This is an article about creating an AI marketing copilot. That’s a bad first sentence from a “grabbiness” standpoint — it’s not a solid literary first impression — as opposed to this second sentence which is intentionally meandering and, sure, run-on, but has a certain musicality to it, I hope, that announces itself as written by and intended for flesh and blood. The first sentence, by contrast, is speaking to machines. It’s a small element of a larger SEO strategy (along with the article’s title, meta description, H1, and URL slug) all meant to clearly signal to search engines what this article is about: creating an AI marketing copilot.
It begs the question of who we’re writing for, at least in my world of corporate content marketing. Generative AI gives “people who hate writing” (as well as people like me who love reading and writing), the option to have machines write stuff intended for other machines to read and rank, all in service of generating web traffic and driving qualified leads. But, at the end of the day, they’re not leads. They’re people. Flesh and blood.
So, despite the fact that my job entails creating B2B content about our Data & AI Consulting Services, I’d initially maintained a certain anger (water is life!) and self-righteousness about GenAI and automation tools. Fine to use for a quick fix or a LinkedIn post, but no substitute for the human touch, especially when absolutely every single ChatGPT output opens with “In the dynamic landscape of…” (seriously, google that phrase, it’s everywhere, and it’s an immediate ‘tell’). ChatGPT generally has a hard time finding an intro “hook,” and is lazy AF.
Why Marketers Resist AI, and Why It’s Worth Getting Over It
Many of the well-publicized marketing use cases for GenAI – summarizing text, writing first drafts of articles, generating imagery – veer heavily into my team’s lane. Many other job roles feel equally annoyed or threatened by these tools: in 2024, 45% of workers expressed worry about being replaced by AI, with higher-educated and higher-paid workers more exposed to AI impact (the most threatened professions include writers/content creators, photographers/graphic designers, software developers, and legal professionals).
Nevertheless, I’d spent the better part of two years writing about how many rote, repetitive tasks could be offloaded to AI, with greater velocity, to the extent that AI can serve as a trained thought partner and assistant. So, I bit the bullet and dove into developing a custom AI copilot for my Marketing team, to see if combining features and capabilities would help me learn to stop worrying and love the robots.
While many of the copilot capabilities I’m about to share are possible to create on individual LLM platforms – ChatGPT, Claude, or Google Gemini, for instance – I’m going to show you how to do this stuff within TELUS Digital’s proprietary enterprise AI platform, Fuel iX. Specifically, I’ll dive into Fuel EX conversational AI assistants designed to help employees with everyday tasks.
Here goes.
Understanding Enterprise AI Platforms Like Fuel EX
For marketers, the reasons to use an enterprise platform like Fuel EX are primarily:
- Democratizing AI: Everyone on your team, regardless of technical skill level, can now easily craft custom copilots for any task or domain.
- Specifying knowledge: You can direct your copilots to utilize specific data (i.e., your documents uploaded to the Knowledge Base), which become easily searchable via retrieval augmented generation (RAG).
- Boosting collaboration: Once you create something that works for your job function, you can easily and selectively share copilots with your team.
- Combining tools: You can swap between advanced conversational assistants supporting a wide range of LLM models for text and image generation, analysis, summarization, etc., all in one centralized interface.
Add to that the following enterprise-wide benefits to your boss and your boss’s boss:
Flexibility: Maybe you like using ChatGPT for some tasks, but prefer Gemini or Llama for others? Or maybe you’re concerned that today’s gold standard LLM might get leapfrogged by some hot new thing three months from now? This space is moving fast, after all. It’s a dynamic landscape.
- With Fuel, you can easily switch between different LLMs to avoid vendor lock-in, without having to pay individual subscription fees or keep up with the latest releases on each platform.
Trust: Maybe your company doesn’t officially allow public LLMs for work tasks, due to the very real concern that you might leak proprietary/private data to our AI overlords? But maybe you use it on the side anyway, because it’s helpful, and who’d know?
This is a massive problem. In our recent webinar “The Executive’s AI Playbook,” AI expert Allie K. Miller warned, “As a company, you have to be thinking about security vulnerabilities. Every single executive is missing the shadow use of AI inside your organization. Over half the employees at your company are using AI that is not approved by your IT team at all!”
- With Fuel, you’re working in a closed, private system, so your data isn’t being shared publicly or used to train these LLMs. The user prompts and responses are also visible only to the individual user. We’ve already implemented Fuel with 50,000+ enterprise users for these very reasons.
Control: Maybe the higher-ups don’t want thousands of employees using multiple AI tools because it’s just too darn hard to track who’s using what and how much it all costs?
- With Fuel, multiple GenAI platforms can be looped into a single control panel where administrators can set user privileges, monitor usage, and control costs.
Price: Speaking of costs, maybe it’s just too expensive to approve every employee signing up for a monthly or yearly subscription to their LLM or GenAI image generator of choice, regardless of how much these tools get used?
- Fuel EX uses usage-based pricing, so enterprises only pay for what they actually use, making it more cost-efficient than paying per user. (TLDR: It’s typically cheaper).
For all these reasons, and because Fuel EX is my company's approved enterprise platform, I opted to set up my first AI copilot on this platform.
Let’s discuss how to set one up and test it out in a limited capacity before launching copilots across your department or broader organization.
Part 1: Setting up an AI Copilot
I created an enterprise Content Marketing Copilot that I now use regularly for work, but there's a lot of proprietary/private company data in there that I didn’t want to share with the world or heavily redact in this article and video. So, instead, I’m going to walk through a similar marketing workflow using a separate copilot I created as a proxy — this one for my personal homesteading/DIY site, Thunderbird Disco Homestead — because I don’t much care if you see that data and nobody besides me is gonna get mad about it.
While this is obviously a smaller-scale use case, the roles I ultimately ask this copilot to play and tasks I ask it to perform mimic how I use my Content Marketing Copilot for larger-scale, enterprise use cases: I basically treat it as a trusted mentor combined with an overqualified intern I don’t feel bad pushing around.
And while both these copilots are geared towards Marketing use cases, similar opportunities abound for Business Development, Engineering, Project Management, HR/Talent, Legal, or any other functional task. I invite you to take some imaginative leaps and mentally substitute in your own work streams, pain points, or the unique software platforms you use, and consider how you might set up your own copilot for individual contributions or team-based work.
In the videos below, I’ll share the quick-and-easy way I set up a copilot in minutes, and my colleagues Andrew Carter (Engineering Practice Advisor) and Zach Richardson (Associate Director, Solutions Architecture) provide some color commentary into the deeper tech of these tools (i.e., if you hate reading, just watch the videos).
Basic Settings
When you select “Create your own copilot” the Name, Description, and Photo fields are self-explanatory window dressing. The System Prompt, however, really matters.
If you’ve used GenAI tools, you’re likely familiar with system prompts, so i won’t spend too much time on this, but in brief: effective system prompts help AI copilots produce relevant and well-informed responses. They save time, ensuring you don’t have to provide the same foundational information and instruction to your assistant every time you start a new chat or ask a new question. It’s like providing a clear, introductory set of Roles, Responsibilities & Guidelines to your fancy new intern when they’re onboarded.
A great system prompt contains the following building blocks:
- Persona: Give your AI copilot a specific character or role to adopt when responding, to set the tone, perspective, and baseline expertise it’ll aim for in its responses.
- Context: Provide the most basic background information or circumstances you want the copilot to understand on a foundational level.
- Task/Steps: Lay out the basics of what you’ll want this copilot to do for you, and how you want it done. Specifically…
- Limits/Guidelines: If desired, set some boundaries (e.g., the length, depth, or complexity of its responses; legal/regulatory compliance guardrails, etc.). I gave mine no limits because I had no particular reason to add constraints for this smaller-scale use case, but also because I’m a rule-breaker, I’m limitless, I’m a loner, Dottie, a rebel.
- Format: Finally, guide your copilot to respond with a certain voice, tone, format, or style. Think in terms of vibe or any other parameters of the output you’re looking for (e.g., formal vs. casual tone, paragraphs vs. bullet points, etc.).
This is actually a relatively short, tame system prompt. You can go nuts and provide pages worth of baseline instruction, but the building blocks above remain.
Advanced Settings
Next, toggle on or off a wide variety of specialized features that your copilot can use in its upcoming tasks. We’re constantly adding features to Fuel EX (new options for Generative UI and Code Execution were recently added since the video was recorded).
You can simply toggle everything on (hey, why not?), but each feature impacts response time, and in some cases, a feature might be counterproductive – if you want your copilot to ONLY reference specific data or content you provide, for instance, you might not want it to Search Internet, for instance (that’s why not!).
I won’t go in-depth on every toggle here, but the big unlock for me was using the Image Analysis capability. Turning on this “Vision” component allows my copilot to view, understand, and analyze images – not merely photos or pictures, but also graphs, charts, dashboards, desktop screengrabs, and other complex visual content.
You can also add a few “Conversation Starters” with the most typical tasks you’ll ask your copilot to perform (for instance, in my Content Marketing Copilot, I have one for “Create a LinkedIn post (of <1500 characters) summarizing/sharing this article…”)
Knowledge Base
This is where it got really interesting for me. With “Knowledge Base,” you can upload files (documents, images, PDFs, data sets, etc.) to ensure your copilot has not only the world knowledge it’s been trained on, plus access to more recent information from the internet (if desired), but also the specialized information and context needed to assist you, personally. What are the manuals, handbooks, guidelines, histories, or other core “bibles” specific to your job role that you want this copilot to have at its disposal?
Think of this as your copilot’s long-term memory, which it can continually reference and draw upon. For my Thunderbird Disco copilot, I uploaded my novel and a longform Rolling Stone article (for tone/style but also for particular topics I plan to reference in future content creation). For my Content Marketing Copilot, I added our brand guidelines, and sets of high-performing articles (to use as examples for structure, style, etc.). You could consider uploading employee/HR handbooks, legal documents, company histories, code requirements, product/process roadmaps, or anything else specific to your org or job task that the copilot should consider gospel.
You can add or subtract from the Knowledge Base any time, so here’s a huge hack: cut/paste especially relevant copilot chats into a .doc or .pdf file and upload these into your Knowledge Base, transforming these short-term conversations (that are otherwise not retained) into long-term memory. More on this later…
Access & Sharing
Finally, with Fuel EX you can decide whether to keep a given copilot private (for personal use only) vs. restricted (to share with an internal team) vs. public (to share with your entire organization). My “Thunderbird Disco” copilot is Private, for instance, while my Content Marketing Copilot is accessible to everyone on the WillowTree Marketing Team.
Now that your copilot is set up, let’s walk through a multi-faceted use case for a content marketer like me (again, consider how you might apply or customize this workflow for any other job role, including BD, engineering, project management, legal/HR, talent acquisition, etc.).
Part 2: Using Your AI Copilot for Complex Tasks & Roles
Rather than asking my copilot to perform one discrete task over and over again (which is indeed a valid and typical use case) and limiting my requests to the expertise stipulated in the system prompt, I treated my copilot as a skilled partner with broad, cross-functional expertise.
I asked it to shift back and forth between a variety of different “roles” within and outside of a Marketing function, including:
- Data Analyst
- Strategist
- Developer/Coder
- Customer Service Rep
- Editorial Assistant
- AI Guide
Let’s get into it.
Data Analyst
At the start of this chat, I want my copilot to understand how the site has been performing over time. Rather than trying to explain this via words/text, I simply upload analytics data dashboards from my CMS and Google Analytics 4 (GA4), plus affiliate marketing data from third-party partners.
I’m simply capturing screenshots of what I’m looking at, or exporting/uploading PDFs. Super easy.
The Fuel EX Image Analysis tool can read these images/files and in a matter of seconds gain an accurate picture of:
- Traffic trends over time
- Top-performing content
- Audience demographics
- Traffic sources
- Engagement metrics (e.g., time on page, bounce rate)
- Conversion rates
I’m not trained in data analysis — I know just enough about analytics to be dangerous — but now my copilot has a solid historical view of all of the above and can cross-reference and marry information from disparate sources to provide me with expert actionable insights.
Again, this Marketing use case could easily be adapted to other job functions. My colleague, Andrew Carter described how he, as a software developer, uses Image Analysis to ingest multiple windows simultaneously — a Jira instance, a Read Me txt file, a command line tool, and whatever other windows he had on his screen — and recommend solutions for complex problems.
This is akin to the difference between telephoning that wise mentor and verbally describing complex data from multiple sources vs. asking that mentor to simply come look over your shoulder. Much faster and easier.
Strategist
Now that my copilot has a solid sense of site performance to date, I’m going to shift gears and essentially ask it to serve as a strategist, generating a 12-month content marketing strategy for growing my audience, engagement, and conversion rates. I then ask it to drill down on the next quarter, providing a detailed project management plan for executing the marketing campaigns it’s laid out for me.
It's often assumed that GenAI tools excel at strategizing, but a quick word on Fuel EX’s Advanced Settings: Zach Richardson explained how copilots use the “Reasoning” function to do more advanced “thinking” based on the data and context provided, leading to more customized strategies.
Rather than “chain-of-thought” (CoT) prompting (which is itself an “advanced” approach that mimics human thought, breaking down a complex problem into manageable steps, one at a time), Fuel instead can run a “planner-executor model” (which first generates a more holistic plan and then carries out each step accordingly; this approach is the basis for the next-gen automation and “agentic AI” capabilities our strategists and Data & AI Research Team (DART) are exploring).
Tactician (Developer/Coder)
Okay, so now I’ve got my comprehensive 12-month plan, including a number of low-hanging, short-term tactics to help improve my site’s SEO and performance. One early step recommended was to add a particular piece of content to every page — an author bio in the footer — which requires some custom HTML coding.
I am not an HTML developer — I do not know enough to even be remotely dangerous. Yes, if this were an advanced enterprise use case, I’d reach out to a colleague who does custom HTML. But this is a relatively basic task and I’m also a picky SOB who’ll want to tweak the format/style/photo cropping therein, so I’m going to see if I can do it myself and streamline the process, with a little help from my copilot.
A lot of MarTech software, like Adobe Experience Cloud or Braze, is intended to empower marketers in a similar way. The platforms allow us more control over how content is customized and displayed without bottlenecking the process or requiring extensive collaboration with engineers. Here, I’m using the copilot as an intermediate step for a technical task that’s beyond my abilities but not necessarily worthy of adding developer hours or hiring a freelancer.
We know LLMs are pretty good at producing code, though they’re not always perfect or elegant with it (what’s the engineering equivalent of “In the dynamic landscape…”?). So when something goes slightly awry, for a non-engineer like myself, I quickly get confused and frustrated. I do a quick Google search for my problem, but the only answers I find are years-old responses based on a prior version of my CMS. So…
Customer Service Rep
Rather than throwing up my hands and bugging someone in engineering to give me a tutorial or fix it for me — and certainly rather than phoning my CMS’s customer service line or navigating its old-school chatbot — I ask my copilot for help.
I again use Fuel’s Image Analysis tool to get myself unstuck: I can simply upload a screengrab of what I’m looking at and be like, “Here’s what I see. Tell me what to do next.” The copilot sees what I see, understands the problem, and gives me step-by-step instructions to unstick myself in real time.
I was trying to navigate a particular CMS to inject HTML code, but you can substitute in any software or platform you’re trying to navigate (e.g., the intricacies of Salesforce, Jira, Adobe Experience Manager, Final Cut Pro, etc.). I see these powerful Vision tools having massive implications for the future of customer service call centers and CX self-service.
Creative Thought Partner
So now I’ve got the HTML working, such that my footer content appears correctly, the image is cropped how I want it, all the formatting and linking are on point, etc. At this point, I asked my copilot for some editorial advice (“Should my bio be written in first- or third-person POV, considering the rest of the site is written from first-person?”).
Writing/editing is something I DO feel comfortable handling myself — “Danger” is my middle name — but I simply wanted a different perspective.
This is another way to view a copilot: as a peer or sounding board in an area you feel abundantly comfortable. Writing can be a solitary enterprise, and as my reticence and general egotistical territorialism started melting away, I felt more comfortable using the AI assistant as a thought partner.
We veered further into divergent creativity, and I became more interested in the nature of this AI copilot. What could I do to make these interactions more productive and satisfying over time?
Part 3: Optimizing Your AI Copilot for Repeat Use
Some Questions I Asked My Copilot to Guide Usage
I found myself asking the copilot questions about its abilities and limitations, ultimately looking for suggestions on how to improve my use of it in the future. For instance…
If I have multiple AI copilots in Fuel EX, do they cross-reference and retain context from each other?
- No. Each interaction and copilot is independent, which is typically desirable. A user can maintain separate spaces for personal vs. professional experimentation, for various enterprise workstreams, for different tones of voice, for distinct knowledge bases, etc.
Does a single copilot retain context from chat to chat?
- No. Each interaction starts fresh and there’s no “persistent memory” across chats. Each chat session is like meeting a stranger who understands their role/purpose, has read my site, and studied the documents I’ve uploaded, but otherwise doesn’t remember any previous conversations we’ve had.
Okay, what if I want the copilot to retain information from chat to chat? We’ve covered a ton of ground today, how do I help the copilot remember this chat? Can I upload this chat into the Knowledge Base?
- Sure, you can basically copy and paste an entire chat into a document or PDF and then simply upload it to the copilot’s Knowledge Base (though you'll eventually hit a limit so be judigious)! Now, this information acts as the copilot’s “long-term memory.” This hack will save tons of time in future chats.
Whoa, this feels like a huge unlock. So the copilot then “knows” everything we’ve talked about in this chat?
- Not exactly. Based on keywords I provide in future chat prompts – or any direct guidance to look in certain documents within the Knowledge Base – the copilot’s retrieval augmented generation (RAG) can essentially search through its long-term memory to find relevant information. So if I want it to reference or recall something from a specific conversation, I need to give it a little guidance (e.g., “During our December 6 chat in your Knowledge Base, you provided some advice on growing organic traffic using…”)
As our chat meandered toward its conclusion, my questions got a bit more philosophical: I found myself asking about the nature of human-AI interactions, the anthropomorphization of artificial intelligence (our desire to assign human characteristics to AI, even when we know it’s not applicable or appropriate) as well as the power of natural language, ethical considerations of AI use, and the limits and nature of AI capabilities.
I'm interested in better understanding (and testing) the boundaries of this odd and new means of interacting, and the copilot is down to clown.
Conclusion
My top 5 biggest takeaways from setting up and optimizing this AI copilot were:
- Working in an approved enterprise platform offered a new level of what we humans might call “psychological safety”: I felt much more comfortable sharing proprietary information and experimenting creatively with the AI without fear of getting myself or my company into trouble.
- The Knowledge Base allowed me to easily customize the LLM with personal and organizational information it would not otherwise be able to access. The ability to add documents and individual chats to this Knowledge Base over time means that the copilot will get smarter about my individual job role as we continue working together.
- The Image Analysis function allowed me to quickly use the copilot to integrate complex views and serve as a personal analyst who understands my unique data. It also serves as a personal customer service rep who can unstick my annoying technical problems.
- When I have a question about the copilot itself, I can simply ask, in plain natural language, for advice on how to improve our interaction (“Help me help you!”). This led to many unique and creatively exciting “a-ha” moments where I felt I was partnering with the technology rather than fighting against it.
- Yes, it’s a little scary, and it’ll get scarier. But humans have felt this way with every new technological leap throughout history. I believe the best we can do is strive to experiment and understand the newfangled with equal measures of caution and wonder.
Ready to start building copilots for your team? Learn more about Fuel EX and give us a holler here to set up a demo: https://www.fuelix.com/contact
Until then, I’ll continue experimenting and reporting back from my unique location in the dynamic landscape…
