👋 Greetings {{first_name|earthlings}},
Jules here, co-founder of Spark AI, taking over this edition’s newsletter.
Spend ten minutes on LinkedIn and you could be forgiven for thinking half the agency world now has a team of AI agents running the company while the founders manage it from a sun lounger.
They don’t. Or at least, the sensible ones don’t.
Most of the businesses who have taken our AI Maturity Score™ sit at Stage 1 or 2 out of 4. The AI agents in this issue are the bridge from 2 to 3 – they are narrow in function, usually internal, and low-risk. If you’re not sure what stage you’re at the Spark AI Maturity Score™ takes five minutes. And if you took it last time, your results are further down – as promised.
Right now, AI agents are no substitute for people thinking, judging, building relationships and making creative calls. Humans still need to be at the centre of the work humans do best. But when pointed at the right tasks, agents are powerful tools for taking on time-consuming repetitive work.
This week, I’m sharing practical examples and guidance around how agents can work inside an agency.
This edition draws on a recent Spark Session I ran. If you would rather watch than read, you can watch the full video here. As a bonus for newsletter readers, I’ve included a competitor-watch checklist you can download and use (see section 9). It covers the framing, briefing and reporting you need to think about before you build it.
Here's what's coming up:
What is an agent?
Before we look at where to deploy agents, let’s go back a step. You may be wondering - what actually is an agent?
Most teams have already used an AI assistant. You might know it as a custom GPT, or a Gemini Gem. Microsoft calls its version a Copilot agent, which is confusing, because it isn't one. An assistant responds when you give it a prompt, and uses the instructions and information you provide to help with a specific task. It waits for a person to start the conversation and replies back in the chat.
An assistant has three parts.
Instructions - The job you want it to do.
Knowledge – The information it needs to do that job your way, including templates, SOPs, examples, client context and previous work.
Standards – A clear definition of a good output, ideally supported by strong and weak examples.
An agent adds three more steps.
A trigger – The event that starts the work. It might run every morning, or when a project status changes.
Connections – Access to the systems where the information lives, such as your inbox, CRM, project management platform, call transcripts or shared drive.
An output – Something useful created outside the chat. That could be a draft email, a research document, a task added for review or a report delivered to your inbox.
So an agent is a narrowly defined job that runs on a schedule or a trigger, reaches into your systems and leaves the result somewhere useful. If you want to read about the six steps of building an agent then check out this blog by my colleague Asta.
Practical agent use cases
Your clients are paying for your team’s judgement and experience. Automating client work risks weakening the reason they chose your agency in the first place. Instead, point your agents at the internal work that helps you get to a great answer yourself.
First, create a morning briefing:
This is a great place to start, Every day at a set time, an agent can read your calendar, emails, call transcripts, CRM and project management system. It can produce a short summary showing what is happening that day, which deals need attention and which actions are waiting on somebody else. It saves the first half-hour of the day spent piecing things together. You start with a clear view of the priorities.
Next, set up a client research agent:
Once you’re comfortable with the basic principles, you can tackle something more ambitious. When a client brief hits a folder, kick off parallel research across the 4 Cs: customer, competitor, culture. Have the responses dropped into a new document, saved in the client folder and emailed to you when it’s ready. This kind of thing is fairly straightforward to set up in Google Workspace Studio or Copilot Studio.
Then, think about where to go next.
Guidelines for deploying agents:
Give your agents read only access at first, and keep its outputs as drafts. Choose tasks that are easy to reverse and make sure somebody remains responsible for checking the result.
Read access, drafts only, someone accountable – that's the Governance & Accountability pillar of the Spark AI Maturity Model in practice. It's one of the areas people score lowest on, and it decides whether agents like these stick or get switched off, or more likely ignored and forgotten after a fortnight. Check your stage before you build.
The copywriting agency that kept the copy human
The temptation is to point AI at the thing you find hardest to do, but I urge you not to. We see this in our workshops all the time. The first thing a brand strategist will try to do with AI is help it come up with strategic routes for a brand they're working on. But that's exactly why your client is paying you - because it’s hard! That's the bit that has to stay with you. Instead, point AI at all the things that help you get there or hold you back.
A great example is a copywriting agency we know. They now get 95% of client copy signed off first time. And their rule is simple: every piece of client copy is still written by a human. But, the agency has pointed AI agents at everything around the writing.
When a brief arrives, an agent checks it against the agency’s briefing template and examples of strong previous briefs. It spots the gaps and asks the client for the missing information. Then, once the brief is complete, one research agent gathers relevant external material. Another searches the agency’s existing client knowledge for useful information about the brand and audience.
So by the time the time the copywriter gets stuck in they already have a complete brief and a prepared research pack, so they can get on with what they do best: writing.
Then when the first draft is ready, another agent checks it against the brief, the client’s guidelines and previously approved work. It returns a first critique within minutes. No more waiting around for client feedback. The writer stays in control and can decide what feedback to act on.
All this means is they can turn around projects faster than ever, while achieving 95% first time approvals from their clients. Not only that, their copywriting team enjoys their work more. They're presented with all the information they need to get started straight away, so they spend more of their day actually writing.
So, what's the takeaway? Don't use AI to do your job. Instead, use AI to help you do a better job.
Think – AI –Think
There is one more reason people need to stay at the centre. Humans experience something called automation bias. We tend to trust a computer’s answer because it looks polished, even when it is wrong.
At Spark AI we teach our clients a simple motto: Think – AI – Think.
Decide what good looks like before the agent starts. Let it perform the agreed work. Then apply your own judgement and take responsibility for what happens next.
One thing to try this week: build a competitor-watch agent
Competitor-watching is an essential part of life for all businesses. This agents helps you monitor known competitors and scout for new ones. The agent runs every week, and adds anything new to a dated running document, so it never repeats itself.
It can take time and a fair few iterations to give it the right context so it can perform really well,. Put some time aside to try it this week. Once it is working well, it replaces repeated manual checking and helps you spot market changes earlier.
Use the checklist below as your outline. You can also paste the checklist and your company context into your LLM, then ask it to turn them into clear instructions for you to review.

Download the checklist as a PDF →
Want to see these agents in action?
Watch the latest episode of Spark Sessions where I build five examples of agents live using Claude, ChatGPT and Gemini.
Where the first batch of AI Maturity Score results have landed so far
Last issue Emma promised to show you where the numbers are so far on how AI-mature we all are. we're running our AI Maturity Score™ tool for the whole of this month, so if you haven't got your score in yet, have a go here and see how you compare 👀. Here's the picture from the first fortnight of responses.
Nearly nine in ten responses are at Stage 1 or 2 – lots of AI activity, not yet changing how the business runs. Which is exactly who this issue is for: simple internal agents are how you get from 2 to 3.
Nearly eight in ten of you already save three or more hours a week with AI. Yet nearly everyone is still at Stage 1 or 2. The time is being saved. It just isn't being turned into anything the business can feel yet. That's the gap the Score is built to close – and it's why this issue is about agents that work on the business, not in it.
Free · 12 questions · 5 minutes · report emailed instantly
🛠️ All the tool updates
ChatGPT had an agentic fortnight
The biggest news is GPT-6 Astra, OpenAI’s answer to Claude Fable 5.1. It is built for long, multi-step professional work and is particularly strong at automation and producing documents. The rollout has begun with a limited number of organisations.
ChatGPT Work tasks can now start when something happens in a connected service, such as a Gmail message arriving. This means a task can begin when new work arrives, rather than waiting for a set time.
ChatGPT Work can also complete tasks on websites you are signed into. Your login details remain hidden from the model, and it asks for confirmation before taking actions with greater consequences.
Finally, ChatGPT Sites can now be shared privately with named people outside your organisation. That makes them more useful for client-facing tools and reporting portals that you do not want to make public.
Ask Gemini has moved into Google Chat
Google started rolling out Ask Gemini inside Google Chat on 26 August. It can search Gmail, Drive and Calendar and help manage tasks or meetings without leaving Chat.
A useful agency test would be a busy project space. Ask it to identify the outstanding actions and files requiring attention, then give the result to somebody to review.
One thing to know before you switch over – previous conversations from the Gemini side panel will not move across, and Gems will not be available inside the Chat side panel.
The rollout covers eligible Business, Enterprise and Education plans.
10x the context. Half the time.
Speak your prompts into ChatGPT or Claude and get detailed, paste-ready input that actually gives you useful output. Wispr Flow captures what you'd cut when typing. Free on Mac, Windows, and iPhone.
Want to build useful agents with your team?
We’ve launched the Agent Build Sprint, a two-day programme that gives your team the skills to build and use agents inside your business.
On day one, we help you choose the right internal tasks, map how they should work and build the first versions around your own systems and standards. Your team then tests them on live work for a week.
When we return, we fix any weak spots, put the right controls in place and document how everything works.
You leave with two or three useful agents, plus the skills to build and improve more yourselves.
Interested? Hit reply and tell me which internal task is taking up too much of your team’s time.
Not sure which stage you're at? Get your AI Maturity Score – five minutes, report in your inbox.
That’s all for today, see you next time.
Jules
Co-founder, Spark AI
What did you think of our email today?
About Spark AI
Spark AI is the AI performance partner that builds organisation-wide AI capability for agencies and brand teams. How? Through role-specific AI training, leadership coaching, AI consultancy and agent building.
We’ve worked with 80+ agencies, published the #1 bestselling book on AI for agencies and publish the AI in agencies benchmark, and teach AI Strategy at Oxford University.






