AI can’t fix numbers you don’t trust.

An AI layer inherits the numbers under it. Feed it figures nobody trusts and it will write clear, confident sentences about figures nobody trusts. The prose improves. The decisions do not.

That is why Automate is the fifth step of the Decision System and not the first. The numbers are calculated from your data by fixed rules. The AI layer reads them, writes the weekly brief and names the action. It does not compute them, and it cannot repair them.

What happens when the order is reversed

RAND, studying why AI projects fail, reports that “by some estimates, more than 80 percent of AI projects fail—twice the rate of failure for information technology projects.”

Google Research interviewed 53 practitioners building high-stakes AI and published the work at ACM CHI in 2021. They named the pattern data cascades: “compounding events causing negative, downstream effects from data issues.” Of those 53 practitioners, 92 percent had lived through one. The paper calls data cascades pervasive, invisible, delayed and often avoidable. Invisible is the word that matters here. A wrong figure does not announce itself at the point it enters. It surfaces later, in a decision nobody connects back to it.

A wrong figure does not announce itself at the point it enters.

The confidence gap

In a Gartner survey of 1,203 data management leaders, 63 percent either do not have the right data management practices for AI or are unsure whether they have them.

Drexel University’s LeBow College of Business and Precisely surveyed 505 data and analytics leaders at major global enterprises for the 2026 State of Data Integrity and AI Readiness. Across that population, 88 percent say their data is ready for AI, while 43 percent name data integrity as an obstacle. The confidence and the complaint come from the same survey.

Ayanna Howard, writing in MIT Sloan Management Review in 2019, put the mechanism plainly: AI “can easily turn valuable data into decision-making mistakes and false truths.”

What has to be true first

The first four steps are what make the fifth safe to run.

• Decisions written down, so the brief has something to be about.
• Metrics nobody acts on removed, so nothing unowned is carried forward by default.
• Tools connected, so finance, sales and operations are reading the same figures.
• Screens built around questions, so an answer already exists for the layer to report on.

Thomas C. Redman, in MIT Sloan Management Review in 2017, estimates the cost of bad data at 15 to 25 percent of revenue for most companies, and estimates that two thirds of those costs can be identified and eliminated permanently. That is what bad data costs. It is not what we promise to return. Steps 01 to 04 go after the same costs, with fixed rules.

Put the AI layer first and it writes a confident brief on numbers nobody checked. Put it fifth and it has something worth reading.

Sources

1. “The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed”, RAND Corporation, 2024
2. Sambasivan et al., “Everyone wants to do the model work, not the data work: Data Cascades in High-Stakes AI”, Google Research, ACM CHI 2021
3. “Lack of AI-Ready Data Puts AI Projects at Risk”, Gartner, 26 February 2025
4. “2026 State of Data Integrity and AI Readiness”, Drexel University LeBow College of Business with Precisely
5. Ayanna Howard, “Demystifying the Intelligence of AI”, MIT Sloan Management Review, 2019
6. Thomas C. Redman, “Seizing Opportunity in Data Quality”, MIT Sloan Management Review, 27 November 2017