TL;DR
- Expensive engineering was doing the filtering. Nobody noticed, because the filter lived in the budget, not in the organisation.
- Cheap prototypes do not just waste money. They increase attachment. A build you commissioned is evidence you were right, whether or not you were.
- The standard prescription, protect time for problem definition, assumes slack that mid-market services firms do not have.
- When people in those firms name their obstacles, the obstacles are structural, not analytical. Compensation, KPIs, career paths, unbilled hours. Not an inability to think of good ideas.
- The gate has to be rebuilt deliberately, as a decision owned by the people who carry the consequences, not as a veto held by an advisor.
Harvard Business Review published a piece on 12 August called AI Makes Building Easy. Choosing What to Build Is Harder. The argument is that as AI commoditizes execution, competitive advantage increasingly comes from deciding what is worth building and why. I agree with all of it. The diagnosis is right and it is not a controversial thing to say anymore.
What I want to add is a sharper version of the same point, and then a reason the prescription does not land where I do most of my work.
The sharper version is this. Expensive engineering was doing the filtering, and almost nobody noticed. Six months of developer time meant somebody in finance asked what the thing was for. The answers were often bad. Sometimes the business case was written backwards from a decision already made. But the question got asked, on a schedule, by a person whose job was to ask it. Two weeks of prototyping asks nothing of anyone. The discipline was never in the organisation. It was in the budget line. Remove the cost and you remove the filter, quietly, because an absent gate does not announce itself. Nothing goes wrong. Things just start going forward.
I have watched this from close range in transformation work with mid-market services firms. The go and no-go structure is something you have to design deliberately. It is not sitting there in the organisation waiting to be used. In one engagement, the most useful early output was not a product decision at all. It was noticing that the ideas the company generated internally had inherited the shape of its services business. Large ticket sizes, long sales cycles, complex multi-stakeholder sales. Those are reasonable attributes for a consulting engagement. They are fatal for the fast-moving product the company had said it wanted. The gate caught the mismatch. Whether the gate then got used the way a gate is supposed to be used is a separate question, and a harder one. Designing a decision point and actually stopping at one are different problems.
The prototype is not neutral evidence
This is the part I think gets missed, and it is where I would push the HBR framing further.
On 14 August, HBR published research by De Freitas, Israeli, Nave, Timoshenko and Toubia on the innovation problems AI cannot solve. Their finding on ideation is specific and it is the most useful thing I have read this year. Generative models steer teams toward familiar ideas and make people more fixated on those ideas. In screening, polished AI-generated pitches get mistaken for better ideas. Their framing is that AI acts on the human bottlenecks already buried in the process, and it does not always dissolve them. Sometimes it reinforces them.
That is their research, and it is about AI-generated ideas. What follows is my extension of it, and I want to be clear about which is which.
The same mechanism applies to prototypes, with a firmer grip, because the effort was yours. A working prototype is not neutral evidence about whether an idea is good. It is evidence that you were right, whether or not you were. The build becomes the sunk cost, and cheap does not mean free. A two-week build costs less money and produces more conviction than a six-month one, because ease reads as confirmation. It went smoothly, so the idea must be sound. Speed gets misread as signal.
Then there is what the artifact does in a room. A paper prototype looks like a question. Something clickable looks like an answer. Once there is a working demo in a board meeting, the conversation quietly moves from whether to build this to when we launch. Nobody decides to skip the validation step. The demo just makes the step feel like it already happened.
This is the specific danger for a services company leader right now. A prototype took two weeks instead of six months, and the speed of the prototype is being read as evidence that the idea is good. It is evidence of nothing except that the tooling works.
The pattern is visible at scale
In June 2025, Gartner forecast that over 40% of agentic AI projects would be canceled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. It is a forecast rather than a measurement, and it is now more than a year old, so treat it as a directional read rather than a result.
The useful part is what is missing from that list of three. Model capability is not on it. All three named causes are scoping and governance failures. A better model does not fix any of them. Drop the best model available into a project with no defined outcome and no owner and you get a more articulate version of the same problem.
Why the standard prescription does not land in the mid-market
The HBR prescription is to protect time for problem definition and build teams around complementary perspectives. Both are correct. Both presume an innovation function with slack in it.
A two hundred person services firm running at high utilisation has no slack. Every hour spent on problem definition is an unbilled hour, and somebody’s utilisation number absorbs it. That is a compensation and structure problem before it is a thinking problem, and no amount of better framing fixes it.
I have seen this stated plainly by the people who would have to do the work. When employees in these organisations map the transition themselves, in their own words, the obstacles they raise are overwhelmingly structural rather than analytical. Compensation. How performance will be measured. Career paths. Additional workload sitting on top of billable targets. Who owns the roadmap. What happens to them personally if the initiative gets stopped. Nobody’s stated blocker is an inability to think of good ideas. The same groups will generate credible product concepts inside a single session. The analytical work is the easy part and everyone in the room knows it.
So when a company says it needs help figuring out what to build, that is frequently not the constraint. The constraint is that nobody can afford to be the person who spends three unbilled weeks discovering the answer is no.
Rebuilding the gate without appointing a veto
Cheap building removed the natural filter, so the filter has to be reinstalled on purpose. But reinstalling a gate is not the same as appointing somebody to hold a veto, and an outside advisor should not hold one.
The version that works is two-part. Whoever is advising surfaces the trade-offs legibly and flags opinion as opinion rather than smuggling it in as finding. The people who carry the consequences make the call. A recommendation written as “it is the opinion of the consulting team that” does more work than the same recommendation dressed up as a conclusion, because it is honest about its own standing and therefore harder to wave away. A finding you cannot argue with is a finding nobody owns.
The other half of that is knowing what to ask for. One of the fields on the prototype template I use asks teams why the thing might fail. It is a plain question and it produces more information than anything else on the page.
That is the next post. How a team answers “why might it fail?” tells you more than the idea does.