Most AI implementations are failing according to the stats, and it is a human problem.
There's a framing error I keep seeing, and it explains most of the failures. Know what you are using and WHY. Using AI as a tool isn't the same as using AI for an agentic workflow.
Human in the loop means you're working with AI as a tool. A developer using Claude code to write code. A writer using ChatGPT to draft. You're in the execution alongside it. That's fine - that's a great use of AI.
Human on the loop is different. That's when AI is doing the work and you're directing it as a manager. You set the mission, review the output, adjust, and govern. The AI runs. You steer, you are the manager.
Agentic AI - any AI that operates with real autonomy in your business - requires you to be human on the loop. The moment you treat it like you're in the loop with it or out of the loop, you've lost the plot.
The mental model is pretty simple, to have truly agentic workflows you're not a peer - you're the manager. And this mental model is what will drive success or failure. It is what will drive insane token burns, or deliver results. Know the difference and which you are using.
THE FIVE-STEP ROLLOUT
1. Define the workflow
If the workflow isn't documented, you don't have a workflow. You have a habit, something that exists in someone's head and gets done differently every time.
You cannot hand an undocumented process to an AI and expect consistent output. The AI will improvise. Improvised AI in a customer-facing or revenue-critical role is a liability. Document the workflow first. Then, and only then, figure out where AI fits.
2. Define a specific, bounded role before you build anything
AI cannot be everything. Pick one job. Scope it completely.
An AI receptionist that handles inbound calls, books appointments, and routes escalations to the right person is a role.
A digital AI employee that helps with the business is not a role - it's crappy sales.
Undefined scope produces inconsistent results - for human teams or AI teams. The AI will do something, it just won't reliably do the right thing, because nobody told it what the right thing is.
3. Training is the real work — and it takes longer than you think
Here's where every "set it up in 30 minutes" demo falls apart. Scoping takes a day. Setup takes a day. Training takes weeks. Those crappy sales - all leave off the part about weeks of training and oversight.
The AI needs to know what good looks like. It needs examples of correct outputs, guidance on edge cases, clear rules for when to escalate and when to proceed. It needs to be run through scenarios, corrected, refined, and run through them again.
This is onboarding. Any manager who has hired a junior employee knows that the first few weeks are expensive - high attention, high correction, lower output. Agentic AI is no different, you're doing it in prompts and system instructions instead of one-on-ones.
MIT research found that 95% of GenAI pilots fail to reach production. The 5% that succeed share a common trait: they designed for friction, embedding AI into specific workflows with structured learning loops and governance.
4. A manager has to own the oversight
There is no set-and-forget with agentic AI. Not yet. Even if manager is not in your title, you're still leaning on the same skillsets. Reviewing outputs, looking for opportunities for improvement, ensuring alignment - all a real ongoing role that takes time and skills.
If nobody owns it, the agent degrades. Quietly. It keeps producing output - but that output drifts further from what you actually want, and nobody notices until something goes wrong. Burning too many tokens, hallucinating results, that is a manager's responsibility in the same way it is if a junior employee was making the same mistakes.
5. Timelines are longer than expected
The mistake isn't that AI takes longer than a two-week pilot suggests. The mistake is thinking the pilot proved something it didn't.
A pilot proves the concept works. It doesn't prove the agent is trained, stable, and ready to run unsupervised under real conditions.
McKinsey data puts the average enterprise AI project at 17 months from start to production. For projects that actually succeed, prototype-to-production averages around 8 months. Scope for the real timeline. Don't treat the demo as done.
Those 30 minutes to pick a name and avatar for your digital employee ads - misleading at best though I tend to think more outright dishonest.
WHAT THIS LOOKS LIKE IN PRACTICE
Moving beyond working with AI as a human in the loop and to agentic workflows as a human on the loop looks and feels a lot like hiring an entry level but highly capable employee that you need to grow.
Getting those questions answered clearly, before you build anything, is most of the work. It's also the step that most people skip - which is why most projects don't reach production or deliver intended results.
If you want help in planning this out, that is exactly what my ClearPath Map is designed to do. It's a structured research and scoping engagement that produces a clear map of where and how AI fits in your operation, what a bounded first role looks like, and what the real timeline and requirements are.