I keep seeing the same story play out.
An organization starts strong. Real executive buy-in, a clear process to automate, a team that's genuinely excited about building something with agents. The architects sketch out the orchestration. The first sprint delivers a working prototype. Everyone's nodding.
Then the demo stumbles. The agent hallucinates on an edge case. It nails 85% of inputs but trips on something weird — an input nobody anticipated, a request phrased in a way the prompt didn't account for. Confidence wavers. And suddenly people are floating "maybe we should go back to a rules-based approach" or "let's pause and revisit in six months."
I've started calling it the flinch. And it's where good projects go to die.
It's a prompting problem, not a model problem
Here's what I keep telling teams when they hit that wall: the gap between 85% and 97% accuracy almost never means the model can't do it. It means you're talking to it wrong.
The pattern is always the same. Someone built one massive prompt that tries to handle the entire process in a single pass. Every exception, every edge case, every piece of conditional logic — all crammed into one set of instructions. Then they're surprised when the model gets confused.
That's not how this works. You break the problem apart. You give each sub-agent a specific, well-scoped job. You write tight prompts for each step. You add checks between stages. Think of it less like writing one giant instruction manual and more like building a team where each person has a clear role.
The teams that push through and do this work? They get to 97%. And 97% with a human reviewing the remaining 3% is a massive outcome for most business processes. That's not a compromise. That's a win.
You're delivering into a future, not into today
This is the part that really gets me. I regularly see projects with 6-to-18-month delivery horizons. Think about that. If you build on traditional, static architecture, you're delivering a solution in late 2027 that was designed with legacy thinking from before the agentic wave.
The models are getting better fast. The orchestration tools are maturing. The governance frameworks are catching up. Six months from now, the agent ecosystem will be meaningfully more capable than it is today. That means the agentic foundations you lay now only become more powerful over time. You're not betting on today's technology — you're riding a curve that's still accelerating.
Building on legacy patterns for a project with a year-plus timeline is like wiring a new office for fax machines in 2010. Sure, it works. But you're investing in the wrong direction.
Every serious IT project or digital transformation right now needs agentic AI baked in from the start. Not as a phase two add-on. Not as a nice-to-have once the "real" system is done. As a core design choice.
Cold feet is a leadership problem
When organizations flinch, it's almost never the engineers pulling the plug. It's leadership losing confidence because they expected perfection on the first try.
But we've been here before. Nobody assumed their cloud migration would have zero hiccups. We accepted the bumps because we understood the destination was worth it. Agentic AI deserves the same patience — more, honestly, because the upside is bigger and the underlying tech is improving faster than anything we've seen.
Yes, it takes confidence to set your organization on an AI-driven path. I won't pretend otherwise. You're placing a bet on the technology, on your team's ability to learn, and on where the market is heading. But from where I sit, working with organizations across industries: this is the clearest path to staying competitive in what's coming.
The companies that push through the flinch — that treat the first stumble as useful data, not a reason to retreat — are building real advantages. The ones that pull back? They'll restart the same project 18 months from now, having lost both the learning and the head start.
Here's the thing
If you're in the middle of an agentic AI project and it feels hard, that's normal. That's what building something that matters feels like from the inside. The hallucinations can be reduced. The accuracy can be improved. The architecture can be sharpened.
But only if you stay in the game.
Don't flinch.