AI transformation is changing how a company operates with AI, not just adding tools. Most companies are spending millions on AI with nothing to show for it. The problem is not the technology. It is the operating model underneath it. AI amplifies whatever it lands on, so when decision rights are unclear, the cadence is weak, and the data is a mess, AI amplifies the mess. That is why so many pilots impress in a demo and then quietly die.
Done right, AI transformation starts with the operating model: where decisions are made, how work flows, and where the measurable ROI actually is. Fix that, then apply AI where the return is clearest. Sequence it the other way and you get expensive experiments instead of compounding gains.
Why most AI initiatives fail
- No clear owner, so the initiative escalates to the CEO and stalls.
- No defined outcome, so "success" means "we'll discuss next steps."
- AI bolted onto broken operations, amplifying the dysfunction.
- Pilots with no pre-commitment to scale, so they never leave the lab.
Operations first vs. AI bolted on
AI transformation goes one of two ways, and the difference shows up long before the model does. Bolt AI onto the operation you have and it amplifies the dysfunction. Fix the operating model first and AI compounds a system that already works.
| AI bolted on | Operations first | |
|---|---|---|
| Starting point | The tool and the demo | The operating model: decision rights, outcomes, cadence, data |
| Who owns it | Nobody, so it escalates to the CEO | One named owner per outcome |
| What success means | "We'll discuss next steps" | A metric, a target, and a deadline agreed before the build |
| What AI amplifies | The dysfunction already there | A system that already works |
| After the pilot | It never leaves the lab | Pre-committed to scale, then kept current as the models change |
How to do AI transformation right
Everyone is pitching AI builds. Most of them are automating broken processes, and many of those processes should not be problems in the first place. So the order matters more than the tooling. Install the operating system first: clear decision rights, measurable outcomes, a cadence that forces closure, and the data discipline to trust what you are feeding the models. That is the VOOCS framework. Then build with production discipline, training and governance built in, pre-committed to scale.
And then keep it current. Foundation models turn over roughly every year, and a vendor can retire one with as little as 60 days' notice, so an AI system nobody maintains is a system already aging out.
None of that is an argument for going slow. Speed is a feature. Two-week sprints and three-month launches were built for a constraint that no longer exists, and teams still plan around it out of habit. Decide, ship, and let the weekly cadence catch what you got wrong. Slow is usually not caution. It is nobody owning the decision.
This is why I built KeyDelta the way I did. Fix what's broken. Build what pays. Keep it relevant. The firm: keydelta.com. The execution framework underneath it: VOOCS.
Two terms I coined for this era
In 2026 I coined two terms for what is happening to software. AI-Generated Software (AIGS) is software AI builds fast: anyone can produce it now. AI-Enabled Software (AIES) is software with a standing system that keeps it secure, current, and evolving. AIGS is what gets built this week. AIES is what is still running, and still evolving, next year. SaaS came with a vendor whose job was keeping it current. AIGS does not. The full definitions live on the KeyDelta pages linked above.
That gap is the reason I built KeyDelta: everyone can build software now, and almost no one can manage what they build.
The data
- MIT, State of AI in Business 2025: 95% of enterprise AI initiatives deliver no measurable value. Source
- SailPoint, 2025: 80% of organizations report AI agents taking unintended actions; fewer than half have governance policies for them. Source
- Menlo Ventures, State of Generative AI in the Enterprise 2025: enterprise generative AI spending reached $37B in 2025, up 3.2x from 2024. Source
- Deloitte, State of AI in the Enterprise 2026: 74% of organizations want AI to grow revenue; 20% say it already does. Only 25% have moved at least 40% of their AI experiments into production. Source
The mistakes I see most
- Starting with the technology, not the business outcome. "Implement AI" is not a strategy.
- Treating AI as a side project of one team instead of a capability the whole business uses.
- Demanding perfection from AI while accepting mediocrity from manual processes that have been broken for years.
- Buying tools before fixing data definitions, so every model produces unreliable results.
- Running pilots with no pre-commitment to scale, so even the wins die in the lab.
- Shipping the build with no owner for keeping it current, so the system starts aging the day it goes live.
The keynote
I speak on this for CEO audiences, PE operating partners, and leadership offsites. The talk is The Pilot Graveyard: Why AI Works in the Demo and Dies in Production, and it is about why AI pilots die before production and how to commit to scale before the pilot starts. See speaking.