Why GenAI struggles to replace your process intern and what actually works

Better context is an incomplete answer

Have you ever noticed something odd?

The same models that can solve extremely difficult exams, write code, reason about physics, and summarize dense legal texts often struggle to replace a junior operations analyst following your process.

Your first instinct might be to say, “Ah, the model is missing context.” But that explanation is too convenient, and more importantly, it does not actually solve the puzzle.

So what is the missing ingredient that turns GenAI brilliance into real automation?

The answer lies in scaffolding, and not just technical scaffolding that shapes the system. Equally important is non-technical scaffolding that shapes human behavior around the system.

Let me explain.

Anyone who has spent time in real operations knows that actual processes are far messier than what appears in your process maps or SOPs. They are full of exceptions, shortcuts, unwritten rules, judgment calls, and “this is how we do it here” habits that rarely make it into documentation.

GenAI experts do not usually come in to fix this mess. In fact, their biggest contribution is often to reveal it.

How? By building a GenAI proof of concept based strictly on your documented process, only to find that it matches human decisions for a surprisingly small fraction of real cases. Suddenly, what looked clean on paper looks chaotic in production.

The instinctive reaction at this point is predictable:
“See? It doesn’t work. Maybe the next version of ChatGPT will be smarter. Let’s move on to another POC.”

Except that this misses the real insight.

What the low match rate is actually telling you is that your process map was incomplete. The nuances, exceptions, and tacit knowledge never surfaced in the one-hour demo where an analyst walked the GenAI team through a few happy paths and a couple of edge cases.

At this stage, many organizations jump to a neat-sounding solution.

“Let’s fix the process first. Make it consistent. Document everything from L1 to L5. Then feed that context to GenAI. Problem solved.”

That sounds reasonable until you try doing it.

Anyone who has seriously worked on process improvement knows this can take months, years, or sometimes forever. Many “process gaps” are actually policy gaps in disguise. Others are pragmatic exceptions that keep the business moving under time pressure or incomplete information. Cleaning all of this up before touching GenAI is rarely realistic.

So what is a more practical way forward?

Instead of treating GenAI automation as an all-or-nothing, end-to-end transformation, think in salami slices.

A workable playbook looks like this:

1. Observe the real work, not just the process map.
   Use process analytics, but also do Gemba. Sit with analysts, watch how they actually make decisions, and listen to why they do what they do. This surfaces the “rules of the jungle.”

2. Understand what today’s GenAI models are actually good at.
   More often than not, current models are capable enough to handle a meaningful portion of what humans already do, just not everything.

3. Pick the cleanest segments first.
   Identify parts of the process where decisions are relatively consistent and rules are clear. Automate those first. Your match rates here will be high because these are genuinely straightforward cases.

4. Adjust policy and process to support partial automation.
   Do not wait for perfect process. Improve it incrementally around what you automate.

5. Repeat.

In practice, what has worked for us is a simple but powerful “cheat code.” Separate the straightforward cases from the messy ones. Let GenAI handle what is already consistent and high-agreement. Let humans continue managing the judgment-heavy, ambiguous cases, but make those cases visible and learn from them over time.

One slice at a time, before the next iteration - process gets cleaner, the automation gets broader, and the overall system becomes more reliable.

In my next article, I will describe the scaffolding patterns, both technical and non-technical, that have made this approach work in practice.