When I turned 50 back in 2023, I decided it was time to join the gym again. But this time I was determined not to repeat my old cycle: join a gym, get injured, quit soon after, then spend months struggling to cancel the membership.
This time, I picked a coworking space with a gym and a trainer to remove excuses. The trainer gave me a diet plan and a daily 1,800-calorie target—plus homework: log everything I ate for a week. Tracking meals turned out to be a hassle, even when using top apps like MyFitnessPal. In fact, I found the app clunkier than a pen and paper food journal. At the time, ChatGPT just launched custom GPTs with computer vision features, so I started experimenting.
I decided to build my own calorie counter called “QuietDietGTP.” The idea was simple: I wanted something where I could upload a picture of a meal, and it would tell me how many calories I had left against my daily goal of 1,800.
I wrote an initial prompt for the custom GPT. I tested it by uploading several photos of my food plates and checking the results against the actual ingredients.
The results weren’t always perfect, but they were good enough to keep me honest and amused!
Over the next few months, I snapped pictures of every meal, tweaking QuietDietGPT with just plain English to get better results. I even taught it to track macronutrients (e.g., fats, carbs, proteins, and fiber)—no coding necessary.
Six months later, OpenAI upgraded its memory, fixing a bug for me for free. To this day, I still use QuietDietGPT daily—it’s so handy, it stuck.
My real-life consumer case study above illustrates how organizations can seize similar opportunities by embracing a progressive strategy. GenAI levels the playing field for non-technical product thinkers: those with customer insight, deep domain knowledge, and a vision for how things could work. This is because the technology makes it possible to rapidly build and test working prototypes, pilot improved workflows, and even implement internal tools using only natural language.
This hands-on approach is fast and inexpensive; it also builds crucial judgment about what’s viable, valuable, and risky before committing significant resources.
There’s a spectrum in building solutions with GenAI:
Convenience: Fast to build, inexpensive, and potential for “good enough” accuracy; ideal for discovery, quick feedback, and iteration.
Perfection: Slower and costlier to achieve but necessary for enterprise-ready applications, especially those that are integrated into enterprise systems, leverage company proprietary data, offer an optimized user experience, and must be secure, highly available, and cost-efficient for regulated or high-stakes scenarios.
User value tends to increase as you invest in more data, better tooling, deeper integrations, robust evaluation, and a refined user experience. But it’s a mistake to start by building the “perfect” solution.
Early experiments surface the features users need most, expose unexpected edge cases, and reveal hidden costs. Watch for trigger points like rising usage, recurring issues, or demands for better data and new integrations — these signal it’s time to invest more in accuracy, reliability, and scale.
A simple rule of thumb: start with convenient, Minimum Viable Product (MVP) grade solutions to validate value with a small number of users. When the returns justify the cost, invest in perfection and broader use.
Do you see sustained weekly active users?
Are active users asking if they can share it with others?
Are failures clustering into patterns you can address?
Is there a regulated or mission-critical need surfacing?
Will the next technical investment create significant user value that outweighs the cost?
Look for repetitive text or image tasks.
Select workflows that tolerate “good enough” answers.
Ensure automation will create clear value without a tremendous amount of upfront work.
Use existing GenAI tools; skip custom models.
Start with prompt-based logic.
Keep inputs and outputs simple.
Test the workflow yourself.
See what works, what doesn’t, and what can be improved.
Target 10-50 trial users.
Collect all failures and “almost-right” outputs.
Maintain a top-10 failure list.
Patch prompts, add guardrails, and tweak the user experience before chasing bigger technical upgrades.
A practical example is a GenAI app that automates transcription of meeting notes and action items. It’s possible to upload meeting recordings and summarize key decisions, action items, and deadlines with GenAI. This saves hours to weeks for teams and reduces human errors in manual notes. This is a great “good enough” solution, as imperfect summaries are better than none.
Consistently growing usage and demonstrating savings: If more teams are adopting your MVP and you can clearly point to time/cost savings or productivity gains, this signals the need for a robust infrastructure to support broader adoption across the enterprise.
Failures are recurring and correctable: Occasional errors are acceptable in MVPs, but if failures start recurring in predictable ways, and you know how to fix them, it’s time to harden processes, prompts, and controls rather than patching with quick fixes.
Emerging business risks demand reliability, traceability, or compliance: As your MVP touches more sensitive data, decisions, or customer-facing workflows, the cost of mistakes grows. This is when requirements for audit trails, transparency, and compliance become paramount.
Stakeholders demand assurance: If leadership, audit, compliance, or security teams start raising questions or if your clients expect reliability, you need robust monitoring, version control, and documentation.
Integration needs expand: Once your MVP needs to plug into core systems or automate complex, cross-functional workflows, the case for a maintainable, scalable enterprise platform becomes clear.
Integrate new information instead of refining natural language prompts: Connect your system to reliable company or online data sources. This gives your AI up-to-date, factual answers instead of guessing or relying solely on prompt engineering.
Set up automated evaluations and guardrails: Use automatic checks to catch mistakes or inappropriate outputs. Think of this as building safety nets so the AI doesn’t go off track.
Transition to human-in-the-loop review: Format the AI’s answers clearly so they’re easy to check. Let humans double-check and approve important results before they’re used.
For a user experience that requires more structure: Develop a web, mobile, or desktop app that uses GenAI where it’s most important and not for everything.
Cost & latency management: Use cheaper models for routine tasks, premium ones for critical moments.
Define what level of error is acceptable up front for each workflow: Decide ahead of time how many mistakes are okay for this task. This prevents surprises and helps everyone set realistic expectations for early results.
Don’t use sensitive data at the MVP stage; publicly available data is safer early on: Avoid feeding private or confidential info to the system in the early stages. Using public data protects your organization if the AI makes mistakes or data is exposed.
Set up simple incident/rollback plans: Have a backup plan in case something goes wrong, like switching back to manual work. This makes it easy to recover quickly without confusion if issues pop up.
Let early users know what “MVP” means — clarify known limitations: Tell test users that it’s still an experimental version and may not be perfect. Be clear about what the tool can and cannot do, so users aren’t caught off guard.
No/low-code stack examples: Prompt tools (ChatGPT, Claude, Gemini, etc.), simple orchestration (Zapier, Make, Relay.app), and solutions like AnythingLLM, Notion AI, or ChatThing that let you search, organize, and automate over your documents and knowledge bases.
Templates: Ready-made structures or frameworks that can be used as starting points when building GenAI solutions. These help teams quickly set up critical components and ensure best practices are followed. For example:
Problem framing templates: help define the user, task, constraints, and error tolerance for a project, making sure the scope and requirements are clear from the start.
Prompt pattern templates: offer reusable formats for crafting prompts, including instructions, example completions, and ways to handle edge cases, ensuring consistency and quality in user interactions.
Minimal telemetry templates: standardize how to collect essential usage data such as timestamp, latency, outcome, retries, and handoffs, making it easier to monitor and troubleshoot solutions.
Don’t wait for the “perfect” plan. As the old adage goes, perfect is the enemy of good. The biggest advantage GenAI gives you is learning by doing at a very low cost: you’ll only discover where perfection pays off by creating and deploying “convenient enough” prototypes. So let usage, and not theory, show where perfection is worth it. The future belongs not to coders alone, but to those who combine domain knowledge, rapid iteration, and creativity.
Interested in trying out QuietDietGPT yourself?
https://chatgpt.com/g/g-mu8agHqIt-quietdietgpt
Let’s Chat. Contact me if you’d like us to discuss how GenAI can help you set your project up for success.