Rishi Katdare, Senior Leader in Networking and Edge for Global Financial Services at Amazon Web Services.
In a procurement and supply chain effort I reviewed, the AI ambition sounded practical. An assistant could analyze transaction history, answer procurement questions and support buying decisions.
The plan became more revealing when it moved from ambition to execution. Product descriptions were inconsistent, units of measure were unclear, business definitions were missing and fields required subject matter experts to explain what the data meant. In the proof of concept, the recommendation layer was not the hard part. Most effort sat in data cleaning, business interpretation and agreement on the business question being answered.
That is where many AI business cases become unreliable. They describe what the system will do, but not what the organization will stop doing once the system works.
The missing artifact is not another ROI model. It is a stop doing list. The idea itself is not new. What changes in the AI era is what the list has to prove.
The Stop Doing List Comes Before The Savings Claim
Executives are being asked to fund AI with operational language. Reduce response time, automate repetitive tasks, increase productivity and support better decisions. Those goals are useful, but they do not prove that work will be retired.
The discipline forces leaders to identify the meeting that will end, the report that will disappear, the queue that will shrink, the approval that will narrow and the metric that will prove the work stayed gone.
Without that discipline, a company can improve activity while leaving the operating model intact. A procurement assistant may answer questions faster while the inquiry queue remains. A customer service tool may draft stronger responses while supervisor review expands. A finance assistant may summarize variance drivers while the same reconciliation meeting survives. The company has added a faster layer above work it still refuses to remove.
A company cannot claim productivity from AI while paying for both the old process and the new system.
AI savings should not be booked only from task speed. Speed matters when it removes cost, reduces cycle time, improves control quality or changes the decision path. Speed without removal becomes parallel processing, where the business funds the old way of working and the new layer above it.
Work Retirement Is A Management Decision
A stop doing list turns AI from a technology investment into a management test. Leaders have to name the work, owner, burden, decision it supports and condition for removal.
It is easier to approve a tool than to eliminate a process. It is easier to celebrate adoption than to close a queue. It is easier to add a dashboard than to retire the report it was supposed to replace. The discipline forces the business to confront why the work existed.
I have watched reports survive long after better data existed because the report had become part of the management ritual. The output changed, the dashboard improved and the meeting still happened because the organization had not decided which decision the old report was serving or whether that decision still needed the same ceremony.
Some work will disappear, some will move into exception management, some will shift from execution to review and some will become policy design, data stewardship or control validation. That distinction has to be explicit. A company can automate task work and create review, audit cleanup or exception triage elsewhere. In that case, the business has automated one layer of work and displaced another.
That risk grows as AI moves from producing answers to acting inside workflows. An agent may resolve routine work faster while creating new supervision around exceptions, downstream actions and uncertain cases. If those costs are not counted, the organization can report automation gains while quietly re-creating the labor elsewhere. The stop doing list therefore has to account for both the work removed and the new work created around the system.
Procurement, finance, risk and customer service make this visible because they depend on data meaning, policy interpretation and exceptions that do not fit cleanly inside a workflow. AI can improve those environments, but the business still has to decide which decisions can move faster and which judgment points remain nonnegotiable.
Adoption Is Not The Finish Line
Usage data can show that people are using the system, and execution data can show that an agent completed a task. Neither proves that the operating model has changed. Harder evidence is whether a queue shrank, a report stopped, an approval path narrowed, a review layer was removed or an exception burden declined without new risk elsewhere.
A familiar pattern is AI activity rising while supervision stays intact. Managers still ask for manual confirmation, exceptions still move through the same approval path and no one can name the work retired.
The owner of the AI business case cannot be only the technology team. Technology can provide the model, integration, workflow, guardrails and evidence trail. The business has to decide what work no longer deserves to exist.
That decision is harder than approving the technology. A review step may exist because of an old failure nobody wants to revisit. An approval may protect someone politically more than operationally. A reconciliation may survive because leaders trust familiar manual effort more than a new system output. AI exposes those truths, but it does not resolve them automatically.
The stop doing list is where the economic value becomes real, once the investment connects to removed work, changed controls, narrower reviews, smaller queues and evidence that the old burden did not return somewhere else.
AI will change work, but the value depends on whether leaders are willing to retire work, not only accelerate it. If old work is not removed, savings are not realized. They are assumed.
When the next AI investment comes forward, ask for the stop doing list. What work will no longer exist, who owns the removal and what evidence will prove it stayed gone? If you cannot answer that directly, the business case is incomplete. You have not designed productivity. You have funded another layer of work.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?