Process first, prompt second: why automation fails without a redesign

Many companies treat artificial intelligence as a magic button that will one day make everything faster, cheaper and smarter.

The reality is different. AI projects usually fail not because the technology is poor or the tool was wrong, but because someone tried to automate a messy process.

If information goes missing along the way, one person makes all the decisions, and everyone does it differently each time, AI will only deepen that same chaos — just faster.

Why does the process have to come first?

Imagine you want to deploy a chatbot for customer service. It can answer questions, route requests, even suggest a product.

But if the company has no clear rules about:

• where a customer request gets logged;

• who picks it up when the bot can no longer help;

• what information has to be passed on so an account manager can continue;

• how to measure whether the conversation succeeded;

then the bot won’t help. It will simply create one more channel where information gets stuck.

So the first question is never “which AI tool should we use?” but “how does this process work today, and what do we want to change?”.

What does a good automation candidate look like?

Not every process should be automated first. You can spot the best candidates by a few traits:

• The process repeats. It happens often, in much the same way, and doesn’t need a unique decision every time.

• The incoming data is structured. Emails, forms, orders, documents — they have clear fields that can be read.

• There is a clear rule for what happens next. If someone has to ask their manager “what do we do in this case?”, the process still depends too much on human judgement.

• The result is worth measuring. Hours saved, faster responses, fewer errors — there has to be at least one clear metric.

Preparing quotes from standard components, for instance, is an excellent candidate. Analysing unstructured customer complaints is not.

How do you prepare a process for automation?

First you need to understand how the process works today. Not formally, but in reality — what each person does, where information gets stuck, where time is lost.

Here is a simple approach:

1. Write out the steps. Describe how the process runs from start to finish. Not the ideal version — the actual one.

2. Audit the data. Where information arrives, where it is stored, in what form, and whether it can be read reliably.

3. Decision points. Where a person has to decide, and where a rule could be written instead.

4. Problems and a measure of benefit. How long the process takes today, how many errors occur, what you want to achieve.

This work can look dull, but it is what separates successful projects from expensive experiments.

The most common mistakes

Across projects we keep seeing the same mistakes. Here are a few worth avoiding:

• Automating too much at once. One small process that genuinely works beats three that half-work.

• Believing AI understands context on its own. AI can’t see context you can’t see either. If you don’t know what matters, neither will it.

• Forgetting the human check. Automation needs a clear point where a person reviews the result and makes the final decision.

• Not measuring before and after. Without a baseline you won’t be able to say whether the project paid off.

Where to start this week

If you are weighing up which process to automate first, don’t start from a list of tools. Start from one question:

“Where do we lose the most time on repetitive work today?”

Then find that process, describe it as it actually is, and only then ask whether AI can help.

Automation doesn’t cure disorder. It accelerates it. So let’s fix the process first.

Once the process is clear, your team needs to understand it too. That is where AI training for companies built around your real work helps.