AI Value Hypothesis: Use AI to Start the Idea… Then Try to Kill It

An AI value hypothesis does two things that a normal use case rarely does: it gives an idea permission to begin before certainty exists, and it defines what evidence should make us change direction before enthusiasm becomes commitment. I have been learning that distinction by building Open Source Consulting in public. What began largely as an experiment in career visibility and a place to think has gradually become something more structured, not because I had the destination mapped, but because each iteration created enough evidence to justify the next one.

For most of my career in consulting, I wanted a platform where I could put unfinished ideas into the open, test them with peers and experts, and see which ones survived contact with other people. I never found a version simple enough to sustain. Publishing looked like a discipline in its own right: editorial calendars, SEO, analytics, distribution, audience building, design, newsletters, all the machinery that people who work in digital marketing understand instinctively and that, for me, often looked like a black box.

So I started with the smallest possible proposition: make the platform useful as a thinking environment first. If it helped me articulate ideas, expose assumptions and create professional visibility, it was already doing a job. Anything beyond that would have to earn its way into the system.

That sounds obvious in hindsight. It was not obvious at the beginning. And that is precisely why I think the distinction between a use case and a value hypothesis matters far beyond content.

An AI value hypothesis starts before certainty

A conventional business instinct says that an idea should be justified before resources are committed. That is sensible when the commitment is large. It is much less useful when the cost of learning has collapsed.

AI has changed that cost structure. It can help research a domain, challenge a premise, draft alternatives, connect adjacent ideas, simulate objections and turn a vague thought into something concrete enough to test. This creates a useful inversion: we can often afford to start learning before we are confident the idea is valuable.

That does not mean we should fund everything. It means the threshold for a small experiment can be much lower than the threshold for a transformation. The later Earn the Right to Transform framework develops this into an explicit commitment ladder: evidence should earn the move from hypothesis to experiment, pilot and transformation.

The first hypothesis behind Open Source Consulting was therefore modest: if I publish useful thinking consistently, I may learn which ideas deserve deeper development while improving my professional visibility. It did not require a business model, a service catalogue or a five-year roadmap. It required a publishing rhythm and a willingness to learn from what happened.

The weekly rhythm mattered more than any single article. I treated the work in sprints, with new and continuing hypotheses entering the week and retrospectives forcing the system to ask what had changed. That discipline created continuity without pretending that the destination was known.

This is close to the logic of the Lean Startup: ideas are treated as hypotheses, experiments create validated learning, and evidence informs whether to persevere or change direction. But AI adds something important to that loop. It lowers not only the cost of building the experiment; it also lowers the cognitive cost of questioning it.

Use AI to start an idea before you have convinced yourself it is valuable. Then use AI just as aggressively to test whether it deserves to continue.

That second sentence is the part I increasingly care about.

The danger is not experimentation. It is falling in love with the experiment

AI makes plausible ideas unusually easy to produce. A team can generate customer personas, market maps, operating-model options, financial assumptions, prototype flows and executive narratives faster than many organizations can schedule the steering committee that will review them.

This is useful cognitive leverage. It is also a new source of self-deception.

Coherence is not validation. A polished explanation can hide a weak causal mechanism. A convincing prototype can hide an irrelevant problem. A dashboard can create the appearance of measurement without proving that the metrics correspond to value. And when people have invested time, identity or political capital into an idea, the ability to produce more supporting material can make it easier to defend the idea rather than challenge it.

The practical answer is not to become more cautious about exploration. It is to make falsification part of the same system that generates possibilities.

In my own work, that means asking AI not only to develop an argument but to find contradictions, compare it against external evidence, identify overlap with ideas I have already published, challenge optimistic interpretations of metrics, and distinguish a positive signal from actual validation. I have written before about how AI agent decision quality improves when repeated human corrections become part of the system. The same principle applies to ideas: every correction should improve the mechanism that decides what deserves another iteration.

This is not an argument for delegating judgment to AI. Human judgment remains the validation gate. The point is to use machines to make that judgment better informed, more frequent and less dependent on memory or enthusiasm.

The first diamond was divergence

There is another framework that helps explain what happened more accurately than a neat retrospective story would.

The UK Design Council’s Double Diamond separates design and innovation into two broad movements. The first diamond expands and then converges around the problem: Discover and Define. The second expands and converges around possible solutions: Develop and Deliver. The model is deliberately simple, but its central distinction between divergent and convergent thinking is useful here.

For a long time, Open Source Consulting was mostly in divergence.

Publishing created material. Material created patterns. Some of those patterns came from things I already knew but had never connected explicitly. Others came from AI finding relationships that I would probably have found only inside a much larger collaborative environment, with researchers, marketers, analysts, designers and domain experts bringing different fragments of context into the room.

That changed the economics of sensemaking. I could follow more threads than my individual cognitive bandwidth would normally allow. The result was not simply more content. Frameworks began to emerge from recurring patterns across the content.

This is where a thinking platform became more interesting than a blog.

The distinction matters. A blog can be a distribution mechanism. A thinking platform can be part of a learning system. The article is not the unit of value; the accumulating evidence, patterns, contradictions and reusable concepts are.

I saw one unexpected example when the work started becoming relevant to an upcoming lecture at the University of Applied Sciences where I completed my bachelor’s degree. That outcome had never been part of the original hypothesis. It was an organic branch of the experiment.

A rigid roadmap might have classified that as scope drift. A hypothesis-led system can treat it as new evidence and ask whether it deserves further exploration.

Then my own cognitive limit became the bottleneck

Divergence feels productive until the number of plausible directions becomes larger than the number you can responsibly pursue.

By the time the publishing system had accumulated roughly fifty articles, the constraint had changed. There was enough material to see what was distinctive, enough repetition to identify durable themes, and enough external response to begin forming stronger hypotheses about who might find the work useful. But my own capacity to coordinate the research, editorial work, website, analytics, distribution, career priorities and next experiments was becoming the bottleneck.

That is when the AI agents stopped being a collection of tools and became an operating system.

I organized them around explicit roles and shared Objectives and Key Results. Content strategy, editorial production, growth, analytics, website architecture and other responsibilities received clearer ownership. I wrote about the deliberately ridiculous names in AI Agent Roles: Why I Gave Mine Ridiculous Names, but the useful part was never the names. It was the decision to make boundaries and responsibilities explicit.

The OKRs changed the behavior of the system. Instead of treating every possible direction as equally interesting, the agents had a shared reference point for what mattered now. That moved the work toward the second funnel of the Double Diamond: less indiscriminate exploration, more deliberate convergence around propositions that had accumulated evidence.

The Open Source Executive Briefing emerged in that transition. It was not simply another content format. It represented a stronger hypothesis: perhaps the value was not only in publishing individual essays, but in synthesizing evidence into a recurring executive product for a more specific audience.

That hypothesis still has to earn its continuation. A name, a Friday cadence and a newsletter template do not make it a validated product.

A use case is not a value hypothesis

This is where the personal experiment becomes transferable to enterprise AI.

A use case usually tells us what could be done. Consider a familiar example: use generative AI to accelerate proposal creation. That statement identifies a task and a technology. It may even include estimated time savings.

But the value logic is still missing.

A stronger hypothesis would be closer to this: if AI reduces proposal-preparation effort while preserving decision quality and client relevance, senior consultants can redirect meaningful capacity toward qualified client work; if that capacity is actually reallocated, the change should improve a defined commercial outcome enough to justify the workflow, governance and integration cost.

Now we have something that can lose.

Maybe proposal creation is not the constraint. Maybe people save time and simply absorb more internal work. Maybe the organization produces more low-quality bids. Maybe quality deteriorates. Maybe the governance overhead removes the economic advantage. Maybe the benefit exists only for junior drafting while the scarce senior review time remains unchanged.

A use case describes a possibility. An AI value hypothesis describes the causal chain that would make the possibility matter.

A business case should come later, when enough of that chain has evidence behind it to justify a larger commitment.

BCG’s July 2026 research captures the enterprise version of this problem. Its survey found that nearly nine in ten CEOs reported some AI-related cost or revenue benefits in targeted areas, but the report also describes a large execution gap when organizations try to scale. Only 14% clearly defined P&L impact for all AI initiatives, and BCG recommends defining the value path, baseline, KPIs, owners and value logic before an initiative goes live, then continuously revisiting assumptions as evidence emerges. That is a living value hypothesis, whether or not we use the term.

Build–Measure–Learn and the Double Diamond solve different problems

I would not collapse these two models into one.

The Double Diamond helps decide where to expand and where to converge. It protects against solving a poorly understood problem too quickly and then protects against committing to the first plausible solution.

Build–Measure–Learn helps decide whether a proposition has earned another iteration. It creates a feedback loop between action, evidence and the next decision.

AI can accelerate both.

During divergence, it increases the range of sources, patterns, counterarguments and solution paths a small team can examine. During convergence, it can compare evidence against objectives, identify contradictions, test alternative interpretations and help construct smaller experiments around the riskiest assumptions.

That is how I now understand the role of agents such as Don, my growth manager, and My Quant, the KPI analyst. They are not permanently positioned on a linear production chain. They become especially important when the chain needs to be interrupted.

If external trends shift, if distribution behaves differently, if KPIs challenge our expectations, or if a success looks suspiciously convenient, their job is to help break the internal narrative and ask whether the direction still deserves confidence.

The value chain temporarily comes apart so that it can come back together in a more targeted form.

That is a feature, not a defect.

The emerging AI-native value chain’s ongoing validation is what gives Open Source Consulting a reason to exist

At the moment, I can see a basic chain forming inside Open Source Consulting:

There are supporting loops around it: distribution generates response; analytics challenge interpretation; growth research tests whether attention is changing; retrospectives convert mistakes into operating rules; new evidence can push the system back into exploration.

This begins to resemble an AI-native professional-services value-generation chain because AI is not merely speeding up one step. It is changing the economics of coordination across the chain.

Research that would once have required more people can be explored continuously. Synthesis can happen across a much larger body of material. Repeated patterns can be converted into frameworks. Different agents can maintain continuity across editorial, analytical and operational responsibilities. Human attention can move upward toward validation, trade-offs and consequential decisions.

But I want to be precise about the evidence.

These are observations about a system becoming more capable. They are not proof that Open Source Consulting’s scalable professional-services model will stay as it is.

The strongest signal so far is more personal and more modest. Online visibility was something I had struggled to make systematic throughout my career. I depended heavily on friends and colleagues who understood digital marketing and analytics better than I did. When I first began publishing, some of them reached out without being asked, encouraged the experiment and generously shared expertise that would normally be professional work.

With OSC, I became responsible for that capability myself. The first OKR cycle set conservative goals; the system exceeded them. The second cycle is deliberately more ambitious and is already producing encouraging signals. But every positive number now creates a second question: what else moved, what alternative explanation exists, and are we measuring the thing we actually care about?

That is why I am increasingly reluctant to celebrate a metric in isolation.

A spike in impressions may be reach. It may also be one unusual post. A subscriber may be genuine audience growth. It may also be a friend. A website visit may be campaign response. Without attribution, it may only be temporally associated. An inbound conversation may indicate authority. It does not automatically validate a service.

The system should make optimism harder to fake.

So far, no major hypothesis has failed; and that is not evidence of perfection

There is an obvious credibility problem in writing about falsification when I cannot point to a spectacular hypothesis that the system has killed.

There has not yet been one.

There have been adjustments, stronger and weaker signals, changes in emphasis and increasingly explicit decision rules. But no foundational OSC hypothesis has yet met a clear negative validation strong enough to require a major pivot.

I do not want to manufacture one for narrative symmetry.

Readiness to stop is not the same as evidence that you have stopped. The useful discipline is to make sure that, when contradictory evidence arrives, the operating model is capable of acting on it.

That means defining disconfirmation conditions before they become emotionally inconvenient.

For an enterprise AI initiative, I would want at least six things to be explicit:

The last two are where the investment discipline enters. Evidence does not have to prove the final destination. It has to be strong enough for the next level of commitment.

Match the commitment to the evidence

This is where validation is often misunderstood.

If “validate the hypothesis” means “prove the entire business case before doing anything,” then it becomes another form of corporate paralysis. New ideas rarely arrive with complete evidence. AI capabilities, costs and user behavior are changing too quickly for certainty to be a practical prerequisite.

The better rule is to match the size of the commitment to the strength of the evidence.

A weak but plausible hypothesis may deserve research or a cheap prototype. A promising signal may deserve a controlled workflow experiment. Repeated evidence may justify integration and operating-model changes. Transformation-level commitment should require a materially stronger value case than a demo requires.

McKinsey’s July 2026 research on moving from AI adoption to enterprise impact reinforces the distinction. Its survey separates organizations into enablement, automation and reinvention horizons and finds that individual AI adoption does not automatically translate into enterprise value. The organizations furthest ahead redesign workflows and operating models around value rather than simply distributing tools. The research is a useful reminder that activity and value are different measures.

That also connects to a broader argument I made in AI Is a Challenge to the Shape of the Firm: AI becomes strategically consequential when it changes how work and value creation are structured, not merely when it makes an existing task faster.

The more consequential the change, the higher the evidence threshold should become.

Use AI twice

The lesson I am taking from the OSC experiment is not “start a blog.” It is not even “apply Lean Startup to AI.”

It is to use AI twice.

Use it first to reduce the cost of curiosity. Start the idea. Examine a possibility that would previously have required too much research, coordination or specialist input to justify exploring. Let cognitive leverage make small experiments economically rational.

Then use it again to reduce the cost of doubt. Ask what would make the idea wrong. Compare the success metric with adjacent metrics. Search for contradictory evidence. Separate correlation from attribution. Bring in an agent whose job is not to support the initiative but to challenge its current direction. Make the system remember the assumptions you will otherwise conveniently forget.

That creates a different sequence from the conventional transformation roadmap:

The loop can move quickly because the commitment remains proportional to what is known.

And the discipline can become stricter precisely because AI makes the analytical overhead cheaper.

That is the paradox I find most useful: AI allows us to begin before we know enough, but it also removes many of our excuses for continuing when the evidence no longer supports us.

Take one AI initiative currently labelled “high priority” in your organization. Rewrite it without naming the technology first. State the current constraint, the expected outcome, the causal mechanism, the evidence that would justify another level of investment, and the result that would make you stop.

If that exercise is difficult, you may have a compelling AI use case.

You do not yet have an AI value hypothesis.