77% of AI’s Hardest Costs Are Invisible. The Most Expensive One Comes Before You Start.
What Stanford Found
Stanford’s Digital Economy Lab spent five months interviewing executives at 41 organizations to document 51 enterprise AI deployments that delivered measurable value. The Enterprise AI Playbook, published in April 2026 by Pereira, Graylin, and Brynjolfsson, reports that 77% of the hardest challenges in AI deployment are invisible: change management, data quality, process redesign. The authors call these “invisible costs” because they sit outside the model spend that the business case is built on.
A separate finding in the same report shows that 61% of successful projects had a prior failed attempt whose cost never appeared in the final ROI. Earlier research the authors cite found that for every dollar of tangible technology investment, companies spend up to ten dollars on intangibles. The visible dollar funds the model. The invisible dollars fund the organizational work that makes the model usable.
Those findings help management set realistic expectations about what AI success actually requires. The most interesting observation to me, however, wasn’t part of the formal findings. In case after case, AI was applied to work the organization already understood: invoice processing, alert triage, support automation, code migration, content production, clinical documentation. In each case, the company addressed a known problem, a known customer, a known workflow, and a known definition of done.
AI does not create clarity. It magnifies the clarity or confusion that already exists in your systems. In the Stanford sample, clarity was present. The front end of innovation, by contrast, is what Peter Koen famously called the “fuzzy front end.” The fuzziness, however, is a function of how most companies execute the work, not a property of the work itself.
The Risk Iteration Was Never Designed to Mitigate
Companies can gain clarity at the front end of innovation by determining the business objective, the target customer, the job to be done, and where customers struggle before generating ideas. This provides the team with customer instructions about the outcomes that matter most, so the creative work of ideation is aimed at real targets rather than assumptions.
I have found that the most helpful definition of the front end of innovation, if you want to turn innovation into a repeatable business process, is “the process of discovering the target customers’ unmet needs and then generating ideas to address them.” The sequence matters. Discovery comes first.
This sequence addresses the two types of product innovation risk: market and solution risk. Market risk is the risk of misunderstanding what customers need so they don’t buy what you make. Solution risk is the risk that, even though you understand what customers need, you can’t make or deliver it. The two are independent. A team can succeed on one and fail on the other.
Why AI Makes This Worse
The iterative methods that have shaped innovation practice for the past thirty years (Lean Startup, Customer Development, Design Thinking, the Business Model Canvas, Effectuation) primarily mitigate solution risk. They take an idea, build a Minimally Viable Product (MVP), test it, refine it. Design Thinking begins earlier, with user empathy, and reduces market risk through qualitative understanding. What none of these methods produces, however, is the precise, solution-free, prioritized statements of customer needs that enable teams to engineer product-market fit at concept creation.
The data reflects this. CB Insights’ March 2026 analysis of 385 VC-backed startup failures since 2023 identified the top root causes of failure as poor product-market fit (43%) and bad timing (29%), both of which are market-risk failures. Furr and Ahlstrom, in Nail It Then Scale It, report that 95% of new business failures are attributable to market risk: misunderstanding customer needs and the competitive situation. Faster iteration on solution features is not an effective way to reduce market risk.
Because the front end of innovation is not a known process for most companies, AI raises the risk of innovation failure and the waste of resources rather than lowering either. Stanford documented organizations compounding advantages inside workflows they already understood, which is the work AI accelerates most reliably. Applied to work the organization does not yet understand, AI will accentuate that confusion. Mitigating market risk and mitigating solution risk are different tasks. Iterating on the solution is generally ineffective for reducing market risk. Speed on the wrong axis does not close the gap.
Most of the organizations Stanford documented already possessed what innovation initiatives typically lack: validated understanding of the problem, the customer, and the definition of done. For companies pursuing growth through new offerings, building that understanding is a necessary part of the “invisible” work that comes before AI can help.
Lean JTBD OS™ is the front-end innovation operating system that closes that gap. It does not replace the iteration methods that have served the field. It adds the upstream discovery discipline they were not designed to provide, so that everything those methods do downstream is aimed at a validated target. I will be launching an ISO-certified course on Lean JTBD OS in collaboration with the Global Innovation Management Institute (GIMI) later this spring or early summer. If this resonates with you, you can sign up for priority notification here.
