In innovation projects, the technology is the easy part 😉

The hard part is everything underneath: the data, the measurement, the discipline.

In innovation projects, the technology is almost never the hard part. Carnegie Mellon’s Software Engineering Institute and Accenture just put numbers behind that, with a new maturity model for AI adoption 🙌

Insights:
☞ 95% of generative-AI initiatives return nothing measurable.
☞ 8% of companies have actually scaled it.

The failure sits underneath the tech,
in the gap between the vision on the slide and the result you will experience.

Accenture’s Manish Sharma helped build the model, and he says something unequivocal: “focus on high-level strategy without considering the engineering rigor that organizations need to actually scale”.
☞ So the deck gets written and the thing never gets built.
☞ 61% of organizations carry a formal AI strategy they have never finished implementing.
☞ Fewer than a third tie it to value and return.
The few that break through turned strategy into disciplined execution and measurable outcomes.

The project's lead, Ipek Ozkaya, puts it in one line: “Our industry often assumes discipline can be automated away.”

What the model actually measures: it has you set a baseline before you start, then draw your metrics from the workflow you are changing, things like accuracy, latency, adoption rate, cost per outcome, and ROI against a target you fixed up front. It defines its criteria in plain terms, such as reproducibility, auditability, and cost efficiency. And it does not let you average your way up: your weakest practice is your score.

When Decathlon digitized its stores, the visible part was the new services. Click and collect. Order a color the shop doesn’t stock and have it shipped home. Software that finally understood the products, the uses and the sizes, to augment our store staff.
The part that carried all of it was the product data. Structured cleanly, across every brand and every product, so a tent meant the same thing in every system it touched. Which fields. Which words. Which story each product had to tell. I started from precise measures and the precise results I expected, and built back from there.

Technology is a lever, not a goal.

This is what I do, on every project, and most of all the ones built on AI. I turn the ambition into a plan with numbers on it. I price the bet before the budget goes out. I measure the uncertainty instead of pretending it away. I make it repeatable, so the next project is not a fresh gamble. Discipline is not the brake on scale. It is the thing that scales.

I find the same pattern in my own practice of alpinism. You can’t run a serious climb without ambition, a plan with numbers, a team, and room to pivot. Ed Viesturs is the perfect example of it. He climbed all fourteen 8,000-meter peaks without bottled oxygen. He chose his partners against written criteria before anyone committed, and tested the team at low altitude first, because a team that fails on an easy climb will not hold on a hard one. He studied the odds: he saved Annapurna, the deadliest of the fourteen, for last, turned back twice when conditions were wrong, and went only when his preparation made it survivable.

Run your projects the same way. Define the starting point and what success looks like, in numbers and in words, before you start. My one rule: fail fast, kill cheap.

Find the model here :

Sources

• 95% of generative-AI initiatives return nothing measurable: MIT NANDA, The GenAI Divide: State of AI in Business 2025 (Challapally et al., 2025). The figure covers generative-AI pilots with no measurable bottom-line impact.

• 8% have scaled across the enterprise: Accenture, The Front-Runners’ Guide to Scaling AI (2025).

• The maturity model, its measurement framework, the Manish Sharma framing, the Ipek Ozkaya quote, and the survey figures (61% with a strategy not fully implemented, fewer than a third prioritizing value and ROI): SEI and Accenture, AI Adoption Maturity Model v1.0 (2026), drawing on the SEI January 2026 survey of 600 practitioners. Manish Sharma, Chief Strategy and Services Officer, Accenture; Ipek Ozkaya, project lead, CMU SEI.

• Annapurna, deadliest of the fourteen eight-thousanders, roughly one death for every three who summit (about 32% historically, lower today): National Geographic; NASA Earth Observatory.

• Ed Viesturs, first American to climb all fourteen eight-thousanders without supplemental oxygen, Annapurna last in 2005 after turning back in 2000 and 2002. “Getting to the top is optional. Getting down is mandatory.”: No Shortcuts to the Top (2006); National Geographic, 2005.

• https://www.accenture.com/us-en/insights/ai-data/front-runners-guide-scaling-ai

• https://www.sei.cmu.edu/news/sei-and-accenture-release-ai-adoption-maturity-model-to-help-organizations-scale-ai-with-predictable-outcomes/

• https://www.decathlon.media/dossier-presse/2015/innovation-magasin-borne-plus.html