The Decision Backlog AI Created

This is the most under-discussed effect of AI adoption, and it is showing up in organizations that did everything else right. They chose good tools. They trained people. Usage is real and the output is genuinely useful. What they did not anticipate is that the useful output arrives faster than the organization can decide anything about it.

The result is a backlog nobody planned for and nobody is measuring: a growing queue of choices the organization has been handed and has not made.

Generation capacity and decision capacity are different numbers

Every organization runs on two rates. The first is the rate at which it can produce candidate work: analyses, drafts, proposals, options, recommendations. The second is the rate at which it can commit to that work, meaning the rate at which someone with the authority to close a question actually closes it.

Only the second rate produces outcomes. The first rate produces raw material. An option that has not been chosen has not changed anything, no matter how good it is or how quickly it appeared.

AI raised the first rate enormously. A capable team with modern tools can now generate in an afternoon what used to take a quarter. The second rate did not move at all, because it was never a function of analytical capacity. It is a function of how many people are permitted to decide, how quickly those people can turn their attention to a question, and how much confidence they have that a decision they make will hold.

Throughput in any system is set at its constraint. Capacity added anywhere other than the constraint does not become speed. It becomes inventory. In manufacturing that inventory is visible, it takes up floor space, and everyone can see it piling up. In an organization the inventory is a set of open questions in a shared drive, and it takes up nothing at all until it takes up everything.

Fred Brooks made the same structural argument in 1975 about a different resource. Adding people to a late software project makes it later, because the added capacity has to be coordinated, and coordination cost grows faster than the output the new capacity contributes. The lesson generalizes past programmers. Capacity added on the wrong side of a constraint does not merely fail to help. It imposes a management burden of its own.

What a decision backlog looks like from inside

From the outside, an organization with a decision backlog looks productive. Something is always being produced. From the inside it feels like effort disappearing.

The first symptom is aging. Every option carries assumptions, and assumptions have a shelf life. A recommendation built on this quarter’s cost structure, this quarter’s staffing, this quarter’s competitive position, is a different recommendation four months later. Nobody rejected it. It simply stopped being true while it waited, and now the responsible answer is to redo the analysis, which the organization is very well equipped to do, quickly, producing another option that will also wait.

The second symptom is re-litigation. While an option sits in the queue, the people who built it move on to other work. The person who eventually picks it up was not in the room where the reasoning happened. So the reasoning is explained again. Then it is questioned, reasonably, by someone encountering it for the first time. Then a new consideration is raised, which requires a new pass, which restarts the clock. The organization is not deliberating. It is refreshing its memory, repeatedly, at full cost each time.

The third symptom is duplicated work. When a decision stays open, downstream teams cannot stop working. They make assumptions so they can keep moving, and different teams make different assumptions. Whichever way the decision eventually lands, part of the work built on the other assumption is thrown away. That waste is charged to execution, not to the delay that caused it.

The fourth symptom is the one that surprises people. The queue itself becomes a source of work. Someone has to triage it. Someone has to report on its status. Someone has to answer where a given item stands, and someone has to re-explain the item to the person who asked. None of that activity moves a single decision forward, and all of it consumes exactly the senior attention that the decisions were waiting on.

The last symptom is silent. Options are abandoned. Not rejected, which would at least be a decision, but allowed to expire because no one was ever clearly empowered to choose them and no one is accountable for the fact that no choice was made. There is no line item on any report called decisions not made. The cost is real and the accounting for it does not exist.

Why adding more AI makes it worse

The intuitive response to a stalled decision is to strengthen the case. Run more analysis. Model another scenario. Produce a cleaner comparison. The implied theory is that the decision is stuck because something is not yet known.

Usually nothing is missing. Ask what specific fact would resolve the question, and in a genuinely stalled decision no one can name one. What is missing is a person who is confident they are allowed to close it.

Herbert Simon described the underlying economics in 1971. Information consumes the attention of its recipients, so a wealth of information creates a poverty of attention. Every additional option, however good, draws down the same scarce resource the queue was already competing for. Producing more analysis for a decision layer that is already saturated does not clarify the choice. It taxes the faculty that has to make it.

Raja Parasuraman and Victor Riley documented what happens next in their 1997 study of how people actually use automation. When output exceeds the capacity to evaluate it, behavior splits into two failure modes. Some of it is disused, ignored regardless of quality, because there is no capacity to engage with it. The rest is misused, accepted without real scrutiny, because scrutiny is what there is no capacity for. Both failures come from the same source. Neither is fixed by improving the output.

An organization in this state has purchased the ability to produce more of what it already cannot absorb.

The centralization trap

There is a second-order effect that makes the backlog self-reinforcing, and it is the reason this is an authority problem rather than a process problem.

When decisions concentrate at the top, the natural response is to give the top better information. AI is exceptionally good at that. The executive who was already deciding everything now decides everything with better analysis. Objectively, the decisions may improve. Structurally, the constraint tightens, because the organization has just made its most concentrated decision point more capable and therefore more justified in staying concentrated.

When an organization deploys AI into an environment where authority is concentrated, the technology becomes another mechanism of centralization. The senior leaders who already controlled every decision now have more data to support those decisions, and the people closest to the work have even less reason to develop judgment of their own.

From Builders Build, Chapter 9

That last clause is the compounding cost. Judgment is a practiced capability. Melvin Conway observed in 1968 that a system reflects the communication structure of the organization that produced it, and the same is true of the decision paths an organization builds into its tools. Route every question upward and the tooling will encode that route, which makes the next question harder to answer anywhere else.

Lisanne Bainbridge named the resulting irony in 1983. Automating the routine portion of a task leaves people responsible for exactly the cases the automation cannot handle, while removing the everyday practice that built the skill those cases require. The same irony operates on decision-making. An organization that routes all judgment upward is training its middle to have none, and it will need that judgment most on the day the queue is longest.

Who actually has to decide

Every organization has two authority structures. One is written down. The other is what happens.

The written structure lives in delegation matrices and approval thresholds. It says a director can commit up to a certain amount, a vice president above that, and it is usually accurate about the extremes. The actual structure is revealed by history. It is what a director does when a decision inside her stated authority carries any visible risk, which in most organizations is to route it upward for cover, because a decision that went badly two years ago taught everyone what the real threshold is.

The gap between those two structures is where the backlog forms. The formal owner does not decide because she does not believe the authority is real. The escalation target does not decide quickly because the question was never supposed to reach him and he now has forty like it. Both people are behaving reasonably. The queue grows anyway.

There is a straightforward diagnostic. Take a category of decision that is currently backed up, list the last five of that type, and write down who actually made each one. Not who was authorized. Who decided. If the answer is the same person for all five, and that person is two or more levels above the stated owner, the delegation matrix is a description of an organization that does not exist.

The backlog takes two shapes depending on which structure fails. When knowledge sits at the front line and authority sits at the top, decisions wait on people who need to be briefed before they can act. When authority formally sits with someone who does not believe they hold it, decisions wait on nobody at all, which is the more expensive version, because there is no one to chase.

Raise decision throughput first

The sequencing matters more than any individual technique. An organization that raises generation throughput before decision throughput pays for the same output twice, once when it is created and again when it is refreshed after aging in the queue. Take the constraint first.

Start by counting the queue. List every decision currently open, how long it has been open, and the name of the person who is supposed to close it. Most organizations cannot produce this list on request, and the inability to produce it is the finding. A backlog that has never been counted has never been managed. The count also settles arguments quickly, because the age column is difficult to explain away.

Then assign every item a name and a date. A decision without a named owner is not pending, it is unassigned, and unassigned work does not age gracefully. The date matters as much as the name. An open decision with no deadline competes for attention against work that has one, and it loses that competition every week.

Give each open decision a default. This is the highest leverage move available and the least used. If a decision has no default, silence resolves to nothing and the item stays in the queue indefinitely. If a decision has a default outcome attached to a date, silence resolves to the default, and anyone who disagrees has to say so on the record before the date. That converts an indefinite wait into a bounded one and surfaces genuine disagreement, which was often the real reason for the delay.

Push what can be pushed. Sort open decisions by how hard they are to reverse. Decisions that can be undone cheaply belong with the people who hold the knowledge, and they should be made there without escalation. Decisions that are genuinely hard to undo belong higher, and they deserve the senior attention that is currently being spent on the reversible ones. Most decision backlogs are made of reversible decisions being handled as though they were permanent.

Finally, cap generation until the queue is moving. This is the least popular recommendation and the most reliable. Running more analysis into a saturated decision layer produces cost, aging, and re-litigation. Holding generation steady while the backlog drains costs nothing except the appearance of activity.

None of this is an argument against the technology. These systems are producing genuinely valuable work, and the organizations that get compounding returns from them will be the ones that can act on what they produce. The constraint is not the model. It is the number of people who are permitted to say yes, and how quickly they can say it. Fix that first and everything else the technology can do becomes available. Leave it alone and the organization has bought a faster way to fill a queue.

Sources