There is a hope, almost universal in boardrooms right now, that AI will compensate for everything that is currently broken. The data is messy, but AI will make sense of it. The workflows are tangled, but AI will route around them. Nobody owns the customer record, but AI will reconcile it. The hope is understandable. It is also exactly backwards.
AI does not repair a broken operating model. It runs on top of one — faster. Point a capable model at unreliable data, undefined ownership and tangled workflows, and you do not get clarity. You get confusion produced at machine speed, with the added authority of something that sounds confident. The mess is no longer slow enough to catch.
AI is an amplifier, not a cure
The simplest way for a CEO to think about AI is as an amplifier. It takes whatever signal you feed it and makes it louder. If the underlying operating model is healthy — clean data, clear ownership, coherent workflows — AI amplifies a good signal into a real advantage. If the operating model is unhealthy, AI amplifies the noise.
This is what makes a premature AI programme so dangerous. A broken manual process is at least visible: people complain, errors are caught, someone notices. Automate that same broken process with AI and the failures move below the surface, executed thousands of times before anyone realises the logic was wrong from the start. The cost of the disease does not fall. It scales.
AI applied to a broken operating model does not remove the dysfunction. It industrialises it — and hands you a confident report on the way.
"AI Without Data" is a named disease
In our Disease Library, the AI and data family contains one of the most commonly funded and least examined diseases of this cycle: AI Without Data. The pattern is consistent. A company commits to an ambitious AI initiative while its data foundation remains fragmented, of unknown quality, and owned by no one in particular. The AI is built on sand and presented as a strategy.
The symptoms are recognisable. Models that work in a polished demo and degrade in production. AI outputs that no one fully trusts but everyone is asked to act on. A growing AI spend with no corresponding growth in the business. These are not failures of the AI itself — the models are usually fine. They are failures of the foundation the AI was placed on, and they are the predictable result of treating AI as a substitute for the data work, rather than a layer on top of it.
What has to be healthy before agents
Before an enterprise puts autonomous agents anywhere near real work, three things must be in place. Not perfect — but present, owned and governed.
- A data foundation. Trustworthy, accessible, reasonably consistent data with known quality. If the people in the business do not trust the data, the agents built on it cannot be trusted either.
- Clear ownership. Someone accountable for each critical data domain and each workflow an agent will touch. Agents act; accountability for those actions cannot be left undefined.
- Governance. Rules for what an agent may decide alone, what requires a human, how its actions are logged, and how it is audited. Governance is not a brake on AI — it is the only thing that makes AI safe to scale.
Notice what these three have in common: none of them is an AI problem. They are operating-model problems. Which is the entire point. The work that makes AI succeed is the unglamorous foundation work that should have been done regardless — and the reason most AI programmes underdeliver is that they skipped it in the rush to the agent.
The right order — and the right question for the board
The disease lifecycle does not change because the technology is new. The order still holds: diagnose the operating model, prescribe the foundation work, treat the data and ownership and workflow gaps, and only then layer AI on top of something that can actually carry it. Reverse that order and you are not accelerating — you are automating the very dysfunction you were trying to escape.
So when the board asks, "What is our AI strategy?", the more valuable question to put back on the table is, "Is our operating model healthy enough to be amplified?" If the honest answer is no, the first move is not an AI programme. It is the diagnosis and treatment that makes one worth funding. AI rewards the companies that did the boring foundation work — and quietly punishes the ones that hoped it would do that work for them.
AI amplifies whatever you already have — so fix the data, ownership and workflows first, because layering agents on a broken operating model only automates the dysfunction at scale.