AI · 2026-08-04 · 12 min read · by Juan Carlos Zuloaga
Transformation programmes fail in a predictable order. First the ambition is set too wide. Then the pilot has no owner. Then the results cannot be measured, so the budget quietly disappears at the next planning cycle.
A roadmap that survives is narrow, sequenced, and tied to numbers the CFO already tracks.
Here is the four-phase structure we use, and what goes wrong in each phase.
You cannot claim improvement without a before.
Measure four things for the processes you intend to touch: time per unit, error rate, cost per unit, and cycle time from request to delivery.
Most companies discover in this phase that nobody agrees on the current numbers. That disagreement is the first finding, and it is worth the two weeks on its own.
Failure mode: skipping the baseline because "we know it's slow." Six months later you cannot defend the budget.
Choose a process with high volume, low regulatory risk, and a single owner. Ship a change. Measure it against the baseline.
The purpose is not the saving. It is credibility. Internal scepticism about AI is rational — most people have watched a tool get bought and abandoned. One working thing changes the conversation more than any presentation.
Failure mode: picking the most strategically important process first. It is also the most political, and it will not ship in six weeks.
This is the phase nobody budgets for and everybody needs.
AI performs on structured, accessible, reasonably clean information. Most mid-size companies have the information spread across inboxes, spreadsheets, a CRM nobody updates, and one person's memory.
The work here is unglamorous: define the fields, fix the forms, connect the systems, agree on one source of truth per entity. In our Clean Clean and Chassé Dance Studios engagements, this phase is what made everything afterwards cheap.
Failure mode: treating data work as an IT project rather than a process decision. The choice of what to record is a business choice.
Now you turn the one win into a repeatable pattern.
That means: documented prompts and workflows, an internal owner per system, a monitoring habit, and a decision rule for when a human takes over.
Add a simple review cadence — monthly, thirty minutes, look at the numbers, kill what is not working. Systems without a kill switch accumulate quietly until nobody trusts any of them.
Only now do you widen scope. Take the pattern that worked and apply it to the two adjacent processes that share the same data and the same owner.
Resist the temptation to scale across departments at once. Different departments have different data quality, and the pattern rarely transfers cleanly.
The honest scaling rule: expand along data boundaries, not org-chart boundaries.
You need three documents, not thirty.
An acceptable-use note — what staff may put into which tools.
A data map — what personal data exists, where it goes, and the legal basis. In the EU this is not optional.
A decision log — what was tried, what happened, what was retired.
The decision log is the one most companies skip and most regret. It is the institutional memory that keeps you from rebuilding the same failed thing in eighteen months.
Not a transformed company. A company with three or four processes that run measurably better, an internal team that knows how to run the loop, and a leadership group that can tell a real use case from a demo.
That is a foundation. Companies that reach it compound. Companies chasing a step change in year one usually have nothing running by year two.
If you want to pressure-test your own roadmap, our AI and digital transformation service page explains how we run this: https://inspiralgrowth.com/services/ai-digital-transformation