AI operating governance · Field note

AI adoption starts with the constraint, not the tool

TL;DR

When an AI request arrives as a tool, the question of where the value should appear is left for later. The paper recommends naming the recurring work and the step that most limits it, then funding the smallest change that moves that step, whether or not it uses AI. It gives a way to find that step, five steps with a named decision-maker for each, and the cases where this approach misses value.

What the paper develops

A leader on a steering group, the group that decides where money and people go, receives an AI request. A team wants to try a tool, a model, or an agent. A pilot starts, people use it, and a dashboard reports activity. The question of where the value should appear tends to come later. That sequence is this paper's observation of how such requests arrive. No source counts how often it happens.

McKinsey's 2025 survey found that 88 percent of respondents regularly use AI in at least one business function, and 39 percent reported an effect on earnings across the enterprise. RAND interviewed 65 practitioners, who named misunderstood problems and technology-first choices among the causes of failed AI projects. Gartner reports that at least half of generative AI projects were abandoned after proof of concept, with unclear business value among the causes. A survey cannot show cause, the interviews are not a count of projects, and none of these sources tests the recommendation that follows.

Before anyone picks a tool, ask which recurring work the request would change and what most limits that work today. A request names a task the tool would speed up, and the limit may sit somewhere else. This paper calls the limiting step the constraint: the step that most limits what the work delivers. The idea comes from Goldratt's Theory of Constraints, which holds that a system's output is limited by its constraint, so speeding up any other step changes nothing. The paper borrows the idea for AI requests and does not prove that it holds for office work.

Compare changes to the workflow, the data, the people, fixed rules, vendor features, and AI. Choose the smallest one that can move the constraint, and fund any missing prerequisite first. If nobody can name the constraint, follow about ten recent cases and pick the step where cases lose the most time, are redone most often, or wait for a scarce person. That step is a candidate, checked with the people who do the work.

A gate is a step that ends in a written output and a named decision-maker. The business owner, the leader accountable for the work's results, names the constraint at gate 1. A group of four chooses the change at gate 2, and the business owner breaks ties. The business owner and the people running the test design it at gate 3. The steering group decides at gate 4 whether to expand the change, revise it, stop it, or try a different one, and a missing result expands nothing. The business owner keeps checking value at gate 5 once the change is part of normal work. A one-page record, the evidence pack, holds each step. The paper's planning estimate for a small test is a few hours of four people's time before anything is built.

A constraint-first path can miss uses that nobody yet sees as a constraint, so the paper keeps a small, capped allowance for exploration. The gates take time, so a low-risk use that is easy to reverse needs only a short conversation. A factory idea can mislead in an office, so gate 5 checks whether the right constraint was named.

Before the next steering meeting, replace one tool name with the recurring work it is meant to change. Write the constraint, the baseline, the owner, and one result that would justify another funding step, then list the simpler changes that could move the constraint. Start with the constraint. Then choose the tool.

What to do next

Take one AI request before the next steering meeting. Replace the tool name with the recurring work. Write the step that limits it, the baseline, the owner, and one result that would justify a next funding step.

Inside the white paper

  • How to find the step that limits recurring work
  • How to compare AI with simpler changes first
  • Five gates, their owners and costs, and objections

Sources and notes

  1. Antoine Montard, Dago Diedrich, and Tanguy Catlin, "Where AI will create value—and where it won't," McKinsey Quarterly, April 29, 2026 — McKinsey argues that the larger value from AI lies in reshaping offerings and business models, and that executives have focused largely on pilots seeking incremental productivity gains. It is an argument, not a measured finding.
  2. Alex Singla, Alexander Sukharevsky, Bryce Hall, Lareina Yee, Michael Chui, and Tara Balakrishnan, "The state of AI in 2025: Agents, innovation, and transformation," McKinsey & Company, November 5, 2025 — McKinsey's survey reports that 88 percent of respondents regularly use AI in at least one function, that nearly two-thirds have not begun to scale AI across the enterprise, and that 39 percent report an effect on enterprise earnings. It is a survey, so it cannot show cause.
  3. James Ryseff, Brandon F. De Bruhl, and Sydne J. Newberry, "The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed: Avoiding the Anti-Patterns of AI," RAND Corporation, 2024 — RAND interviewed 65 experienced data scientists and engineers about failed AI projects and names five root causes, including misunderstood problems, models that did not fit the business workflow, and a focus on technology over a real user problem. These are interviews, not a count of projects.
  4. Arun Chandrasekaran, "Why 50% of GenAI Projects Fail - And How to Beat the Odds," Gartner, January 26, 2026 — Gartner reports that at least half of generative AI projects were abandoned after proof of concept by the end of 2025, with poor data, weak risk controls, rising cost, and unclear business value as causes, and says projects without success measures become easy targets when budgets tighten. The page says Gartner analyzed hundreds of implementations and shows no method.
  5. Zeynep Tuğçe Şimşit, Noyan Sebla Günay, and Özalp Vayvay, "Theory of Constraints: A Literature Review," Procedia - Social and Behavioral Sciences 150 (2014): 930-936 — A 2014 open-access review of the Theory of Constraints: every system has at least one constraint, the system's output is limited by it, and fixing one makes another step the limit. It covers production and other systems, not AI, so the paper borrows the idea.
  6. Johannes-Tobias Lorenz, Joshan Cherian Abraham, Robert Levin, and Douglas Ziman, "From promise to impact: How companies can measure and realize the full value of AI," McKinsey & Company, April 24, 2026 — McKinsey says many organizations deploy AI in ways more visible than valuable, recommends a single evidence pack that tracks benefits and total cost, and separates five layers of measurement. It is a consultancy's framework, not a tested standard.
  7. NIST, "AI Risk Management Framework Core," excerpt from AI RMF 1.0, 2023 — NIST's framework asks that the people who set the context for an AI system bring broad domain and user expertise, and that someone decide whether a system meets its purposes and should proceed. It is voluntary and does not prescribe this paper's gates.