Chapter 14 of 15
The Kill Decision
How to evaluate a struggling AI initiative honestly — separating sunk cost from future value, analyzing three scenarios, recommending a course of action, and communicating it in a way that preserves organizational trust and builds learning.
Overview
Somewhere between twelve and eighteen months into most AI programs, at least one initiative will reach a point where the original business case is not holding up. The metrics are not where they should be. The team has been making adjustments but the trajectory has not changed. The quarterly review feels more defensive than confident. And nobody has said the word "kill" yet, because killing an initiative feels like admitting failure.
This is one of the most consequential moments in an AI program. Organizations that make the kill decision well, when they should, for the right reasons, with clear communication, emerge with stronger programs, higher organizational trust in AI investment, and a culture that can support continued innovation. Organizations that do not make it, that keep funding struggling initiatives past the point of rational justification or communicate the decision poorly, pay compounding costs for years.
This capstone gives you the framework for making and communicating the kill decision well.
Why This Decision Is Hard
The kill decision is emotionally and organizationally difficult for reasons that are entirely predictable and entirely resistible, if you understand them.
Sunk cost fallacy. The human brain is wired to treat money already spent as a reason to continue spending. "We've put $3 million into this — we can't just walk away." The correct response is: the $3 million is gone whether you continue or not. The only relevant question is whether future investment will generate sufficient future value. The $3 million is not a reason to continue. It is irrelevant to the forward-looking decision.
Think of it like this: A commercial real estate investor buys a property for $5 million that declines in value to $3 million. Their advisor recommends selling before values decline further. The investor says: "I can't sell — I'd be taking a $2 million loss." But the loss already happened. The question is not whether to realize it on paper but whether holding longer will increase or decrease the damage. The property's future value is the only relevant variable. The original purchase price is not. AI initiative decisions work identically.
Political attachment. Initiatives have sponsors. Sponsors have reputations. Killing an initiative feels like indicting the sponsor's judgment. In organizations where failure is not handled gracefully, sponsors resist killing their initiative because they fear the personal consequence more than the continued cost to the organization.
Narrative momentum. Vendors have told your stakeholders this is going to work. The team has told the organization this is important. The initiative appears in board materials. Killing it requires revising a narrative that has institutional momentum.
Fear of the message. "We are terminating this AI initiative" can be misread by the organization as "AI does not work here" — which is particularly damaging if you have multiple other AI initiatives that are performing. The communication of the kill decision is as important as the decision itself.
Understanding these obstacles does not eliminate them. It gives you the ability to name them in the decision conversation and set them aside.
Step 1: The Honest Assessment
Before you reach a decision, you need a clear, honest picture of the current state. This is harder than it sounds because the people closest to the initiative are the least likely to have an unmediated view of it.
The Initiative Status Report
Require the initiative lead to complete this document before the evaluation begins. The framing matters: this is not a defense brief — it is an honest status assessment. If the initiative lead cannot write an honest assessment, that is itself useful information.
Current State Summary:
| Metric | Baseline | Target | Current | Trend (last 3 months) |
|---|---|---|---|---|
| [Primary value metric] | ||||
| [Secondary value metric] | ||||
| [Cost metric] | ||||
| [User adoption] |
Gap Analysis:
| Gap from Target | Root Cause (Honest Assessment) | Evidence |
|---|---|---|
| [Metric 1 gap] | ||
| [Metric 2 gap] |
What Has Been Tried:
List the adjustments made in the past 90 days to address underperformance. Describe the result of each adjustment.
| Adjustment | When | Result |
|---|---|---|
The Initiative Lead's Honest Assessment: In one paragraph, what does the initiative lead actually believe is likely to happen in the next 6 months if the current approach continues?
The Independent View
Do not rely only on the initiative team's assessment. Assign someone outside the initiative — your measurement owner, a Finance Business Partner, or an external advisor — to produce an independent view of the same metrics. Discrepancies between the initiative team's view and the independent view are important data.
Step 2: The Three Scenarios
For any struggling AI initiative, there are three possible forward paths. Each has a different investment and value profile. Analyze all three before making a recommendation.
Scenario A: Continue as Planned
The thesis: The initiative is underperforming but the root causes are identifiable, addressable, and temporary. The fundamental business case remains sound. With the identified adjustments, the initiative can still reach its targets.
When to take this seriously:
- The underperformance can be specifically attributed to a root cause (not "the AI needs more time")
- That root cause has a concrete fix with a specific timeline
- The fix has not already been tried and failed
- The adjusted business case still meets the original investment hurdle rate
What a credible Scenario A looks like:
| Element | What It Must Contain |
|---|---|
| Root cause | Specific, not generic ("data quality in the Eastern region is insufficient" not "we need better data") |
| Fix | Specific named action, named owner, specific deadline |
| Timeline to impact | Specific date by which the fix will have produced measurable effect |
| Adjusted forecast | New metric projections under the fix assumption |
| Decision trigger | If metrics have not reached X by Y date, the scenario changes |
A Scenario A without these elements is a wish, not a plan. Do not approve Scenario A without them.
Scenario B: Pivot
The thesis: The initiative in its current form is not working, but there is a different, better-defined version of the initiative that could work. The learning from the failure of the current approach is valuable and should be preserved; the direction should change.
When to take this seriously:
- The core technology is working but the use case was wrong
- The use case is right but a different implementation approach would work better
- The value exists in a different function or for a different user group than originally targeted
- The investment to pivot is meaningfully less than the value of starting fresh with the learning you now have
What a credible Scenario B looks like:
| Element | What It Must Contain |
|---|---|
| Current failure | Clear statement of what specifically is not working |
| The pivot | Specific description of what changes — use case, user, technology, or implementation approach |
| Evidence the pivot will work | Not hope — reference cases, pilot data, user research |
| Incremental investment | What the pivot costs beyond what is already committed |
| New business case | Value projections for the pivoted initiative with honest confidence intervals |
| Sunset plan for current approach | Specific actions to wind down what is not working |
Pivot decisions fail most often because they are not specific enough. "We are pivoting to focus more on operations" is not a pivot — it is a rebranding. A specific pivot has a specific new use case, a specific new business case, and a specific new timeline.
Scenario C: Kill
The thesis: The initiative is not salvageable in its current form, and a credible pivot does not exist. The remaining investment required to reach value is not justified by the remaining probability of achieving it. The organization is better served by stopping, learning, and redeploying those resources.
When to take this seriously:
- Multiple adjustment attempts have not changed the trajectory
- The root cause of underperformance is structural (the business case assumptions were wrong, not the execution)
- The remaining investment needed to reach value exceeds what is justified by the probability-adjusted value
- The team working on the initiative has lost confidence in the path and is not being realistic about the probability of success
The kill is not a failure acknowledgment. It is a capital allocation decision. The question is not "did this fail?" The question is: is future investment in this direction the best use of available resources? Often the answer is no. That is a financially disciplined answer, not a failure.
Step 3: The Scenario Comparison
Score each scenario on four dimensions:
| Dimension | Scenario A: Continue | Scenario B: Pivot | Scenario C: Kill |
|---|---|---|---|
| Probability of reaching value (0-100%) | |||
| Additional investment required ($) | |||
| Time to measurable value (months) | |||
| Opportunity cost (what else could we do with these resources?) |
The scenario comparison is a framework for conversation, not a mathematical output. Numbers assigned to "probability of reaching value" are educated estimates, not actuarial calculations. The value of the exercise is forcing the conversation to be specific about each scenario's real prospects, rather than remaining at the level of optimism or pessimism.
The Expected Value Calculation
For each scenario, compute an expected value:
Expected Value = (Probability of success × Value if successful) − Additional investment required
If Scenario A has a 30% probability of reaching a $5M value target with $2M of additional investment: EV = (0.30 × $5M) − $2M = $1.5M − $2M = −$0.5M
If Scenario C (kill) returns $0.5M in redeployed resources and team capacity, its EV is +$0.5M.
In this example, Scenario C is financially superior to Scenario A — even though Scenario A has a non-trivial probability of success. The expected value calculation makes explicit what intuition sometimes resists acknowledging.
Think of it like this: A poker player with a losing hand who has already put $200 in the pot faces a $300 raise. The $200 is irrelevant — it is gone. The question is whether the cards they hold have sufficient expected value to justify the $300 call given the remaining players and likely hands. An experienced player folds and keeps the $300. A less experienced player calls because they are "already in" — and loses the $300 too. AI initiative decisions are structurally identical to this calculation.
Step 4: Making the Recommendation
After the scenario analysis, one scenario will be clearly superior or there will be a close call between two. Either way, produce a written recommendation. This is not optional. A verbal recommendation in a meeting is too easily misunderstood, too easily revisited, and too easily blamed on whoever was in the room.
The Recommendation Document
One page. Five sections.
1. Current State (3-4 sentences) The initiative's current position: where it is against targets, what the trend is, and the honest characterization of why it is where it is.
2. Scenarios Considered (1-2 sentences each) Brief description of each scenario evaluated and the key variable for each.
3. Analysis (4-6 sentences) Why one scenario is superior to the others. Include the expected value calculation. Be specific about the assumptions that drive the analysis.
4. Recommendation State it plainly: "We recommend [Scenario A / B / C]." No hedging. No "we might consider" or "subject to further analysis." A recommendation is not a suggestion.
If you are recommending Scenario A (continue) or Scenario B (pivot), include specific conditions: "If the following milestones are not met by [date], we will return with a Scenario C recommendation at that time." This prevents the continue or pivot decision from becoming a license for indefinite additional investment.
5. What Happens Next (bullet list) The specific actions in the first 30 days following the decision, with named owners and dates.
Step 5: The Communication Plan
Killing an AI initiative creates an organizational narrative. If you do not shape that narrative, rumor will. The people most affected will fill the vacuum, and the organization will interpret events as signals about the future.
The communication you control is always better than the narrative vacuum you leave.
Audiences and Messages
The Board and Executive Team
Timing: Immediately, in the same communication as the decision.
Message: "We are terminating Initiative X. Here is the specific reason. Here is what we learned. Here is how we are redeploying the resources. This decision strengthens our overall AI program by [specific reason]. The other four initiatives are performing as follows [brief status update]."
Emphasize: the decision is a demonstration of disciplined portfolio management, not evidence that AI does not work. Boards that have been through previous technology cycles will recognize and respect this framing.
The Initiative Team
Timing: Before any broader communication. Private, direct, respectful.
Message: "We are making a change to this initiative. I want to talk to you about it directly before you hear it from anyone else."
Acknowledge the work done. Be specific about what was learned. Be clear about what is happening next for each person on the team. The handling of the people closest to the initiative will be the most visible signal about organizational values in this moment.
The Broader Organization
Timing: Shortly after the initiative team, same day if possible.
Message: Calibrate to the visibility of the initiative. If it was widely known, a brief announcement. If it was largely internal, manager-to-team communication may be sufficient.
Frame: "We launched this initiative to test whether [specific value]. We learned that [specific finding]. Based on that learning, we are [specific action]. The resources are being redeployed to [specific use]. We remain committed to AI investment and continue to run [list other initiatives]."
Do not be defensive. Do not blame the vendor, the market, or "timing." Do not be vague. Vagueness is filled by anxiety.
Vendors and Partners
Follow contract terms for notice. Where possible, have the conversation before sending the formal notice — relationship handling in how you wind down will affect your standing in the market and with vendors for future initiatives.
The Learning Debrief
Within 30 days of the termination decision, conduct a formal learning debrief. This is not a post-mortem, which often becomes a blame exercise. It is a structured examination of what the initiative revealed.
Debrief Questions
About the business case:
- What assumptions were in the original business case that turned out to be wrong?
- Were those assumptions testable earlier? Why were they not tested earlier?
- What would a better-constructed business case have looked like?
About execution:
- What execution choices made the outcome worse than it could have been?
- What would we do differently in the implementation approach?
- What signals did we have early that we did not act on?
About governance:
- At what point did we have enough information to make the termination decision, and why did we not make it then?
- What governance structure would have surfaced the problem earlier?
About the learning:
- What do we now know about this problem space that we did not know at the start?
- Does this learning have value for other initiatives or future investments?
- Is there a better-defined version of this problem that is worth pursuing in a future cycle?
Write a two-to-three page learning summary and distribute it to all initiative sponsors and the AI program lead. This document is one of the most valuable outputs of a killed initiative. The investment is not entirely lost if the organization genuinely learns from it.
The Organizational Signal the Kill Sends
How you handle the kill decision sends a signal to your organization about what kind of AI program you are building.
If you make kill decisions when the evidence warrants them, communicate them honestly, learn from them systematically, and redeploy resources to better uses — you are signaling that your AI program is rigorous, that investment is subject to evidence, and that failure of an experiment is not a career risk for the people who ran it.
That signal makes every future initiative easier to staff, easier to fund, and more likely to be honest with you when things are not working.
If you do not make kill decisions when you should, if you keep funding failing initiatives because of sunk cost, political attachment, or fear of the message, you signal that AI investment is not subject to normal financial discipline. That signal attracts more of the same: larger bets, lower accountability, and a culture where nobody tells the executive what is really happening.
The kill decision, done well, is not the end of an AI initiative. It is evidence that your program is mature enough to be trusted with more.
Key Takeaways
- Sunk cost is not a reason to continue. The only relevant variable is whether future investment justifies the future probability of future value. Everything already spent is irrelevant to the forward decision.
- Analyze all three scenarios — Continue, Pivot, Kill — before making a recommendation. Skipping the analysis produces a decision that will not survive scrutiny.
- The expected value calculation makes the decision concrete: (probability of success × value if successful) minus additional investment required. Run this for each scenario.
- Write the recommendation as a one-page document with five sections: current state, scenarios, analysis, recommendation, and next actions. Verbal recommendations in meetings are insufficient.
- The communication plan is as important as the decision: the board and executive team first, the initiative team next (privately, before anyone else), and the broader organization shortly after.
- Conduct a structured learning debrief within 30 days. The learning from a killed initiative is often worth more than its cost if the organization captures and uses it.
- How you handle the kill decision signals whether your AI program is rigorous. Organizations that kill when they should build cultures of honesty and accountability. Organizations that do not, build cultures of wishful thinking.