All insights
Strategic Brief·11 min read

She Knew the Pilot Was in Trouble in Week Three. She Did Not Say So Until Week Eleven.

A leader diagnoses a failing AI implementation weeks before they escalate it. The delay runs through a calculation they make under depletion: what speaking up costs today, measured against what staying quiet costs the team waiting for them to move first.

By Jessica Caresse White·
Abstract timeline showing two diverging paths between an early signal point and a late escalation point.

Quick answer

Most enterprise AI pilots that fail were diagnosed as troubled by at least one person in the room weeks before anyone said so formally. The delay between private diagnosis and public escalation runs through a specific cost calculation a depleted leader makes under pressure: the interpersonal and political cost of calling the problem now, weighed against the operational cost of waiting. Both are real. The calculation happens in about ninety seconds, usually in a weekly status call, and it almost always resolves toward silence. That silence becomes the team's permission to wait too.

TL;DR

What the data says before we go deeper:

  • The failure rate is structural

    More than 80% of enterprise AI projects fail to deliver their intended business value, roughly twice the failure rate of comparable IT projects (RAND Corporation, 2024). Among generative AI pilots specifically, 95% fail to deliver measurable financial returns (MIT NANDA, 2025).

  • Abandonment accelerated sharply

    42% of companies abandoned most of their AI initiatives in 2025, up from 17% the year before, a 147% increase in a single year (S&P Global Market Intelligence, 2025). The average organization scrapped 46% of its proofs of concept before they ever reached production.

  • Leadership is where it fails, not the model

    RAND's 2024 analysis of AI practitioners found that leadership and strategy failures account for 84% of all AI project failures. The model is rarely the problem. The governance, ownership, and decision-making around it are.

  • Burnt-out leaders make risk-averse, avoidant decisions

    Research published in Organizacija (November 2025) found that the resource-depleting dimensions of burnout push leaders toward avoidant or hypervigilant decision-making. A separate body of research on mental fatigue and risk preference finds that individuals under cognitive depletion consistently make more conservative choices on complex decisions requiring long-term thinking, and show greater tendency toward decision avoidance (Nature Scientific Reports, 2022; Organizacija, 2025).

  • Teams are waiting for the leader to speak first

    When employees believe others already know about a problem, they are significantly less likely to speak up themselves. Researchers studying team meeting behavior found this pattern strong enough to name it the voice bystander effect, and noted that responsibility for raising a concern becomes something everyone assumes someone else will handle (organizational behavior research, 2019).

  • Each quarter of silence compounds the cost

    Gartner priced individual enterprise AI initiatives at $5 million to $20 million each (Gartner, 2024). Each quarter a stalled pilot remains without a decision represents both sunk cost and growing opportunity cost as competitors who are shipping extend their lead (SoftwareSeni analysis, 2026).

Where this comes from

Before I get into the research, I want to tell you where this comes from for me. In August I had three client engagements in simultaneous transition: one ramping up, one ramping down, production bugs being firefighted across both. On one Wednesday I ran three one-hour training sessions, two kickoff calls, and a production deployment in the same day. I was in pure execution mode. On that day, if a pilot had shown me a warning sign, I would have noted it and moved on. The capacity to stop and act on a diagnosis requires a different cognitive state than the one that gets you through a full calendar. I know this because I pushed through to Saturday that week and caught a cold the following Monday. That push had a physical cost and a decision cost: the things I did not stop to decide. I cleared out most of my calendar the Monday before that Wednesday specifically because I knew what was coming. I wanted space to plan and pace myself. The gym still went on Wednesday, consciously, as a trade for getting through the deliverables. But if I had not built that Monday buffer, I would not have had the cognitive runway to manage what came next. I tell you that because what follows is about a VP who is not building that buffer. And the pilot is paying for it.

What is actually happening at 2:15 on a Thursday

It is Thursday at 2:15pm. The status call has been running for eighteen minutes. The implementation partner just showed the same dashboard they showed last week, and the week before that. Adoption is flat. The error rate on output has not come down. The team lead from the partner side gave the same explanation they gave in week seven. The person holding this call recognized this pattern in week three. They watched a data-readiness gap surface in the first sprint review. By week five they had a name for what they were seeing: the pilot was running on a cleaned sample that did not represent the actual data environment. The model was performing against test conditions that bore little resemblance to production. It is now week eleven. They type 'good progress' in the chat and move to the next agenda item. They observed correctly. They did not escalate. Those two things have different causes and different remedies.

The ninety-second calculation

What happens between recognizing the problem in week three and typing 'good progress' in week eleven is a compressed cost-benefit analysis. It runs fast, mostly below the level of conscious deliberation, and it runs again in every status call in between. On one side: the cost of raising the concern now. This is not abstract. Naming a pilot as troubled in week three means owning the follow-through. It means a conversation with the C-suite sponsor before they have done enough homework. It means being the person who slows down an initiative the organization has publicly committed to. In a hostile or passive-aggressive environment, it also means being quietly blamed if the pivot goes badly. On the other side: the cost of waiting. This is easier to rationalize as smaller, because it is spread across future weeks they have not yet experienced. Maybe the partner will course-correct. Maybe the data team will resolve the access issue. Maybe week twelve looks different from week eleven. Deloitte's 2025 Workforce Intelligence Report identified mental fatigue, cognitive load, and decision friction as the leading burnout indicators, surpassing raw workload volume for the first time. When a leader is operating inside that state, the calculation tips toward deferral. Research on burnout and decision-making consistently finds that chronic depletion produces decision avoidance and decision simplification: defaulting to familiar patterns rather than optimal solutions (Organizacija, November 2025). Typing 'good progress' is a familiar pattern. It is the path of least resistance on a Thursday at 2:15 when there are still three other things open on the second monitor. The calculation in week eleven is identical to the calculation in week three. The difference is eight weeks of compounding cost.

What depletion does to the diagnosis

There is a distinction between recognizing that something is wrong and being willing to act on that recognition. Depletion erodes the second without touching the first. The VP in week three can see the warning signs clearly. What they cannot absorb, at 2:15 on a Thursday with a full second monitor, is the follow-through cost of naming them. Consider what the warning signs looked like in week three. Data access was still listed as 'in progress': approvals, masking, and data lineage unresolved, which meant the pilot was running on temporary data and temporary trust. Security review was scheduled for a later phase that had not yet appeared on the calendar. Integration to core systems was described in the original scope in a way that left the actual write-back unspecified, so the pilot was generating output that went nowhere operational. These are observable facts, present in week three's sprint review. Gartner's Q3 2024 survey of 248 data management leaders found that 63% of organizations either do not have or are unsure whether they have the right data management practices for AI (Gartner, 2025). The data problem is almost always visible early. So is the security gap. So is the integration hand-wave. On that Thursday at 2:15, in week three, the VP saw those facts and filed them. The cost of acting on them, in that moment, exceeded the cost of waiting. Research on mental fatigue and risk preference finds that individuals under cognitive depletion make more conservative choices on complex, multi-variable decisions and show a measurable tendency toward avoidance over action, with the greatest impairment on decisions requiring long-term thinking (Nature Scientific Reports, 2022; Organizacija, November 2025). An AI pilot in trouble is exactly that kind of decision: multi-variable, long-horizon, with consequences that are hard to reverse once named. In a clinical review of 298 senior executives and CEOs, 71% met criteria for significant burnout on structured assessment, 63% reported deliberately concealing the severity from their boards, and 49% said the strain had already affected their decision quality (CEREVITY, 2026). The concealment finding is worth pausing on. The burden of appearing in control compounds the original depletion. And the VP on Thursday's call is doing both: managing the depletion and maintaining the appearance that the pilot is progressing. A 2024 computational modeling study published in the American Journal of Preventive Medicine put the direct productivity cost of executive burnout at $20,683 per executive per year, nearly five times the estimate for a nonmanagerial salaried employee, because what executives do, decision-making, strategic direction, culture-setting, is precisely what degrades first under chronic stress. The VP watching the dashboard at 2:15 is not failing to care. They are operating in a cognitive state where the cost of action is felt immediately and the cost of inaction is felt in a future week they have not reached yet. In week three, that felt like a reasonable read. In week eleven, they are still making the same calculation.

The team is running the same calculation

Every other person on that Thursday call is watching the VP. The implementation team lead sees the same dashboard the VP sees. The business analyst who built the original data requirements knows exactly which gap the numbers are reflecting. The project coordinator has been updating a risk log that nobody has asked to see in four weeks. All of them have made their own version of the ninety-second calculation. All of them have resolved it toward silence. Researchers studying speaking-up behavior in team meetings have documented this pattern under the label voice bystander effect: when employees believe others might raise a concern, they are less likely to do so themselves. Even in boardrooms, responsibility for naming a problem becomes something everyone assumes someone else will handle (organizational behavior research, 2019). The room fills with people who have individually diagnosed the problem and collectively decided it is someone else's move. The VP is the person with the clearest line to the C-suite sponsor. They are also the person with the most political exposure if they call it wrong. So the team waits. And the VP, running the ninety-second calculation at 2:15 on a Thursday in week eleven, types 'good progress' and moves to the next item. The team reads that as the signal to continue waiting. The bystander effect is proportional to the power of the leader relative to the group. When the most senior person in the room signals that the pilot is on track, the remaining participants recalibrate their own assessment toward on track. This is not groupthink in the clinical sense. It is rational deference to the person with the best information and the most authority. When that person is burnt out and deferring to next week, the team defers to next week. Over half of employees worldwide report fearing the consequences of speaking up at work, whether it is challenging an idea, admitting a mistake, or asking for help (Workplace Options, 2024). The VP's silence on Thursday defers the decision and sets the floor for everyone else's willingness to speak.

What the data says about how long pilots run past their useful life

The costs of deferral are not abstract. They compound per quarter. Gartner put the cost of enterprise AI initiatives at $5 million to $20 million apiece (Gartner, 2024). An initiative running eight weeks past its diagnosis point is consuming budget, engineering attention, and organizational credibility. None of that consumption is tracked to the specific Thursday when the calculation resolved toward 'good progress.' Nearly 9 in 10 organizations delayed both agentic and generative AI deployments by an average of almost six months, driven primarily by unresolved data security and data management concerns (AvePoint State of AI, 2026). Those are the same concerns that were visible in week three's sprint review. The data access issue. The security review not on the calendar. The integration hand-wave. Each of those unresolved items is sitting inside the six-month average delay, and each one had a Thursday at 2:15 where someone saw it and typed 'good progress.' The delay at the governance layer and the delay at the leadership layer interact. When the VP who could resolve the governance question stays silent in the status call, the governance question stays open. The pilot stays in its test environment. The six months of average delay is partly structural and partly the ninety-second calculation running in every weekly status call from week three to week eleven. The earliest observable indicator of a pilot that will not ship is a team that can describe the demo but cannot describe the shipping decision. In week eleven, the team on this call can describe the demo. Nobody is talking about the shipping decision. Each week of that silence sits on a Gartner scale of $5 million to $20 million.

The sentence that interrupts the calculation

Back to Thursday at 2:15. The implementation partner is showing the same dashboard for the third week in a row. The VP has two options for what happens in the next thirty seconds. The first option is the one they have been taking. Type 'good progress,' move to the next agenda item, and carry the private diagnosis into week twelve. The second option requires one sentence. One sentence they have already thought through before the call, that names something specific enough to be actionable and narrow enough that it does not demand a full remediation plan on the spot. Here is what that sentence is not: 'I want to flag some concerns about the pilot trajectory and schedule time to discuss next steps.' That sentence opens a conversation without anchoring it. The sponsor hears that the VP is worried. They do not hear what the VP observed, what it means, or what they are being asked to decide. Here is the sentence: 'Adoption has been flat for three weeks and I want fifteen minutes with you before next Thursday's call.' That is it. Observation, specific ask. The consequence is implied by the ask itself. Fifteen minutes means she has seen something worth talking about privately before it comes up in the group. The sponsor can say yes or no. The VP has not committed to a recommended path they have not thought through. They have not blamed the partner. They have not called the pilot dead. They have opened a door that has been closed since week three. I know what it costs to push through on a Wednesday when every deliverable is due and the gym goes first. I cleared my Monday that week specifically so I would have the runway to manage what was coming. Without that Monday buffer, the Wednesday would have broken something that could not be fixed by Saturday. The VP who could name the problem in week three had the information. What they did not have was fifteen minutes of quiet before Thursday's call to write the sentence down. This is why the month's argument holds: AI adoption for a lean team is a time-reclamation question before it is a technology strategy question. The forty-five minutes consumed on Sunday reformatting a performance report draws from the same cognitive and temporal resource needed to prepare that one sentence. You cannot do both at the scale this role demands. So the smaller ask, the one that is completable in the next five minutes, before the next call: open a blank document and write one sentence that names one thing you have observed about the pilot this week. Not a plan. Not a recommendation. One observation. 'Adoption has been flat for three weeks.' 'The data access issue from week two is still open.' 'The integration step does not appear on the current project plan.' That is the preparation. The sentence gets refined from there. But the sentence that changes the trajectory of a failing pilot starts as one plain observation written down before the call begins.

What could go wrong

Being honest about the limits of this analysis:

  • Early escalation backfires without a clear ask

    Raising a concern at week three without a specific next step puts the VP in the position of problem-owner without authority to solve it. If the C-suite sponsor interprets the escalation as alarm rather than diagnosis, it can affect their credibility more than a later, more complete escalation would have. The one-sentence format is designed to reduce that risk: an observation and a request for fifteen minutes is not an alarm.

  • The pilot may genuinely resolve

    Some stalled pilots do recover. Data access issues get cleared. Partner teams course-correct. The ninety-second calculation includes this possibility, and it weights it too heavily precisely because it is the most emotionally appealing outcome. Weighing it accurately requires asking: what would have to be true by week four for this to be on track? If the answer is not visible, that is itself the observation worth naming.

  • The team's silence may not be bystander behavior

    Some team members are genuinely uncertain whether what they are seeing is a warning sign or expected variance. When information is genuinely incomplete, waiting for more data is not avoidant. The distinction is whether the VP is waiting for data or waiting because naming the problem feels costly. Those are different situations that feel identical from the inside when depletion is running.

  • The status call may not be the right channel

    In some organizations and with some C-suite sponsors, the weekly status call is exactly the wrong place to raise a concern. A bilateral conversation before the status call, or a written note sent the evening before, can carry the same information with lower political cost. The gap between diagnosis and action does not have to close in public. The fifteen-minute request is designed for exactly that: a private conversation before the group call.

  • Pilot fatigue compounds the risk of over-correction

    Deloitte identifies pilot fatigue as the organizational exhaustion that sets in when enterprises repeatedly launch AI pilots that never reach production (Deloitte, via SoftwareSeni, 2026). An organization already in that state may interpret an early escalation as one more reason to abandon rather than as an invitation to fix. The VP who escalates in week three needs to know which organizational context they are operating in before choosing their channel.

The J.Caresse point of view

The operations community has spent two years building better AI frameworks: governance checklists, adoption metrics, change management playbooks. Most of them address the structural conditions that let a pilot fail. Almost none of them address the specific moment when a leader who can see the failure decides, in ninety seconds, to stay quiet for another week. That moment, Thursday at 2:15 in week eleven, is where the structural conditions become permanent. A governance framework sits unused when the person with the authority to invoke it is running a cost calculation that keeps resolving toward silence. What actually moves a stalled implementation is the VP getting clear, before Thursday's call, on what they observe and what the smallest actionable ask is. That preparation takes five minutes and requires a cognitive state they may not have on a week where they pushed through to Saturday and caught a cold on Monday. I cleared my Monday that week in August because I knew Wednesday was coming. The VP in week eleven has not built that buffer. Sunday evening went to reformatting a performance report. The same cognitive and temporal resource that preparation requires went somewhere else. You cannot do both at scale. Something has to come off the list first, and it should be the task a bot can do before she finishes her coffee. That is where the AI team earns its place: not in the flashiest use case, not in the one the vendor demoed, but in the forty-five minutes that should have been available for the one sentence that changes the trajectory of a twelve-week implementation.

Key takeaways

What to carry out of this piece:

  • The diagnosis precedes the escalation by weeks in most failing implementations

    RAND's 2024 analysis found leadership and strategy failures account for 84% of AI project failures. The failure is rarely in what the leader observed. It is in the gap between observation and action, a gap driven by a cost calculation made under depletion, not by an absence of information.

  • The ninety-second calculation is predictable and can be interrupted

    The calculation that keeps a leader silent in week three has a known structure: it overweights the immediate interpersonal cost of speaking up and underweights the compounding operational cost of silence. It runs again in week four, and week five, with the same inputs and the same resolution, until week eleven.

  • The team's silence is a response to the leader's signal

    When the VP types 'good progress,' the team reads that as permission to wait. When they name a specific observation and ask a specific question, the team has permission to do the same. That sentence sets the floor for what the room is allowed to say.

  • Each week of silence is priced at Gartner's $5 to $20 million scale

    Gartner's pricing of enterprise AI initiatives at $5 million to $20 million each (Gartner, 2024) converts the escalation delay from an interpersonal discomfort into a capital allocation decision. Each Thursday that resolves toward 'good progress' is a capital allocation decision, even when it feels like waiting for more data.

  • The sentence is the intervention

    One observation, written down before the call begins. Not a plan, not a recommendation, not a full diagnosis. One thing they have seen. 'Adoption has been flat for three weeks.' That sentence is the intervention. Everything else follows from it.

  • Pilot purgatory has a decision at its center

    Nearly 9 in 10 organizations delayed AI deployments by an average of almost six months (AvePoint State of AI, 2026). Inside each of those explanations is a person who had a decision to make and deferred it. The VP who moves first, with a specific observation and a small ask, is the person who converts a stalled pilot into a live question the organization can actually answer.

Private Consultation

Bring these ideas into the room.

If this essay sounds like the conversation you're sitting with, Jessica responds personally to every inquiry.