If you're working with AI I am sure you have seen the same pattern as I have seen lately. The company you work for invests in AI tools, runs a pilot that works, tries to scale it and the scaling hits a wall. The instinct is to blame the technology.

But I think the technology is usually fine.

McKinsey and BCG both put numbers to what practitioners already know from experience. 92% of companies plan to increase AI investment. But actually fewer than 1% achieve full operational integration. The gap isn't a technology gap but a organizational foundation gap. The data quality, process ownership and organizational structure needs to be designed to carry AI at scale.

I ran into this directly working on a several process documentation initiatives. For example, you run 40+ business processes. There are more than 60 stakeholders across multiple business units to gather the information from, and a monthly review cycle consuming between 320 and 480 hours of senior time. How to scale it with AI?

The first thing I normally do is not starting with AI.

In most organisations, the source data is a problem. Often the data is fragmented across systems and JIRA/MIRO boards. People use inconsistent versioning, or documents in draft being used as if they were final.

Here's what AI does with that: it amplifies it. A model working from inconsistent, outdated, conflicting input produces output that is consistent, current, and confident and wrong. At small scale you can catch the errors, but it intensifies when you have 40+ processes with a lot of stakeholders making decisions based on the output.

So before we designed an AI layer, you need to designed a data layer.

Set up a defined hierarchy of trusted sources. Apply version governance with actual rules. And provide a clear answer to the question nobody had asked yet: what is this AI allowed to use as input, and what isn't it? Only once that was settled, you can design the AI layer. Not a single tool but an ecosystem of tools with specific roles, working in sequence.

A trustworthy and well designed architecture is the foundation of every successful implementation. It's not about the AI model alone; it's about defining which systems it can access, which rules and prompts guide its decisions, and whether the underlying data is structured in such away and reliable enough to trust the outcome.

As the saying goes: garbage in, garbage out. That's why getting the foundation right isn't optional it's essential for scaling AI automations successfully.

Rianne Overwijk, Customer Intelligence Specialist

I also make sure to keep humans deliberately in the loop. This is not workaround but one of my core design principles. The full organizational trust for fully automated output isn't there yet. Pretending otherwise would have produced faster output and worse decisions.

Also keep in mind that the RACI changes as a consequence if you start to scale. When AI takes over parts of a process, somebody has to own what it produces. This means new responsibilities, new review triggers, and new accountability. Ignoring that piece is how companies end up with AI that works and outcomes that don't.

This is organizational design work. It's slower and less visible than deploying a new model. Most programmes skip it because it doesn't feel like progress. Nobody wants to tell the board they spent six weeks on data governance before touching an AI tool. But it's the difference between a pilot that runs and a programme that scales.

If your AI investments are producing results in controlled conditions and stalling in the real organisation, the problem probably isn't the technology. It's what the technology is sitting on.

The Focal Sprint starts by building that foundation. So before the roadmap, before the rollout.