The Uncertainty Tax: A Different Way to Think About AI at Work

Consider a company trying to approve a new supplier. Procurement has the proposal, finance needs to check the commercial terms, and operations wants evidence that the supplier can deliver. Someone also needs to establish whether the contract requires additional review. Information moves between documents, inboxes and meetings. Questions come back, earlier answers get lost, and the person with approval authority eventually receives a summary without all the evidence behind it. The final decision might take a short conversation. Getting everyone ready for that conversation is where the effort accumulates.

This is a useful way to examine AI opportunities: look at what it costs an organization to become sufficiently certain to act. Call it the uncertainty tax. It includes the work of gathering evidence, reconciling conflicting information, checking assumptions and obtaining approval. It can also include the consequences of delay: a project that cannot start, a customer waiting for an answer, or an expert repeatedly pulled away from other work to resolve questions that should already have been answered.

Some of this cost is justified. Consequential decisions deserve scrutiny, and removing review indiscriminately can create problems that are more expensive than the time saved. The opportunity is to understand how much of the effort improves the decision and how much exists because the workflow makes people establish the same facts repeatedly. Checking a supplier’s ability to deliver is necessary work. Three teams independently reconstructing the same supplier history points to a workflow that could be designed more effectively.

Much of the appeal of AI is expressed through faster execution: producing documents, summarizing meetings and drafting responses. There is another possibility worth testing. AI may lower the cost of preparing for a decision by helping people assemble and interpret the evidence they need. In the supplier example, it could extract relevant information from submitted documents, compare it against defined requirements, identify missing evidence and prepare a brief that links each finding to its source. The value would come from making it easier to see what is known, what remains unresolved and what requires someone’s judgment.

That distinction matters because a shorter report does not necessarily create a better decision. A polished recommendation can make weak evidence look stronger than it is, particularly when the reader cannot see how the conclusion was reached. The aim should be to reduce the cost of resolving uncertainty while keeping the remaining uncertainty visible. A useful system must be able to show that a requirement has not been verified, that two sources disagree or that the available information is insufficient to support a recommendation.

Doing this well requires understanding why the work is stuck. Sometimes the missing piece is information: nobody has retrieved a document, checked a requirement or reconciled conflicting records. Better intake, retrieval or AI assistance may help. In other cases, the information is available but the organization has not agreed on what matters. Finance prioritizes price while operations prioritizes resilience, and the supplier meets one team’s standard while failing another’s. Elsewhere, everyone agrees on the facts and criteria, but nobody knows who has the authority to approve the decision.

These situations can look similar when they appear as delayed work in a queue, yet they require different interventions. More analysis will not assign decision authority, and a better summary will not settle an unresolved trade-off. Before introducing another tool, the organization needs to understand what would allow the work to move forward. For a supplier review, that means identifying who owns approval, what evidence they require, which criteria are fixed and which trade-offs call for judgment. It also means deciding what happens when evidence is missing or a case falls outside normal limits.

Once those questions have answers, the role of technology becomes more specific. A completeness check may be handled by a simple rule. Comparing information across varied documents may benefit from AI. Accepting an unusual commercial risk may require a named person with the authority to make that call. These choices determine whether technology removes avoidable effort or simply adds another output that someone must inspect, correct and translate into action.

The controls surrounding the model are equally important. They determine which sources it can use, what it must check, when it must stop and what it is permitted to do with a result. This is what we mean by an AI harness: the instructions, tools, checks and permissions that govern how a model operates within a task. A useful harness preserves the distinction between findings and assumptions, makes missing evidence visible and routes exceptions to an accountable owner. It also records the basis for a decision so the organization can examine what happened and improve the process.

This creates four connected layers. AI can reduce the effort required to resolve information gaps. Harnesses govern how AI performs that work and when it must defer. Workflows connect the resulting evidence to decisions, owners and action. Rivington identifies where redesigning those workflows could create meaningful business value. Each layer depends on the others: a capable model can still sit inside a process that loses context, routes work poorly or leaves decisions unowned.

The economic test therefore has to extend beyond the AI task. If preparing a supplier brief becomes faster but every brief requires extensive correction, the work may simply have moved. If approvals arrive sooner but more decisions are later reversed, speed alone gives an incomplete picture. A useful evaluation follows the whole case, including the effort required to prepare it, the time it spends waiting, the review it receives, the corrections it needs and the outcome that follows.

Even apparently straightforward measures require interpretation. More escalations may mean the system is detecting risks that were previously missed. Fewer may mean routine cases are being handled effectively, or that important exceptions have become invisible. The question is whether the workflow reaches defensible decisions with less avoidable effort while maintaining the quality and controls the business requires.

A practical starting point is one recurring decision with a visible business consequence. Follow a small sample of real cases, including routine work and exceptions, and examine what people needed to know, where they found it, what they checked again and what finally allowed someone to act. That exercise may reveal an AI opportunity. It may also reveal a missing intake field, conflicting criteria or an approval with no clear owner. Each finding provides a more concrete basis for deciding what to change.

At Rivington, we examine consequential workflows through their information, decisions, ownership, handoffs and exceptions. Our AI Workflow Diagnostic turns that examination into a practical redesign and implementation plan. The uncertainty tax offers another way to identify where that work should begin: where does your organization spend the most time and effort getting ready to make a decision, and how much of that effort actually improves the decision?