Options, Not Pilots: The Mechanics of AI Capital Allocation
The AI proposals arrive at your executive committee one at a time. Knowledge Management wants to test an agentic research workflow. The corporate group wants automated first-pass diligence. Business Development has found a tool that promises to identify cross-selling opportunities from the firm’s relationship data. IT wants an enterprise platform rather than another collection of point solutions. The litigation group has three projects of its own. A few partners are already running experiments that nobody formally approved.
Each proposal sounds reasonable on its own. Each has an enthusiastic sponsor, a manageable initial budget, and some version of the same reassuring label: pilot. No single proposal requires you to make an existential bet on artificial intelligence. Approve enough pilots, however, and you have made a portfolio decision without ever having a portfolio discussion.
The problem with these pilots is not necessarily that the technology fails. Many will succeed. The problem is that they can succeed while answering the wrong question.
They ask: Can we make this technology work?
They should ask: What did we learn that changes what we should commit next?
In short, we need to stop thinking in terms of pilots and start buying options.
What Are You Actually Buying?
When you allocate $100,000 to an AI initiative, what are you actually purchasing? The conventional answer is capability, efficiency, or automation. Is it really? An early-stage AI project is usually an investment under considerable uncertainty. What you should be paying for is decision-relevant information.
Think of it as buying an option. You spend a relatively small amount today to learn enough to make a more consequential decision tomorrow. This is not a literal financial option, and nobody is going to calculate a Black-Scholes value for an agentic research workflow. The useful principle is simpler: spend small amounts of capital to resolve important uncertainties before spending large amounts of capital and institutional capacity on commitments that will be expensive to unwind.
That changes what counts as a successful experiment. It should not merely tell you whether a tool can perform a task. It should produce evidence capable of changing a consequential decision. The sequence is experiment → evidence → decision → commitment. If the evidence cannot plausibly change the decision, you are not buying much option value. You are gathering support for a commitment you have effectively already made.
Learning Before Commitment
That yields a governing principle for prudent AI capital allocation:
Learning Velocity > Commitment Velocity
Commitment is not measured only in dollars. A month-to-month software subscription might be easy to abandon. Connect the same tool to confidential firm data and the exposure changes. Rebuild a practice-group workflow around it and it changes again. Allow the old workflow to decay and the cost of retreat rises further. What began as a modest technology experiment can become an institutional dependency.
When commitment gets ahead of learning, assumptions harden into architecture before you have adequately tested them. On the other hand, when learning leads, evidence earns the next level of commitment. The objective is not maximum reversibility or minimum commitment. Some things can only be learned by making a real commitment, and delay has costs of its own. The objective is to make no more commitment than you need to resolve the next consequential uncertainty.
Cheap Experiments Should Make Us More Demanding
AI is making many experiments cheaper and easier to build. But cheap to build does not mean cheap to evaluate. KM expertise, IT integration time, information-security review, lawyer attention, clean data, workflow knowledge, training capacity, and executive attention remain scarce. Ten $25,000 pilots are not necessarily a cheap $250,000 portfolio if all ten depend on the same three people in KM and IT and compete for the attention of the same lawyers who must determine whether the results have real value.
So cheaper experimentation should not produce an indiscriminate explosion of pilots. It should let us put more of our important assumptions to the test before making expensive commitments.
This is where those early science classes become useful. An experiment designed only to prove something works is less useful than one designed to discover what might make the next investment a mistake. If an inexpensive experiment tells you that a proposed AI system works beautifully under one set of conditions but becomes unreliable under another, the experiment has bought information that can change what you do next.
Cheap experimentation should make ignorance harder to defend.
When Several Experiments Become One Bet
There is one more problem that does not appear on any individual pilot proposal. Five different groups can bring forward five genuinely different AI initiatives and still be making substantially the same underlying bet.
The KM project, corporate project, BD project, and litigation projects might all depend on the same foundation model, enterprise platform, retrieval architecture, or assumption that lawyers will reliably catch a certain category of error. Viewed project by project, you appear to be experimenting broadly. Viewed as a portfolio, you may be concentrating rapidly around a small number of untested assumptions.
So before approving the next commitment, look not only at the experiment in front of you but also across the table. What assumptions do these supposedly different experiments share? What scarce institutional capacity are they competing for? Where are you unintentionally climbing several commitment ladders at once?
The Option-Value Test
Before approving the next stage of an AI initiative, ask three questions.
1. The Uncertainty Question: What specific, consequential uncertainty are we paying to reduce?
2. The Cost-to-Learn Question: What is the smallest, cheapest, most reversible experiment that will meaningfully reduce that uncertainty?
3. The Pivot Question: What specific decision will change depending on what the evidence reveals?
The third question is the one most likely to expose a disguised implementation. If no plausible result will cause you to stop, redirect, narrow, expand, accelerate, or delay the commitment, the experiment is not functioning as an option. You are rehearsing a decision already made.
The purpose of an AI experiment is to acquire evidence that improves your ability to make the next consequential allocation decision about a specific AI use for the actual job to be done. Some experiments should move rapidly. Some should earn substantial institutional commitment. Others should end, not because they failed, but because they succeeded in telling you not to spend the next dollar there.
Prudent AI treats early AI investments as opportunities to buy information and preserve choices, not as down payments on predetermined implementations. Let evidence earn greater commitment while preserving the capacity to choose differently when it doesn’t.
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
Like this post? 
Buy me a coffee
DennisKennedy.Blog is part of the LexBlog network.