Perspectives

When AI creates capacity, where does the saved time go?

István Borbíró · [publish date] · 5 min read

 

Companion piece to Vera Damerjian Pieters, How Absolute is the "Ground Truth" in Medical AI Annotation? [linked to the article]

A model can be technically sound and still go off at the edges.

Vera Damerjian Pieters, founder of VeraDP, has written about where that begins. At the data stage.

In the quiet decisions made when a tumour boundary is genuinely unclear, when a label is set at eighty percent confidence, when the person annotating image four thousand is no longer the person who annotated image one. She shows how unmanaged ambiguity is born there, and how it travels downstream, surfacing later where the model shows its limits.
When the model shows its limits, someone is standing there.

It is rarely the person who built the tool. Often it is not even the clinician the tool was built for. It is whoever has to reconcile a confidently generated output with a real patient who does not fit it. In cancer care, that is usually the person holding the team together. And the gut feeling does not have a risk value attached.

The AI was meant to create capacity and save time, as every technology before it was meant to. So the question remains.

Where does the saved time go?

The problem before the pipeline

Before any AI pipeline begins, there is a problem in the clinical world it is meant to solve. That problem is usually under-defined, and solved by whoever was closest to it. In the worst case the goal names the innovation we need to use, without matching it to a specific problem we need to solve. We need to use AI. That is not good enough.
Think of cancer care as three layers. The data layer is accelerating, more test results, more detailed health records, more signals, piling up faster than ever before. The clinical layer advances at human pace, because decisions still need synthesis, judgement, conversation, and because treatments can act at the speed of human biology. The coordination layer connects the two. It was never designed. It was inherited. It runs on informal knowledge, personal relationships and workarounds built up over years.

When the data layer speeds up and the clinical layer cannot match it, the coordination layer absorbs the difference. Until it cannot.

Most implementations do not fail because the technical chain was wrong. They fail because this prior problem was never named, and because the structure the system would land on was never in the room. Talking to the clinical team usually means talking to the two physicians who are senior and have five minutes available. The coordination that actually carries the system is mostly implicit, held by roles that were never designed, asked, or trained, only appointed.

Time does not compound. You only spend it once.

Now the part we rarely examine.

When a tool makes one node faster, the time it frees does not bank as a benefit. It moves.

The same tool that makes a clinician more efficient also makes them more interruptible. It raises what the people around them have to absorb. There is solid evidence for this. Ter Hoeven CL, et al. (2016) studied communication technology directly and found it runs along two pathways at the same time. One is a resource pathway, efficiency and accessibility, feeding work engagement. The other is a demand pathway, interruption and unpredictability, feeding burnout. The two are comparable in size. 

So technology is never only a resource. It is a resource and a demand in the same motion. What decides which one wins is not the tool. It is the coordination capacity of the team it lands on.

The minutes saved at one desk are easy to count. The coordination work that flows downstream from every faster output is counted nowhere.

Who absorbs the output

This is where the opening question finds its answer.

The load-bearing absorber in cancer care is not an individual. It is the multidisciplinary team as a coordination structure, distinct from the people who use the tool.

The coordinator who was never in the room when the tool was chosen still inherits its edges. The named but absent role still deals with the outcome of a system that was good up to a point and then went off. They absorb the gap in a structural sense, whether or not they ever touch the interface.

This is the part most implementation conversations miss. They optimise the node and assume the team will absorb whatever the node produces. Sometimes it does. Sometimes the absorption turns a workable tool into a burden. When that happens, the failure gets attributed to personality, to culture, to resistance. The real cause stays invisible. The coordination layer was never specified.

A system does not arrive into empty space. It lands on a coordination structure that already exists, designed or not. It amplifies what it finds, and it moves the work. Capacity or fragmentation, it makes more of what is already there, and it puts the load somewhere new.

The frontier is not another tool

So the next frontier in AI for cancer care is not a better model. It is understanding and designing a measurable human infrastructure that decides whether a model creates value or friction.

Vera's piece shows where this starts, at the data stage, where ambiguity is born and sent downstream. This piece is about what happens when it arrives. Her end of the pipeline and mine, closing on each other.

If you carry coordination in a cancer team, you already know the feeling of inheriting an edge you did not design. The question worth asking, before the next technology lands, is a plain one.

What will the next tool move, and where?

Reference: Ter Hoeven CL, et al. Communication Monographs. 2016;83(2):239-263.