The Scope She Agreed To at 6 PM on a Friday was not a project management failure.

Quick answer

Last week of August I had three one-hour training sessions, a production deployment, two kickoff calls, and live production bugs to firefight, all on the same Wednesday. I knew going in it was going to be that kind of day, so I had cleared most of my Monday calendar just to pace myself. I skipped the gym on Wednesday deliberately. I got everything done. I pushed all the way through to Saturday. And then the following week I got sick. That is the cost I can name. The other cost is quieter. During that stretch I said yes to things I did not have the cognitive capacity to evaluate clearly. Some of that work now lives in systems without anyone formally owning it. I do not know exactly where all of it landed. I am not sure I want to know. I am telling you this because the Friday moment this post is built around, the one where a VP types a boundary and then deletes it fifteen seconds later, is a very, very ordinary failure. And it keeps producing expensive problems downstream because no one is naming what actually happened in those fifteen seconds.

TL;DR

The research on scope, depletion, and AI readiness points to one compounding problem.

• Scope creep is the majority experience

  PMI's 2018 Pulse of the Profession report found that 52% of projects experienced scope creep, up from 43% five years earlier (PMI, 2018). A separate analysis of 1,471 IT projects found an average budget overrun of 27% on projects that crept (Flyvbjerg & Budzier, Harvard Business Review, 2011).

• Depletion measurably degrades the decisions that let it in

  High workload causes decision fatigue by depleting cognitive resources, forcing reliance on heuristic thinking rather than deliberate analysis (Frontiers in Cognition, 2025).

• The undocumented process is the AI readiness problem

  Only 16% of knowledge workers say their workflows are extremely well-documented, and 49% say undocumented or ad-hoc processes impact efficiency at least sometimes (Lucid AI Readiness Report, 2025).

• Tribal knowledge is a production failure, not a cultural one

  Gartner's 2025 research predicts that organizations will abandon 60% of AI projects that lack AI-ready data through 2026, with 63% of organizations reporting they do not have, or are unsure whether they have, the right data management practices to support AI (Gartner, 2025).

• The person running the implementation is often too burnt out to see it

  55% of the U.S. workforce was experiencing burnout as of late 2025 (Eagle Hill Consulting, 2025).

• Projects without formal change management are 35% more likely to exceed cost or miss deadlines (PMI, 2025)

  Scope agreed to informally, verbally, at 6 PM on a Friday, is the definition of a change with no management at all. Projects without formal change management are 35% more likely to exceed cost or miss deadlines (PMI, 2025).

What actually happened at 6:12 PM

She had been in back-to-back calls since 9 that morning. Three coffees. No real break. When the VP of IT dropped the message in the project channel, 'Quick ask, can we fold the store-level exception reporting into this sprint? It'll be faster than standing up a separate workstream,' they were on their third read of a deck they should have approved two hours earlier. They typed 'sure, let's talk Monday about scope.' Fifteen seconds passed. They deleted it. They typed 'yes, makes sense.' They closed their laptop at 7:04. From the outside that looks like a reasonable decision made by a senior leader who understood the project. From the inside something else happened: a person whose cognitive resources were spent, defaulting to the path that required the least resistance in the next thirty seconds. Research on decision fatigue is consistent here. When a leader has been making decisions across a long, high-demand day, the quality of subsequent decisions deteriorates. The mechanism is neurological. Making choices depletes the same mental resource used for self-control and active initiative (Vohs, Baumeister, et al., published in multiple peer-reviewed series). Depleted, people default. They simplify. They approve. They type 'yes, makes sense' and close the laptop. What I want to be clear about is what happened in those fifteen seconds. They wrote the right answer. They deleted it. That is not a process failure. The process was working. The person running it was past their capacity to hold what the process required.

The 27% overrun starts with a yes that had no owner

PMI's 2018 Pulse of the Profession report put a number on scope creep: 52% of projects experience it, up from 43% five years prior (PMI, 2018). A separate analysis of 1,471 IT projects found an average budget overrun of 27% on the projects that crept (Flyvbjerg & Budzier, Harvard Business Review, 2011). Neither figure shows where in the project the boundary moved, or what state the person was in when it moved. Scope creep is most commonly defined as unapproved work added after a project has started. The approval, in practice, often happens the way it happened on that Friday: a Slack message, a depleted leader, a sentence typed in the project channel that no one will ever find again during a change control review. Projects without formal change management are 35% more likely to exceed cost or miss deadlines (PMI, 2025). That figure typically gets read as a process argument: document your change requests, run them through a review board, require sign-off. Correct, and insufficient. The process did not fail on Friday. The person running it had already passed the point where they could hold a boundary, and the process had no way to know that. Store-level exception reporting is not a small add. In a retail operations context, it sits at the intersection of inventory, supplier data, and store-level POS reconciliation. Adding it to a sprint mid-cycle means someone on the team will carry it. The question is who, and whether that person was assigned or simply absorbed it because no one else did.

The work that follows a depleted yes is the hardest kind to audit

When scope expands formally, through a change request and a revised statement of work, it leaves a record. Someone owns it. There is a ticket, an assigned resource, a revised timeline. The work is visible. When scope expands the way it expanded on that Friday, none of that happens. The store-level exception reporting goes into the sprint. Someone picks it up, probably the person who has been at this retailer long enough to hold all the legacy knowledge. They figure out how it connects to the existing workstream. They build a workaround for the part that does not quite fit. They do not document the workaround because they did not have time and because it seemed obvious in the moment. Three weeks later, the workaround is the process. This is the pattern that makes undocumented work so durable. It does not advertise itself. It runs quietly inside what looks like a functioning operation. The Lucid AI Readiness Report (2025) found that 49% of knowledge workers say undocumented or ad-hoc processes impact efficiency at least sometimes, and 22% say it happens often or always. Only 16% describe their workflows as extremely well-documented. That gap between how processes actually run and how they are supposed to run accumulates through scope decisions made in project channels at 6 PM. Each one starts with a yes that had no owner.

Why this is specifically an AI readiness problem

This month's argument is about AI readiness. Scope creep is the entry point. When an organization brings an AI implementation into a retail operations function, the implementation needs to read the process underneath it. If the process is documented, the AI reads the documentation. If the process is undocumented, the AI reads whatever version of it gets captured first, and that version is almost always cleaner than the real thing. Fluid Labs' analysis of 2025 enterprise AI abandonments names this as the most common failure mode: organizations picked the AI use case before understanding the process underneath it. Halfway through the project, they discovered that what they were automating was never actually a process. It was a collection of individual judgments and informal workarounds that had never been systematically captured. McKinsey's 2025 State of AI research found that organizations classified as AI high performers were 2.8 times more likely than average to fundamentally redesign workflows when deploying AI, rather than layering AI on top of existing workflows (McKinsey, State of AI, 2025). Gartner predicts that through 2026, organizations will abandon the majority of AI projects that lack AI-ready data, and their survey of data management leaders found that 63% of organizations either do not have or are unsure whether they have the right data management practices to support AI (Gartner, 2025). Atlan's Context Quality Testing (2025) found a 38% improvement in AI accuracy across 522 queries when governed, structured metadata was added to agent context versus unstructured retrieval alone. The exception reporting workaround the team built three weeks after the Friday yes is exactly this kind of missing context. It is real. It governs real decisions. It is invisible to any system that did not watch someone build it at 9 PM on a Thursday.

The self-betrayal the governance system cannot catch

The standard prescription for scope creep is process: document the change request, require formal approval, build the governance muscle. That is correct. But consider what actually happened on that Friday. They had a process. They wrote the boundary. They deleted it. Fifteen seconds. The process existed and it made no difference, because the person holding it had already consumed the deliberate thinking capacity that the process required of them. Research on decision fatigue points to this precisely. A 2024 study of Arkansas traffic courts (Hemrajani and Hobert, Journal of Law and Courts) found that in high-volume arraignment hearings, dismissal rates declined significantly as a session progressed without a break. The quality of judgment degrades in a predictable direction, toward the default, toward whatever requires the least active choice in the next thirty seconds. For a VP who has been in back-to-back calls since 9, the default at 6:12 PM is the residue of a day that has already spent what deliberate thinking they had. I think about this when I look at my own August. I cleared Monday to pace myself for Wednesday. I made a conscious decision to skip the gym. I still got sick by the following week. The prep helped. It was not enough. And I am not sure I have fully figured out what enough would have looked like, because the demands that week were real and the deliverables were not optional. I do not have a clean answer for that. What I have is the observation that the decision I most needed to protect, the one that would have mattered most to the work downstream, was the one I had the least capacity to protect by the time it arrived. Gartner's survey found that 63% of organizations either do not have or are unsure whether they have the right data management practices to support AI (Gartner, 2025). Those practices reflect decisions made by people, across thousands of Fridays at 6 PM.

What compounds over the life of the project

One depleted yes does not break a project. The compounding does. The store-level exception reporting goes in. The workaround gets built. The workaround becomes the process. Four months later, when the AI implementation begins its process mapping phase, someone documents the process as it runs now. They document the workaround without knowing it is a workaround. The implementation is built on top of it. When the output is wrong, the team diagnoses a data quality problem. Then a model problem. Then a prompt problem. The actual problem, the undocumented exception logic that has been running inside the reconciliation column for six weeks, does not appear on any diagnostic checklist because no one knew to look for it. Gartner predicts that through 2026, organizations will abandon 60% of AI projects that lack AI-ready data (Gartner, 2025). The abandonment will be diagnosed as a technology failure. The actual cause will be that the process was never clean enough to automate, and the person who would have known that was too depleted on a Friday afternoon to say no to one more thing. A 2025 study of Korean workers found that over-extension of work hours mediated approximately 51% of the total effect of occupational stress on health-related productivity loss (Kim et al., 2025). A separate Swedish study quantified the lasting economic impact of clinical burnout at a permanent 12.36% earnings reduction two years post-diagnosis (Gaspar et al., 2024, via Emory Economics Review). The person whose judgment is most at risk is the one the organization has been relying on the longest. Here is where it comes back to that Friday. They closed their laptop at 7:04. The message they sent at 6:12 is still in the project channel. Nobody will trace what happened next back to that message, because it looked like a normal project decision made by a senior person who knew what they were doing. And in a way, it was. That is the part that is hard to sit with.

Who carries it, and who knows where it lives

The person most likely to know where the workarounds live is also the most likely to be the one who agreed to the scope expansion that created them. Those two facts are connected, and the connection is what makes this hard to audit from the outside. When lean-team members absorb undocumented scope, they rarely surface the capacity gap in a status report. The VP approved the add-on. The analyst carries it silently. The sprint completes on paper. The gap does not appear until something downstream breaks and nobody can trace it. An AI implementation built on top of an undocumented workaround does not surface the problem. It gives the workaround a faster engine. The investigation that follows will look at the model first, then the data, and last at the process underneath, because no one documented that the process had a workaround to begin with. The depleted leader who typed yes at 6:12 PM is now the single point of failure for any audit that needs to trace what happened. She holds the institutional memory of a process she agreed to under depletion, in a project channel, in fifteen seconds, on a Friday.

What I have learned from twenty years of retail operations

Twenty years of retail and consumer operations, and the one thing I can tell you with confidence is that the broken process almost always has a Friday at 6 PM somewhere in its history. Someone, past capacity and under pressure, said yes to something they half-understood, and the thing they said yes to became load-bearing before anyone noticed. AI readiness runs on one question: does the person signing off on an implementation actually know what her team does to make the numbers look right? Most leaders cannot answer that. By the time the operations report lands in their inbox on a Monday at 8:47 AM, they are three minutes from their next call. The summary line matches what they expected. The intention to ask about the supplier reconciliation column disappears the moment the calendar opens. The session I cleared on Monday before my brutal Wednesday, that was the right instinct. Protect the cognitive state before the high-stakes day, not during it. The scope decision you most need to get right will arrive when you have the least left to give. You cannot always control that. You can control when in your day you make irreversible commitments. The one thing I would ask you to do is small. Before you type yes into a project channel, ask yourself one question: who carries this by Thursday? Name the person. If you cannot name them, you have not made a decision. You have moved the decision to someone with less context.

Key takeaways

One ask. One question worth repeating in your own head at 4:40 on a Tuesday: who carries this by Thursday?

• The one thing

  Before you type yes into a project channel, ask: who carries this by Thursday? If you can name the person, you have made a decision. If you cannot, you have moved the decision to someone with less authority and less context than you have.