The channel AI could not pass through

Meta description: Faster AI analysis does not improve business speed when input standards and unnecessary approval layers remain unchanged.

The complaint arrives at my attention in a consistent form: the model is fast, the organization is slow. A team gets an answer in seconds and then watches that answer wait — for a review, for a sign-off, for a forum that meets monthly — until the answer is stale and the moment has passed. Leadership concludes the organization is resistant. I have learned to be suspicious of that conclusion. The model’s speed did not create the queue. It made the queue visible, and the queue was never where everyone said it was.

An AI system can finish an analysis in minutes while the company takes two weeks to make a decision. In a hypothetical organization, a report that previously required three days is generated almost instantly. It then passes through five approval stages, just as it did before. Management reports a dramatic technical improvement. The overall cycle time barely changes.

What does the queue look like from the operator’s chair?

From the operator’s chair, the pattern is mundane. Work is ready and cannot move. It sits in a shared inbox, in an approval chain, in the gap between two teams who each believe the other holds the pen. The operator does not experience this as strategy. They experience it as Tuesday. What strikes me, across companies, is how precisely the operators can describe the queue and how little of that description ever reaches the people who fund the AI program.

Why is the bottleneck rarely the person everyone names?

The first issue I would examine is the baseline. What did the earlier, human-produced analysis contain? Which historical results were used? How were assumptions validated? That material should inform the AI process rather than disappear when the tool is introduced. A useful system needs a meaningful reference point against which a new output can be judged.

Once the organization can establish that the AI-supported output is as good as the earlier approach or better, it can revisit the approval structure. In a five-stage process, I would examine whether those five steps could become three without losing meaningful validation. The number is an illustration, not a universal rule. Its value lies in forcing management to ask what each review stage actually contributes.

The named bottleneck is usually a person — a cautious manager, a slow committee chair. I have watched organizations route around the named person and find the queue intact, because the narrow point was structural: a rule about who may see what, a handoff designed for a slower era, a decision that requires three signatures because one signature was once burned. Removing a person is easy. That is why organizations do it first. It rarely shortens anything.

How do added AI tools lengthen the channel?

Read more: more about Miklos Roth.

The counterintuitive part of this pattern: each new tool adds a node. One more system to check, one more output to reconcile against the others, one more review by someone who does not trust the new source yet. I have seen a company buy three AI tools to speed up a process and end with a process that has five checkpoints where it had three. Faster models, same queue — sometimes a longer queue wearing better clothes.

What pattern repeats in the handoff?

The handoff is where the channel narrows. Inside a team, the AI work moves well; at the boundary, it enters a system built for different work, and it waits. The repeat pattern I have learned to expect is that the handoff is unowned — each side believes the other side’s process is the path. Nobody designed the seam, because the seam sits between two org-chart boxes, and org charts do not design seams.

The difference between quality control and approval theater is whether a stage adds necessary judgment. Does it detect an error the previous stage could not detect? Does it validate a source that the model cannot independently establish? Does it have the authority and expertise to reject the result? If three reviewers make effectively the same inspection, the workflow may be carrying historical habit rather than new assurance.

Control quality is not measured by the number of names on a document. It is measured by what the organization learns about the output before relying on it. A faster process that makes errors harder to notice is not an improvement.

The immediate temptation is to eliminate review because AI has done the work. That confuses speed with reliability. Generative output still requires scrutiny, particularly where financial consequences, customer decisions, or operational commitments are involved. The alternative is not a choice between five signatures and no human control.

What question locates the narrow point?

In these rooms, everyone has a theory about where the queue is, and the theories are usually theories about people. I have sat in those discussions and watched the conversation circle the same two or three names. There is a question that moves the discussion from names to structure, and it is the one that does not get asked, partly because its answer would implicate the room itself.

The decisive question is not how many days the model has saved in isolation. It is which parts of the remaining queue actually improve the business decision, and which are consuming time that could be used to improve the analysis itself?

PR excerpt (80–120 words, Miklos’s own words): An AI system can finish an analysis in minutes while the company takes two weeks to make a decision. A report that previously required three days is generated almost instantly — and then passes through five approval stages, just as before. The overall cycle time barely changes. The difference between quality control and approval theater is whether a stage adds necessary judgment. Control quality is not measured by the number of names on a document; it is measured by what the organization learns about the output before relying on it.