Essay 01
Why AI Projects Fail — And Why It’s Not What You Think
By Steve Cannon
I’ve spent six years watching AI projects fail inside real companies. Not fail because the technology didn’t work. Not fail because the talent wasn’t there.
Fail because the organization didn’t have the structure to act on what the model was telling them.
Here’s what that looks like in practice.
My team built a model at a Medicare insurance company losing $1–2M per year on $79M in revenue. What we found:
- 3 of 50 states were causing $9.7M per year in losses
- Recommendation: stop buying leads in those three states
- Modeled opportunity: +$8.7M per year in EBITDA
The model was right. The data was clear.
The company didn’t act. Why? Because leadership couldn’t explain why those three states were loss-making. The model found the pattern. It couldn’t provide the story. And without the story, nobody moved.
Same losses. Year after year.
That’s not a technology problem. That’s not a talent problem. That’s an agency problem.
What the agency problem actually is
The agency problem is simple. It’s what happens when the people making decisions don’t bear the full consequences of those decisions. Managers protect their budgets, their teams, their explanations. A model that threatens any of those things gets ignored — not out of bad faith, but because the incentive structure makes ignoring it the rational choice.
This happens everywhere. The only structure I’ve seen actually solve it is private equity.
Think about what PE actually gives you:
- Controlling interest — the authority to act on what the data says, even when management resists
- Rigorous metrics — the language to defend model-driven decisions
- Patient capital — the time to let the model be right
PE wasn’t invented for AI. It was invented to solve the agency problem between managers and owners. But it turns out those are the same problem.
Most funds just aren’t using their toolkit this way yet.
That’s the gap system8.ai exists to close.
Steve Cannon, Founder & CEO, system8.ai
Essay 02
Don’t Clone Humans Into Agents
By Steve Cannon
Every AI firm right now is selling the same idea.
Map each human role to an AI equivalent. Build the agent that replaces the analyst. The agent that replaces the customer service rep. The agent that replaces the operations manager.
It’s intuitive. It’s visual. It almost always fails.
Here’s why.
A job description is not a workflow. When you try to replace a human role with an AI agent, you inherit everything that role actually involves:
- The messy data inputs the human was quietly cleaning
- The edge cases handled with judgment that was never written down
- The organizational dependencies nobody documented
- The failure modes — which in an agent are harder to predict and harder to explain than in a human
The result is a project that is technically ambitious, organizationally disruptive, deeply expensive — and produces results roughly equivalent to what the human was doing before.
That is not value creation.
The right question isn’t “which roles can we replace?” It’s “where are decisions being made badly?”
Those are different questions. They lead to very different places.
The wins in Phase 2 consistently come from three places:
- Decision optimization — find the decisions made hundreds of times a day and make them better with math. Lead buying. Pricing. Scheduling. Inventory. The human stays. The decision quality improves dramatically. EBITDA moves.
- Workflow redesign — instead of replicating an existing workflow with AI, ask what the workflow would look like if you designed it from scratch today. The answer is almost never the same workflow with a human swapped out. It’s usually something fundamentally better.
- Automation of the genuinely repeatable — data entry, report generation, routine communications. Fast to automate, easy to verify, immediately visible in headcount economics. And it frees people to do work that actually requires them.
The test before any AI initiative
- Are we replacing a role or improving a decision?
- Are we replicating a workflow or redesigning it?
- Is this task complex or just time-consuming?
For a PE fund manager, this matters for a specific reason. Clone-the-human projects look ambitious, consume significant capital, generate organizational resistance, and produce results that are hard to measure.
Decision optimization projects move EBITDA. They build on each other. They generate a documented value creation story that holds up in diligence.
That’s the difference between an AI initiative that ends up in a footnote and one that ends up in the exit narrative.
Steve Cannon, Founder & CEO, system8.ai
