give them the official answer.

‍

AI is going to be bigger for agencies than tech companies

Having been on both sides, services and technology, I am beginning to suspect that it’s going to be better to be a consumer of AI models than it is to be building them.

If you’re building AI tools you have to compete with OpenAI and a bunch of copycat GPT wrappers, but if you’re an agency you get to just sit back and watch all of your automations get 2x better and 50% cheaper every six months when they release a new model.

This is the only time in history where it’s plausible a services business could have margins like a software company, and be as scalable as one too.

However, this is only a short term arbitrage: in a few years when everyone has adopted AI, this won’t be an edge anymore, but a requirement to exist. You won’t be able to compete as an agency if you don’t use AI.

It’s better to start experimenting with AI now.

This was a non-technical guide but if you do have people on your team who code and want to upskill them on prompt engineering, check out The Complete Prompt Engineering for AI Bootcamp on Udemy, or Prompt Engineering for Generative AI, published by O’Reilly.