Start with Understanding – not Hype
Before jumping into implementation, understand the landscape. Many companies rushed into AI in 2023 due to FOMO — in fact, 67% of IT decision-makers admitted fear of falling behind drove their adoption decisions. Don’t adopt AI because others are doing it. Instead of adopting AI because “everyone is doing it,” ask “Does this make sense for us?” and then proceed with clear eyes. Here is the playbook on how one would approach it systematically based on current State of AI
Start here to cut through the hype:
🔗 The State of AI – Cutting Through the Hype
Leverage this framework to guide ideation and match the right layer of AI:
🔗 Learning AI Through Classifications
Or Deep dive with the library of compiled learning here:
🔗 https://github.com/skychew/LearningAI-MyJourney
🔗 https://skychew.com/category/ai/
Conceptually the 5 steps here uses the 3 Pillars of AI Readiness:
- Systems: Identify where AI fits into workflows and decision-making
- Technology: Integrate or build tools (Make.com, LangChain, open-source models, etc.)
- People: Train team on prompt engineering, ethical AI use, tool adoption
How to Implement AI Agents in Your Business
Step 1: Identify High-Leverage Workflows [Systems]
Use the Pareto Principle to map out the 20% of work that drives 80% of business value.
Ask:
- What are our most critical workflows?
- Where do we see bottlenecks, manual work, or delays?
- Look for repetitive or data-heavy tasks that require decision-making.
- Examples: Customer support routing, document analysis, inventory forecasting.
Step 2: Ideate AI Use-Cases (Innovative + Creative Solutions) [Systems, Technology]
Map value drivers to AI possibilities:
- Innovative Solutions: Improve existing processes with AI tools (e.g., automation, recommendations, intelligent routing)
- Creative Solutions: Explore new AI-powered offerings, services, or business models
Choosing the Right AI Solution / Models
- Do we have proprietary data (Layer 0)
- Workflow-Based AI (Prompt chaining, routing) for structured processes.
- Full AI Agents for dynamic, complex decision-making.
- Can we improve simple workflows using pre-trained models(Layer 1)
- Can we automate, enhance, or augment tasks (Layer 2)
Use design thinking workshops or design sprints to:
- Reframe problems with “How might we” prompts
- Ideate “What if AI could do this?” use cases
Step 3: Launch a Pilot AI Solution [Technology]
Start small. Choose a low-risk, high-impact workflow to test AI.
- Test AI agents in small-scale scenarios before full deployment.
- Automate customer queries
- Draft reports using LLMs
- Categorize documents or emails
- Measure success metrics like:
- response time
- accuracy
- cost savings.
- Speed
- accuracy
- quality / customer satisfaction
Step 4: Build Internal Capability [People]
Operationalize success using the 3 Pillars of AI Readiness:
- Upskill employees on AI capabilities and ethical considerations.
- Example: Train team on hard skills prompt engineering, tool adoption etc.
- Ensure human oversight where needed.
Step 5: Scale & Create an AI Culture [People]
Expand from pilot to broader rollout:
- Refine based on feedback and performance data
- Document what works and what doesn’t
- Foster experimentation: make it safe to test and learn with AI
Inspire a culture of:
“Let’s try it – what can AI do here?”
Start Small, Win Fast
- Pick one high-impact area and build a quick win
- Document results, then scale
- AI is not a silver bullet – but with clarity and focus, it becomes a powerful tool.
Training Your Own Model
If you’re considering training your own AI model instead of using existing tools or APIs, here’s a heads-up: most businesses vastly underestimate the resources required.
While the idea of owning a proprietary model sounds attractive, the reality involves heavy costs in data, hardware, and engineering effort. Unless you’re solving a highly specialized problem with access to unique data, you’re likely better off fine-tuning existing models or leveraging APIs.
But if you’re serious, here’s a minimum viable benchmark to give you context:
🔹 Data
- At least 100,000 to 1M+ high-quality, structured examples for basic model training. Expect this number to go much higher for domain-specific or language models.
- Labeled data is essential. Manual labeling or data-cleaning pipelines are often needed. [Structured data]
- Data infrastructure: Data lake or warehouse, versioning tools, pipelines for ETL (extract, transform, load).
🔹 Hardware
- Training even a small LLM (~100M–1B parameters) requires powerful GPUs (e.g. NVIDIA A100s or equivalent).
- Minimum setup: Access to 4–8 GPUs (can be cloud-based), or a dedicated training environment on platforms like Lambda Labs or AWS EC2 GPU instances.
- Cost: Training costs can range from $10k–$500k+, depending on model size and training duration.
🔹 Team & Talent
- AI/ML Engineers: With deep learning and data pipeline experience.
- MLOps Expertise: For deploying, monitoring, and iterating models.
- Data Scientists: To structure and interpret the problem and results.
🔹 Time
- 6–12 months minimum to build, test, and deploy a usable model.
- Ongoing iteration and tuning will be needed to reach performance benchmarks.
Bottom Line:
Training your own model is like building a rocket to deliver a pizza. Make sure the ROI justifies it. For most use cases, starting with open-source or API-based models (like OpenAI, Mistral, or Claude) and fine-tuning them on your data is more than enough — and far more efficient.
If you’re still committed to going down this path, make sure you have a clear problem worth solving, and the infrastructure, team, and budget to back it up.
I put this playbook together based on my experience with digital transformation, venture building, automation, and emerging tech. Think of it as a guide to help teams like yours move from just being curious about AI to actually using it in ways that make a real impact.
I’ve led cross-functional teams through transformation journeys, connected up siloed systems, and rolled up my sleeves with tools like n8n, Make.com, PyTorch, and building web app and mobile app soilutions. I try to stay close to where tech (No-code, AI and Blockchain) is heading—not just for the shiny new stuff, but to figure out what’s actually useful and tied to real business goals.
If you’re trying to bring smarter workflows or AI into your operations (without the hype), I’d love to chat and see where I can help. https://www.linkedin.com/in/skychew
