How to Pick Your First AI Transformation Project: Start With Your Biggest Bottleneck
Ask five people inside a company what the first AI project should be, and the answers tend to cluster around the same wrong criteria: whatever's most visible to the board, whatever a vendor demoed most impressively, whatever a competitor announced last quarter. None of those are bad things to be aware of. None of them are good reasons to pick a starting project, because none of them say anything about whether the project will actually work.
The Criterion That Actually Predicts Success
The projects that succeed as a first AI transformation initiative tend to share one unglamorous trait: they target a bottleneck the business already understands deeply. Not a new capability nobody's tried before — a known, named, already-measured point of friction. A claims queue that backs up every month-end. A manual reconciliation step that takes three people two days. An approval chain that adds a week to every deal regardless of its size. The appeal of starting here isn't that these problems are easy; it's that the business already knows exactly what "better" looks like, which means success or failure is unambiguous rather than a matter of interpretation six months in.
Why "Known Bottleneck" Beats "Novel Capability"
A first project built around a new capability — "let's see what AI can do for customer sentiment" — has no existing baseline to measure against and no established owner who already cares about the outcome. We've written about why transformation programs stall without a clear owner and a defined success metric; starting with a known bottleneck solves both problems before the project even begins, because the owner and the metric already exist. Nobody has to invent a way to measure success — the business has been measuring the pain of this bottleneck for years, usually in a spreadsheet somebody already maintains.
Why This Needs an Honest Data Check First
Picking the right bottleneck is necessary but not sufficient. The data layer underneath a given process has to actually be ready before AI can meaningfully help with it — a bottleneck caused by three disconnected systems that don't agree on the same customer record isn't a bad choice of starting project, but it does mean the real first phase of work is reconciling that data, not deploying a model. Better to discover that during a two-week assessment than six weeks into a pilot that can't produce a trustworthy answer.
The Trap of Starting Too Big
A known bottleneck is the right scope filter, but it's still possible to pick one that's too large for a first project — a company-wide procurement overhaul instead of the specific approval step that's actually slow. The first project should be bounded enough to ship and prove out in a defined window, not a transformation of an entire function. The board-level case for AI transformation gets easier to make the second time around specifically because the first project was scoped tightly enough to produce a clean, defensible result — a messy, oversized first attempt makes the next budget conversation harder, not easier.
What to Actually Do Before Picking
List the three or four processes in the business that already generate the most complaints, the most manual rework, or the most visible delay. For each, ask: do we already know how to measure this well or badly, is there a clear owner who wants it fixed, and is the data this process touches reasonably consistent across the systems involved? The process that answers yes to all three, not the one that sounds most impressive in a kickoff meeting, is the right place to start.
If you're weighing a few candidate projects and want a second opinion on which one is actually ready, let's walk through the shortlist together.
Comments