The Golden Dome on a Sinking Foundation
Companies across every sector are throwing massive budgets at anything with an AI label, often driven more by the fear of missing out than by a clear, strategic roadmap.
Because of the huge pressure on companies as potential clients, for tech providers and agencies, it is an unprecedented era of easy monetization. But a critical shift is happening, at least among ethics-driven professionals: the courage to say “No”.
A consultant recently shared, “I could have billed 500k CHF for an AI implementation last week, but I refused.”
The client’s VP and board wanted AI. Inspection found ERP master data in very poor condition. The pressure was for visible software, while the proposed system depended on information the organization could not reliably supply.
A consulting firm can be rewarded for selling an implementation while the client needs someone to stop a failed initiative before it reaches production. The fee creates visible activity. Repairing master data creates less impressive evidence of progress, even when that is the work a business decision requires.
Software depends on the foundation
AI software is step 5, not step 1. Before advanced analytics, an organization needs data collection, infrastructure, cleaning, governance, and business intelligence. Later capabilities depend on those earlier conditions. This sequence can be called the Data Hierarchy of Needs.
The ERP example makes the dependency concrete. Incomplete, duplicated, or ownerless master data weakens the decision built on a model’s output. The model can run and return an answer while the result remains unreliable, because the records behind it do not describe the business consistently. Putting algorithms on top of broken data only accelerates the mess.
Get calculito’s stories in your inbox
Join Medium for free to get updates from this writer.
The unresolved questions are operational: who owns a customer record, which product is the real one, and what a missing value means. Buying the system early pushes those questions into a later phase, where they become defects, exceptions, and arguments about the model.
The model is only the midpoint
Deployment creates another body of work. Monitoring, retraining, testing, integration, workflow design, training, and trust-building continue after the software is installed. Inputs change, as does the relationship between those inputs and the decision the model is meant to support. The system has to be tested against those changes and kept connected to daily work.
Otherwise, the implementation becomes a golden dome on top of a sinking house of cards. AI outputs sit raw behind an API endpoint while business users cannot use them. Staff bypass tools they do not understand or trust, and the organization records spending without changing operations.
The procurement sequence is specific. Audit the ERP master data for missing ownership, duplication, and missing values. Budget for data ownership, deduplication, governance, and a single source of truth before model work begins. Require monitoring, workflow integration, and staff training in the AI plan, so the implementation budget covers the conditions of use.
The available evidence does not establish whether the 500k CHF figure was a documented commercial offer, whether the VP or board accepted this sequence, or whether an implementation produced a measured result. The lifecycle describes obligations surrounding deployment, not evidence that this client succeeded or failed.
Who carries the failure
Budget holders control the order in which money is released. Directors and employees often inherit the consequences. A director faces the initiative in operating terms, while staff must either incorporate the system into their work or find a way around it. They have less control over whether the foundations were funded before the software was approved.
When the data is weak, the model becomes the visible suspect. When adoption stalls, users become the visible problem. The spending decision disappears behind the technical incident and the workflow complaint.
The people who authorize the purchase control when risk becomes visible without controlling who must absorb it. The sequence of spending decides who gets to call a failure technical and who has to live with it.
Before approving the next AI line item, ask: are we funding software that works, or are we paying for a golden dome to cover a broken foundation?
