The answer up front: Before AI can improve your decisions, your decisions need somewhere to live: clean data, clear owners, and rules for what gets decided by whom. Gartner predicted at least 30 percent of generative AI projects would be abandoned after proof of concept by end of 2025, with poor data quality a leading cause. Infrastructure first, intelligence second.
Imagine hiring a world-class chef for a kitchen where nothing is labeled.
The pantry has flour in a jar marked “sugar.” Half the recipes live in a binder, half in the previous chef’s head, and the previous chef moved to Denver. The ovens work, mostly. Now ask your expensive new hire to cook something extraordinary.
The chef isn’t the problem. The kitchen is. And that, in one image, is what most companies do with AI.
What is decision infrastructure, in plain terms?
It’s three unglamorous things:
- Data that means what it says. One customer list, not four that disagree. Numbers someone actually maintains. If two systems give two answers to “how many active customers do we have,” you don’t have data, you have competing versions of it.
- Owners. Every recurring decision has a name attached. Not a committee, a name.
- Rules of the road. Which calls can anyone make, which need review, and what never gets decided by software alone. That last one is governance, which sounds like paperwork but is really just “the reason your AI can’t accidentally issue a refund to everyone in Missouri.”
Nothing on that list requires buying anything. Which may be why it gets skipped: nobody celebrates labeling the pantry.
What happens when you skip it?
You join a well-documented club. Gartner predicted at least 30 percent of generative AI projects would be abandoned after proof of concept by the end of 2025, and its list of reasons starts with poor data quality and inadequate risk controls. The pattern behind the stat is always the same story: the pilot demos beautifully on clean sample data, then meets the real pantry.
AI is an amplifier. Point it at good infrastructure and it compounds the good. Point it at contradictory data and undefined ownership, and it produces confident, fluent, fast wrong answers. Speed is not a favor when the direction is wrong.
Isn’t this just “clean your data,” the advice everyone ignores?
Partly, but with a crucial difference: you don’t need to clean everything. That’s the version of the advice that kills momentum, because “fix all the data” is a two-year project nobody funds.
You need the data for one workflow, the one you’re about to improve. One pantry shelf, labeled correctly, is enough to cook one great dish. We covered how to pick that workflow in Where AI Actually Creates Leverage: repeats at volume, data reachable, outcome measurable. The infrastructure work inherits the same scope. Fix the shelf you’re cooking from, not the whole warehouse.
What does “governance” mean for a company my size?
Smaller than you fear. For a mid-sized business, governance is a one-page answer to four questions:
- What data can the AI read, and what’s off-limits?
- What can it do on its own, and what needs a human sign-off?
- Who’s accountable when it’s wrong? (A name, again.)
- How would we know it’s drifting? (Someone reviews the outputs on a schedule.)
That page is boring. It’s also the difference between an AI system you can trust with real work and a demo you keep at arm’s length forever. The companies that move fastest with AI aren’t the reckless ones, they’re the ones who settled these questions early and stopped re-litigating them.
So what’s the actual order of operations?
Pick the one workflow worth improving. Label its shelf: get that data into one trustworthy place. Write the one-page rules. Name the owner. Then bring in the intelligence, whether that’s a simple tool or something custom-built.
Run it in that order and the AI lands in a kitchen where it can actually cook. Run it backwards and you’ll spend next quarter explaining to the board why the pilot got quietly shelved.
If you’d like help figuring out which shelf to label first, that’s the first thing our audit looks at: the state of the kitchen before anyone talks about chefs.
One short read like this, twice a month: what’s changing in AI, why it matters, and what to actually do about it. Get the Insights below.
Quick questions
Do we need a data warehouse before using AI? No. You need trustworthy data for the one workflow you’re improving. Scope the infrastructure to the use case, and grow it as the use cases grow.
What’s the minimum viable governance? One page: what the AI can read, what it can do alone, who’s accountable, and who reviews it. Write it before launch, revisit it quarterly.
Our data is a mess. Are we years away from AI? Almost certainly not. Most companies can get one workflow’s data trustworthy in weeks. The trap is trying to fix everything before doing anything.
Source: Gartner press release, July 2024 — “at least 30% of generative AI projects will be abandoned after proof of concept by end of 2025,” citing poor data quality, inadequate risk controls, escalating costs, or unclear business value.
