Most teams get stuck in the same place with AI. They think too big. The instinct is to start with transformation. End-to-end automation. A complete reinvention of how work gets done. The result is usually the same. Nothing ships. Seth, a marketing operations manager at Publicis Sapient, took a different approach. “Don’t start big,” he says. “Start with a real problem.” His starting point was not flashy; it was frustrating and a major time suck. It was the dreaded campaign naming.
The kind of work that slows everything down
Campaign naming sounds simple until you are doing it at scale. Each name needs to follow a strict structure. Industry, platform, vendor and other variables all need to be included in the right order, every time. One mistake can break reporting or create downstream issues. It is repetitive. Rule-heavy. Easy to get wrong. In other words, it is exactly the kind of work AI should handle. For Seth, he was able to implement a simple GPT into his everyday by following a few simple steps.
Step one: find the friction
The best place to start with a custom GPT is not the most advanced use case. It is the most painful one.
Look for tasks that are:
- Repetitive
- Structured by clear rules
- Prone to human error
Campaign naming checked every box.
Step two: write everything down
Before building anything, Seth documented the process in detail. Not just what the task is, but how it works. What inputs are required? What should the output look like? What rules cannot be broken? This included taxonomy, naming conventions and character limits. That documentation became the foundation of everything that followed. A simple but important habit made a difference. The document stayed live, with a running change log. As the system evolved, so did the rules.
Step three: turn rules into instructions
Once the process was clearly defined, the next step was straightforward. Take the documentation and convert it into a structured instruction set for a custom GPT. This becomes the system’s operating logic. It tells the model what to do, how to do it and what not to do. The quality of the output depends on the clarity of this step.
Step four: build for real workflows
A common mistake is building a GPT that does one thing. A stronger approach is to think in terms of workflows.
Seth’s solution did not just generate campaign names. It also:
- Validates naming conventions
- Fixes formatting issues such as hidden characters
- Extracts structured data from messy briefs
- Prepares bulk-ready outputs for upload
It acts as an assistant, not just a generator.
Step five: add guardrails early
During testing, Seth noticed a familiar issue. The GPT started inventing vendors and categories. That is where constraints matter.
To make the system reliable, he introduced strict rules:
- Defined lists for acceptable inputs
- Clear instructions not to deviate
- Structured output formats
AI performs best when boundaries are clear.
Step six: test it like it will break
Instead of assuming the system worked, Seth pressure tested it. One effective method was using another GPT to identify weaknesses. Questions like “What test cases should I run?” and “Where could this fail?” helped surface edge cases early. It is a simple way to strengthen the system before it scales.
Step seven: scale is where it pays off
Individually, the task did not seem like a major problem. Each campaign took about 10 to 15 minutes to name. But scale changes the math. With around 100 campaigns, that added up to more than 16 to 25 hours of work. With the GPT in place, the same workload now takes about 20 minutes. The time savings are clear. The consistency is just as important. Standardized naming improves data quality, which makes reporting and decision-making more reliable.
Step eight: keep refining the system
No GPT stays perfect. As requirements change or edge cases appear, the system needs to evolve. When something breaks, the fix is not guesswork. Update the source document. Regenerate the instructions. Relaunch the system. It is a structured process, not trial and error.
What actually makes a good AI use case
The value of this work is not in the complexity of the solution. It is in the clarity of the problem.
The most effective AI applications tend to do three things well:
- They remove tedious work.
- They standardize messy processes.
- They scale output without increasing effort.
Or as Seth puts it, “Use AI to do the work you don’t want to do so you can focus on the work that actually matters.” That is where momentum starts.
