Kill Criteria and Exit Mechanics: An Option You Cannot Bring Yourself to Abandon Is Not an Option at All

Real-options reasoning sounds clean in a corporate slide deck, promising a cheap trial where you evaluate evidence before deciding whether to commit real capital. But an experiment only works if its main job is to buy reliable facts, and an option is only real if you actually have the stomach to abandon it.

Without plain, pre-committed kill criteria, a pilot project is just an implementation waiting for permission. Once an AI initiative acquires executive sponsors, vendor contracts, middleware integrations, internal communication campaigns, and political capital, stopping becomes hard. Because AI outputs are probabilistic and qualitative, project champions can always argue that one more fine-tuning run, an updated prompt library, or the vendor’s next model release will fix the quality problem. Sunk-cost bias takes hold, and the initial experiment quietly morphs into a permanent fixture.

This is how enterprise technology accumulates zombie systems: failed or marginal initiatives that linger indefinitely. You see them everywhere in enterprise tech, including custom AI search wrappers that staff quietly bypass in favor of raw web tools, while the organization continues paying thousands of dollars a month in vendor seat licenses, compute fees, and maintenance bandwidth.

If your AI strategy relies on real-options reasoning to manage uncertainty, and it should, you must confront an uncomfortable operational reality: the ability to stop is not an automatic feature of software experiments. You must design it as an intentional management capability before the work begins.

The Abandonment Value of an Option

In financial option theory, the value of an option depends directly on the right to walk away. If you are forced to exercise a contract regardless of market conditions, you do not own an option; you hold an obligation.

The same rule governs AI commitments:

An option you cannot bring yourself to abandon is not an option at all.

When an AI project begins, stopping is cheap. The team is small, the financial commitment is bounded, and leadership remains relatively disinterested in the outcome. Six months later, the dynamic changes. The vendor has added features, the project team has staked its reputation on delivery, and middle management has reframed the experiment as a strategic milestone. It begins to look like you are in the middle of an enterprise software implementation.

At that point, evaluating performance objectively becomes almost impossible. Every missing capability is treated as a temporary bug to be fixed in the next release. Every cost overrun is reframed as a necessary investment in baseline infrastructure.

Simply put, the best time to decide what would make you stop is before you care whether you stop.

Concrete Termination Signals

To prevent experiments from degenerating into zombie projects, leadership must define clear, operational stopping triggers upfront that mandate hard decisions. Generic milestones like “improve productivity” or “demonstrate business value” are useless because they can be endlessly redefined to justify continued funding.

Instead, exit mechanics should respond to specific operational sinals:

  • Validation costs rising faster than execution gains: If an automated workflow generates draft output in seconds while verifying that output requires extensive human review or complex secondary verification systems, the net economic gain evaporates. When validation overhead permanently outpaces execution savings, the option has lost its value.
  • Performance plateauing below reliance thresholds: Machine learning systems often achieve 80% accuracy quickly, while the remaining 20% requires exponentially more data, tuning, and ongoing supervision. If model performance hits a plateau that still requires constant human intervention for high-consequence tasks, the initiative should be paused or abandoned rather than continuously funded on the assumption that accuracy will magically improve. In many cases, even 99% accuracy simply is not good enough.
  • Abrupt changes in infrastructure economics: Vendor pricing models, API fees, and compute requirements can shift rapidly. An architecture that makes economic sense at a trial tier may become unviable when scaled across an enterprise. If the marginal cost of execution outpaces the marginal value delivered, the experiment should trigger an immediate review. That means you must be measuring these metrics.
  • Scaling requiring unviable human supervision: If scaling an AI workflow from ten users to ten thousand requires a proportional increase in human oversight to manage edge cases and error rates, the initiative fails the basic test of operational leverage.

Preserving Reconstitution and Artifacts

Exit mechanics are not simply about shutting off API access or terminating a vendor subscription. A sudden shutdown that leaves behind orphaned data, broken dependencies, or unmapped workflows creates its own form of operational debt.

A structured exit must cover three critical elements:

  1. Data and Prompt Portability: Extracting accumulated context, proprietary dataset refinements, evaluation benchmarks, and system prompts so they remain organizational property rather than vendor lock-in.
  2. Preservation of Institutional Learning: Documenting explicitly why the initiative failed, where the model broke down, and what failure modes were uncovered. Without a clear record of termination, the same experiment will inevitably be re-proposed two years later by a different team.
  3. Fallback Reconstitution: Ensuring that the underlying human skills and legacy workflows displaced by the trial are preserved until the option is officially exercised. The true cost of an unmanaged exit is discovering that you have dismantled the cognitive infrastructure needed to run the manual baseline.

Shutting down an experiment should not feel like an admission of failure or a chaotic emergency cleanup. It is the scientific method doing its job for you. It should be treated as the orderly execution of an intentional strategy to recover capacity and redeploy resources toward higher-value opportunities.

The Pre-Commitment Kill Criteria Protocol

Before any AI initiative receives initial funding or advances to the next level of commitment, leadership should pre-commit to four explicit bounds:

  • Expansion Evidence: What specific, measurable evidence of performance, validation efficiency, and user adoption would justify expanding this initiative?
  • Continuation Evidence: What evidence would justify running another bounded experiment to resolve remaining uncertainties?
  • Termination Evidence: What specific operational, financial, or accuracy signals will trigger immediate termination without further debate?
  • Independent Authority: Who holds the independent authority to exercise the stop decision?

If the person holding authority to stop the project is also responsible for promoting it, the option to abandon will almost never be exercised. Authority should rest with an independent portfolio governance lead or risk committee with no political stake in the project’s continuation.

When these bounds are established before funding is released, exercising the option to abandon changes from a politically charged conflict into a routine operational decision.

The Diagnostic Test

As you review your organization’s active AI initiatives, ask your project leads and innovation committees one simple question:

Which of our current AI experiments have explicit, pre-committed conditions under which we will shut them down this quarter, and who specifically has the authority to make that call?

If the answer is “none,” you do not have a portfolio of real options. You have a collection of implementations waiting for permission.

Prudent AI is not a prescription for moving slowly, nor is it about guaranteeing that every experiment succeeds. It governs consequential AI commitments under uncertainty while preserving the capacity to learn, adapt, and recover. Move fastest where experiments are cheap, failures are reversible, and learning is valuable. As AI systems become more deeply embedded and harder to unwind, demand stronger evidence, pre-commit to explicit kill criteria, and preserve the organizational discipline to stop.

Accelerate reversible learning. Pace irreversible commitment.

Dennis Kennedy – CC BY 4.0 license

[Originally posted on DennisKennedy.Blog (https://www.denniskennedy.com/blog/)]

DennisKennedy.com is the home of the Kennedy Idea Propulsion Laboratory

DennisKennedy.Blog is part of the LexBlog network.