👋 Greetings earthlings,

Emma here, co-founder of Spark AI. Last week we ran the first episode of our monthly Spark Sessions, hosted by Emma Jackson (EJ) – one of our Spark AI coaches, who works directly inside agencies on AI transformation. We kicked off by talking to Chris Murphy, Founder and MD of Tuncarp, about their AI operating system. Their AI production capability lets them deliver campaigns that simply weren't possible before – so we went behind the scenes to show you what they've built.

Our latest research found 83% of agency staff describe themselves as capable AI users, but only 15% are operating at the level where AI starts to change how the agency competes. Our programmes exist to close that gap - and we’ve set up our free monthly Spark Sessions to inspire you based on what others are doing brilliantly.

Plus – don't miss the bonus content from the Spark team: the SOP prompts we use ourselves. They'll save your team hours.

Quick links:

• Tuncarp’s creative AI operating system

• What can you do?

• Try this: SOP Builder

• Need to know: Watermarking AI content

• All the tool updates

Tuncarp’s creative AI operating system

When EJ sat down with Chris Murphy, founder and MD of Tuncarp*, he described how their challenge was working across multiple LLM platforms, jumping between tabs and losing track of versions (feel familiar?). So they built something to fix it.

The solution was a custom creative operating system that covers the full journey from brief to approved output. It sits as a layer across multiple LLMs, so rather than accessing each model through its own interface, everything runs through Tuncarp's system. On top of that, they have added functionality built around how their team actually works. A briefing agent asks questions when a brief comes in incomplete. Brand guidelines travel with every prompt. Models can be run side by side and compared without switching platforms. Creative directors and clients review and approve without work ever leaving the system.

*Tuncarp is a 50-person design outsourcing agency based in the Philippines led by founder and MD Chris Murphy. They work across network agencies and direct brand relationships.

Tuncarp’s approach to governance and data

Chris was clear that for larger clients, risk is currently the biggest blocker to AI adoption. To address that, every generation in their system is logged: the prompt, the model, the timestamp. If a client raises a concern months down the line, they can show exactly what was generated and how. The system also scans prompts before generation, checking for potential IP issues such as named characters or uploaded faces. It also handles watermarking in line with incoming EU regulations (more on this later). 

Tuncarp has a creative technologist with a legal background whose job includes staying across all of this. Not every agency can resource that, but the underlying principle is accessible to all: build a record of what you generate, know which models you are using and what their commercial terms are, and make sure someone owns the question of what happens when a client asks you to account for your outputs.

Watch the full webinar:

What can you do?

What makes Tuncarp's system work is the shared foundation behind it: organised, consistent, and accessible to the whole team. Building something like Tuncarp's OS takes time and resources, but the foundation it runs on is something you can start putting in place using tools you already have: map your typical workflow, identify where AI can play a role at each step, and build a custom GPT or Copilot agent to support each one.

That foundation is one of the most direct ways to turn individual-level AI use into embedded capability across the agency. It's also one of the first things buyers look for in due diligence, which we covered in our webinar with Rory Spence, Commercial Director of The WOW Company.

But mapping your workflow means you need to know what your workflow actually is. Many of the agencies we work with struggle at this step - if your processes aren't documented, you're leaving value on the table on both fronts.

Try this: SOP Builder

This will be most useful for the people who run how your agency operates day to day – MDs, COOs, ops leads, project managers. The operational foundations covered here are what determine whether AI becomes embedded capability or stays trapped in individual hands.

Start by asking three basic questions about your standard operating procedures:

1. Are your processes written down?

2. Are they consistent in format?

3. Could someone new to the team pick one up and follow it without asking questions?

If the answer to any of those is no, that is where to start.

What follows is the simplified version we use ourselves at Spark, built by the brilliant Chris D and Asta as part of our own internal foundations work. We split it into two steps and then an extra level of advancement to take it into agent territory. Feel free to use it as is, or adapt it, or take what's useful:

Step 1: Build your SOP ‘Master Template’

Open up ChatGPT, Copilot, Gemini (see update below) or Claude to build your template. We have developed a checklist of everything a simple but solid SOP should cover (see below), feel free to use it but please do add to it! Here’s the prompt to use based on our checklist to get you started:

Review the output, adjust any sections that do not fit how your business works, and save the final version as your SOP Master Template.

If you have brand guidelines, add this line to the end of the prompt:

Pro tip: When the assistant gives you its summary of the process before drafting, slow down and read it properly. Most SOP errors trace back to a misunderstood step right at the start, and they get harder to spot once they're buried inside a finished document. 1 minute of attention at the summary stage saves you a redraft later.

If you want to push it further, paste the finished SOP back into the chat and ask: "If a new starter on day one tried to follow this without asking anyone questions, where would they get stuck?" Make the assistant find the gaps before your team does.

Step 2: Set up your SOP assistant 

Create a Custom GPT, Claude Project or Gemini Gem or Copilot project.

These let you create a repeatable AI workspace with instructions and reference files, so the team does not have to start from scratch every time. Add your SOP Master Template to the knowledge section, alongside your brand guidelines. 

Then, paste these instructions into the GPT, Project or Gem:

Share this assistant with the team, so that anyone can paste rough notes and get back a completed SOP in the same format, every time. Save each approved SOP in one shared folder, with a clear file name, owner, version number and review date.

Use a file name like:

SOP_[Team]_[Process Name]_v1.0_[YYYY-MM-DD]

If you already have a standardised SOP:

Use your LLM to audit it first. Upload your existing document alongside the checklist below and ask it to identify gaps and flag anything that could be clearer or more consistent. Then ask it to adapt our prompt to your own standard.

Once you are happy with it, follow Step 2 instructions above.

Next level: Connect it to your tools (turn your assistant into an agent)

A Project or Custom GPT is helpful, but you're still copy-pasting inputs in and outputs out. The next step is connecting your assistant to the tools where the work actually happens, so it can read the inputs and save the outputs without you in the middle.

This is the shift from assistant to agent. Same brain, but now with hands.

A practical starting point: connect your assistant to your meeting transcript tool (Fireflies, Otter, Granola) and your shared drive (Google Drive, OneDrive, Notion). Then update your prompt with something like:

How to set this up depends on your stack:

• Claude with MCP connectors (Drive, Slack, Fireflies, Notion)

• ChatGPT with Connectors and Agents

• Copilot Studio if you're on Microsoft 365

• Make or n8n if you want a platform-agnostic flow

Don't try to automate the whole SOP library on day one! Pick one trigger (a new transcript landing in a folder), one action (draft an SOP and save it for review), and one notification (a message to the process owner). Get that working, then add to it.

The principle: your team's job is to think and decide. The agent's job is to capture, structure, file, and remind.

Get this foundation in place once and you build on it every time. The bonus: you'll have something to point at the next time a buyer asks about your processes, or you start a serious conversation about AI capability.

Need to know: Watermarking AI content

When EJ raised regulation during the session, Chris reminded everyone to keep an eye out for the new regulations coming into place. From August 2026, Therequires that all AI-generated images, video and audio are marked to identify them as artificially generated. These rules apply regardless of where the agency is based, as long as the output is intended for the EU market.

The technical watermarking sits with the AI tool providers. The platforms your team already uses, Adobe, or other major AI image generation tools, are responsible for embedding provenance information directly into files. What the regulation requires is a layered approach: metadata embedded in the file, an invisible watermark woven into the image at pixel level, and a logging trail so content can be traced back to its AI origin if other marks are removed or degraded.

Agencies are not responsible for building that infrastructure. But if you are producing AI-generated work for clients, you need a clear process for disclosing it. Who checks that content is properly marked before it leaves the building? What is the client told, and how? What happens if a piece of work is queried months down the line? These are not complicated questions, but most agencies do not yet have documented answers to them.

Getting that in place before August is straightforward. It is a much easier conversation to have now than when a client asks first.

All the tool updates

Adobe Firefly AI Assistant is now in public beta. Until now, using Adobe's creative tools meant navigating between applications. Now, Firefly AI Assistant sits across the whole suite as a single conversational interface. You describe the outcome you want, and it coordinates the steps across whichever tools are needed to get there, without you switching between them. It also learns your preferences over time, including your visual style, your most used tools and your recurring workflows, so results become more consistent the more you use it. For agencies already working inside Creative Cloud, this is worth trying. It is available now on Creative Cloud Pro and paid Firefly plans.

Read more about Adobe here

Sora has been discontinued. OpenAI shut down the Sora web and app on 26 April. For most agencies it was never a primary production tool, but some teams had built it into video workflows. The short reason: user numbers collapsed while the app was very expensive to run. OpenAI redirected the compute to its coding and enterprise products instead. The API remains available until September, so there is time to migrate, but it is worth reviewing now rather than scrambling later.

Read more about Sora here 

Gemini can now create files directly from chat. The useful shift is that Gemini is moving from giving you content to giving you finished files. Instead of copying a response into a document or spreadsheet, you can ask it to create a Google Doc, Sheet, PDF, CSV or Excel file from your prompt - just like you can in Claude Cowork, Copilot and ChatGPT already. That means meeting notes can become a document and project notes can become a PDF summary with less copy / pasting and manual formatting. They will still require you to check the content, formulas, formatting and source accuracy before using it of course (Remember our mantra of "think AI think"!). This is being rolled out globally and across different account levels - if you haven’t got it yet you’ll see it springing up in the next few months.

Read more about Gemini here

I hope EJ and Chris's conversation inspired you! Keep an eye out for next month’s webinar - we’ll be promoting it soon 👀.

As you know, we help agencies build AI operating systems for your own teams through our AI Accelerator. Give me a shout if this sounds interesting. 

And as always, if something in this edition raises a question you would like to talk through, please reply to this email. 

See you next time!

Emma

Co-founder, Spark AI

About Spark AI

Spark AI helps you lead your team through the biggest shift since digital, through AI training, transformation and workflows. We've worked with 70+ agencies, published the #1 bestselling book on AI for Agencies, and teach at Oxford University.

👉 Find out more 

Not a subscriber yet?