Four steps to a workflow? Not Quite.

Scroll LinkedIn and you'll find plenty of neat four-step diagrams: Idea. Prototype. Integrate. Deploy

They look great, it's easy to look at, digest and believe however, just aren't how AI transformation actually happens.

What actually stands between an idea and something in production isn't a step. It's four much harder, much less photogenic things. None of them fit neatly into a box, and none of them are optional.

The part most AI projects skip

Most AI projects start with a solution already in mind before anyone's properly interrogated the problem. This is where you end up with an AI system that automates the wrong version of the process, technically functional, practically useless.

Real discovery means sitting with the people who actually do the work and extracting the unwritten rules they apply instinctively, the exceptions they've never documented, the judgment calls that live entirely in their heads.

The step that breaks more projects than any other

If the data behind a workflow is not ready, your workflow will fail. Profiling the real data against the rules a workflow needs to follow, before a line of production code is written, catches this early. Skip it, and the mess surfaces later, in production, where it costs far more to fix and does far more damage to trust in the system.

It’s not a formality, it’s a gate

In regulated environments, governance isn't a final checkpoint. It's a requirement from day one. Questions around accountability, auditability, security, and data residency must be answered before anything reaches production. This is usually where projects that looked finished quietly stall, because nobody resolved it early enough for it to be a non-issue later.

The test a synthetic dataset can't pass

The final distance, and the one most frameworks skip entirely, is proving a workflow holds up against real, messy, live data, not a clean test set built to make the demo look good. A first run against genuine data is often where the real problems show up: an edge case nobody anticipated, a formatting inconsistency, a subtle contamination issue that synthetic data would never surface. Real data is undefeated, and any workflow that hasn't been tested against it hasn't actually been tested.

So, what should you be asking instead?

• Have we properly understood the process we're trying to improve?

• Has this been tested against real business data?

• Has governance been addressed from the start?

• How will we monitor and manage it once it's live?

  
  

If you can't answer those questions, a four-box framework won't get you to production.

A question worth sitting with

Before your next AI initiative gets a project name and a kick-off date, ask yourself honestly: do we actually know the distance between where we are and production, or have we just agreed on four boxes and called it a plan?