Lindsey Witmer Collins is Founder and CEO of WLCM.AI, an Inc. 5000 company building AI solutions for enterprises.
Here’s what nobody wants to admit: 95% of AI pilot programs return zero value, according to MIT. We’ve spent the past couple of years in the “gee-whiz” phase—everyone trying things, sharing prompts, marveling at what’s possible. That was fun. But now we’re at the rubber-meets-the-road phase, where organizations need AI systems that actually deliver value, not just demos that impress in meetings.
The evolution of AI has felt like one big group project so far, where individuals, companies and researchers are organically experimenting. But as the novelty wears off, three distinct disciplines are emerging that separate working AI systems from expensive experiments that go nowhere.
Workflow Engineering: Designing The Factory Floor
AI can take over the machine-like work that bogs down your team: the tedium, the repetitive stuff, the high-volume grunt work. But here’s where most organizations mess up: They don’t think through the handoffs.
Workflow engineering maps out exactly what the human does and what the AI agent does, and how they pass work between them. It’s the process-level design of your AI system—how data gets gathered and cleaned, which model handles which task and how the system gets deployed and maintained over time.
Think of it as designing a factory floor. You need designated workstations and an efficient layout. You can’t just throw AI at a problem and hope it figures itself out.
Where This Goes Wrong: Companies build a bunch of isolated AI tools instead of one complete workflow. Someone might create separate AI assistants for email, scheduling and research. Congratulations, you’ve replaced one kind of tedium (doing the work) with another (managing three different AI tools with three different logins).
Or they go the opposite direction and build one massive agent that’s supposed to handle everything, which degrades performance and creates engineering nightmares. Finding the right-sized use case for AI-inclusive workflows is critical.
Prompt Engineering: Writing The Instructions
This is how we talk to AI—the natural language commands and questions that direct the model to do what we want. There’s an art to asking the right question in the right way, and well-crafted prompts are what separate useful outputs from garbage.
Think of prompts as the instructions for your factory workers; they keep everyone safe and ensure quality output.
Here’s something most people don’t know: Prompt adherence (how well a model follows instructions) varies wildly between models. So does each model’s preference for how prompts should be structured. I run a workflow in my company that uses five different models, each leveraging unique strengths. Some are highly specialized. Model selection matters here.
Where This Goes Wrong: Prompting is the most user-facing part of AI, so people get fixated on finding that one magical prompt when the real issue is deeper, usually at the workflow or context level. Users either overpack their prompts with excessive detail and business logic (which confuses the AI and burns through tokens, hiking up costs) or they keep tweaking prompts when the problem is actually that the AI doesn’t have access to the right information.
Context Engineering: Building Institutional Knowledge
All AI depends on data to function. Garbage in, garbage out. But context engineering goes beyond just the data in your prompt. It manages the broader information ecosystem that your AI pulls from.
This is the well of institutional knowledge your AI uses to make decisions and produce outputs. The more current, complete and intentional that background knowledge is, the more intelligent your AI can be. Tools like retrieval augmented generation (RAG, a method for connecting AI to specific data sources) and knowledge graphs let you feed your AI information strictly from trusted sources you choose.
Think of this as the training and accumulated experience of your factory workers. That’s what enables deep, expert-level work.
Where This Goes Wrong: Organizations throw everything but the kitchen sink into the AI’s context window (its working memory—the maximum information it can consider at once). This creates latency, bloats token usage, drives up costs and produces unreliable outputs because the AI can’t figure out what actually matters. The direction becomes unclear.
The Organizational Mistakes I Keep Seeing
At the broadest level, leaders are “spraying and praying”—mandating AI adoption with no clear objectives or instructions, then hoping it works out.
More specifically:
• Leaving It To Employees: Organizations let people adopt whatever AI tools they want and use them however they think might work. Most employees aren’t AI engineers. This is risky and creates chaos.
• Not Understanding The Workflows: You can’t improve a process with AI if you don’t understand the process. I see companies automate the wrong things, create “franken-workflows” that don’t actually save time or build systems so complex that nobody uses them.
• Ignoring Scale: Organizations build an AI system that works for the pilot program, then everything falls apart when they try to expand. They don’t plan for keeping context clean and current. They don’t anticipate rising token costs as usage grows. They don’t build in testing and fine-tuning (which is critical). They end up with something that doesn’t work but costs a fortune.
AI As An Organizational Capability
AI shouldn’t just be a tool you plug in. It should be an organizational capability—something baked into how your company operates.
Building AI systems that actually work means more than finding good models or writing clever prompts. It requires deeply considered architecture that supports systems that learn, adapt and scale alongside your business. It means understanding the interplay between workflow, prompts and context. It means planning for maintenance, iteration and growth.
The companies that figure this out won’t just have AI tools. They’ll have a genuine competitive advantage. The ones that don’t will have expensive pilot programs gathering dust and a finance team wondering where all that AI budget went.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
.jpeg){kind=small}
