AI on top of a broken process

Definition. This failure pattern occurs when AI speeds up one task while the larger process remains slow, costly or unreliable.

A model can write a summary quickly while the request still waits for an approval. It can classify a form accurately while the source data is entered inconsistently. In both cases, the AI step may pass its tests and the customer may notice no improvement. The problem is not necessarily the model; it is where the model sits in the flow of work.

Follow the work from start to finish

Take a recent request and trace it from arrival to a usable result. Mark the time spent doing work, waiting for someone, moving information between systems and correcting errors. Ask the people at each hand-off what they actually do, including workarounds that are missing from the formal procedure.

Compare three views:

• The AI step: Can the system complete its assigned task accurately and within its boundaries?
• The full process: Has end-to-end time, cost or error rate changed?
• The customer's result: Is the output usable, complete and delivered when needed?

If the AI step improves but the other two do not, investigate the hand-offs around it. Process flow efficiency — productive time divided by total elapsed time — can help show how much of the journey is waiting. It is a diagnostic measure, not a universal pass/fail threshold. See Six Metrics for an AI Process for other measures to pair with it.

Find the constraint

The bottleneck may be a queue for approval, information copied manually between tools, repeated requests for missing data or a rule that sends too many cases to an exception path. AI may make one of these problems more visible by producing work faster than the next step can accept it.

Look at the origin of bad data. Cleaning a dataset once will not help for long if the same inconsistent values continue to enter the process. Similarly, a better prompt cannot remove an approval requirement. Fix the source, hand-off or decision rule that creates the delay.

Ask the process owner to choose the change. It may be a simpler form, clearer authority to make the decision, an integration, a different approval route or a narrower AI task. Some changes need agreement from managers and control functions; include them early enough to shape the operating design. The Frozen Middle article covers what happens when that decision has no clear owner.

Recheck before expanding

Test the revised workflow with cases that represent ordinary work and exceptions. Measure the same end-to-end outcome as before, including quality and rework. If the AI step is good but the process remains poor, keep tracing the work instead of expanding the model's reach.

This differs from a pilot with no exit. A pilot may have good results and no decision about its next step; here the next step may be clear, but the process itself still needs repair. Both can happen at once. A good review identifies which constraint to address first.