
AI model loyalty in companies barely lasts a few weeks
Databricks' index, built on usage data from more than 10,000 companies, shows rapid switching from one AI model to another, following release cycles. That volatility changes how teams should evaluate models and build around them.
Blog du Modérateur reports on a Databricks index built from the usage of more than 10,000 companies, which highlights how quickly businesses move from one AI model to another, in step with each new release. The key point is volatility: corporate loyalty to a given model appears measured in weeks rather than in years.
The report does not specify which models are involved, the size of the switches, or the sectors and geographies covered. It gives no breakdown by use case either, so it is impossible to say whether the churn is concentrated in coding, support, or content production. These gaps matter, because a team's response depends on them.
What this means for a small marketing team
For anyone building acquisition or monetisation workflows on top of a model, this observation is a planning constraint. The vendor you integrate in January may no longer be the default choice in March. That does not mean changing providers constantly; it means assuming change as a normal event and designing for it.
AI is now embedded on both sides of the funnel: generating ad variants, drafting organic content, scoring leads, powering support or personalisation on a paid product. When the underlying model moves, quality, latency and cost per task can shift before anyone updates a strategy document.
Several practical consequences follow. Outputs are not stable enough to be judged once and forgotten. Prompts that work well on one model can behave differently on another. And unit economics built on a vendor's price list need a margin of safety, since prices and product tiers move at the same rhythm as releases.
Building for swapping, not for permanence
A pragmatic approach is to separate the workflow from the model. Keep prompts, evaluation cases and expected outputs in a versioned layer that any provider can plug into. Keep a small set of reference tasks, ideally tied to real business outcomes such as click-through on ad copy or conversion on a landing page, and re-run them when a model changes.
A short checklist for this week:
- List every place an AI model touches acquisition or revenue, including forgotten scripts and no-code automations.
- Pick three representative tasks and record current output quality, cost and latency as a baseline.
- Store prompts and expected outputs outside the vendor's interface so switching is an afternoon, not a project.
None of this requires a large team. It requires treating the model as a replaceable component rather than a fixed foundation. The Databricks index suggests the market is already behaving that way; the question is whether internal processes follow.
Also worth reading
-
Web IQ : Microsoft ouvre l'index de Bing à tous les agents IA sur Azure
Microsoft rend Web IQ, son API de grounding basée sur l'index de Bing, accessible à tous les clients Azure depuis le 17 novembre 2026, avec un mode Auto par défaut et une baisse de prix non chiffrée. Pour les équipes SEO et contenu, la question du référencement glisse vers celle de la citabilité par les agents.
-
Meta's Muse agent raises data concerns among employees
Meta's new AI agent Muse is popular with users but has triggered internal concern over personal data protection, according to Siècle Digital. The report does not detail the incidents, their scale or any response from Meta.
-
Microsoft et 40 organisations pour vérifier les agents d'IA
Microsoft a annoncé à Ignite une initiative de transparence réunissant plus de 40 organisations pour permettre une vérification indépendante des agents d'IA. L'éditeur a aussi mis en préversion un score de risque et des contrôles de confinement, sans prétendre révéler le comportement du modèle sous-jacent.


