Every consultant and platform is selling 90-day AI reskilling. The data says skill decay starts at day 45 without continuous practice. Either the program has a 12-month retention scaffold or you’re paying for theatre.
Open any enterprise AI reskilling proposal from the past twelve months and you will find the same number: 90 days. It is the length procurement approves without escalation. It is the sprint that fits inside a quarter. It is also, according to a growing body of evidence, a fiction.
Hermann Ebbinghaus demonstrated the forgetting curve in 1885. The principle has not aged out: without spaced retrieval, humans lose 50–80% of new knowledge within 30 days. Apply that to a 90-day AI reskilling bootcamp and the arithmetic is brutal. By the time the program wraps, the skills taught in week one are already degrading. By day 180, without reinforcement, the average learner retains less than 28% of the material.
The vendors know this. The 90-day model persists not because it works, but because it is what procurement will sign. A 12-month program requires a different budget line, a different approval chain, and a different relationship between L&D and the business. Most organizations are not set up for that conversation — yet.
But the organizations that are having it are seeing dramatically different results. The 12-month scaffold model — where initial intensive training is followed by monthly lab exercises, quarterly skill assessments, and continuous micro-reinforcements — retains 92% of skill at the six-month mark. The 90-day sprint retains 28%.
You cannot reskill someone in a two-week bootcamp. In 2026, the most successful programs are 6 to 12 months long.impress.ai workforce intelligence report, April 2026
The cost difference is smaller than most L&D leaders assume. A 90-day sprint that produces 28% retention and requires re-training is more expensive per competent employee than a 12-month scaffold that produces 92% retention once. The real question is whether your organization measures cost per seat or cost per ready employee.
If you measure cost per seat, the 90-day sprint wins every time. If you measure cost per ready employee, it never does.
Source: Composite from Ebbinghaus (1885), updated with enterprise cohort data from impress.ai and internal Learnlytica benchmarks. Scaffold includes spaced retrieval, monthly labs, and quarterly skill assessments.
We killed our 90-day Databricks sprint in Q1. Replaced it with a 12-month program: 6-week intensive, then monthly lab challenges, quarterly proctored assessments, and a peer-teaching requirement. Retention at month six went from 31% to 89%. The CFO stopped asking why we spend money on training. She started asking how fast we can scale it.
Vikram’s team now measures one number per cohort: days from enrollment to first independent production deployment. The average dropped from 174 days to 68 days after switching to the scaffold model. The program costs 40% more per seat — but 62% less per production-ready engineer.
Before you sign the next reskilling contract, run these five questions past the vendor. If they can’t answer all five, you’re buying a 90-day sprint dressed up as transformation.
After internal data showed sub-40% retention at day 180, Correlation One is restructuring its enterprise data-science programs from 14-week sprints to 9-month scaffolded cohorts. The market is starting to correct.
New BCG research documents the “retention cliff” in enterprise AI training: skill proficiency drops 54% in the 60 days following a bootcamp with no follow-up. Programs with monthly reinforcement retained 4.1× more skill.
Under the EU AI Act, organizations deploying high-risk AI systems must demonstrate ongoing workforce competency — not just initial training. One-and-done 90-day programs will not meet the compliance bar.
NSDC’s draft framework for AI skill credentials requires re-certification at 6-month intervals. Vendors without longitudinal assessment infrastructure will not qualify as approved providers.
Learnlytica’s platform tracks skill retention at 30, 90, 180, and 365 days — with spaced labs, proctored assessments, and decay alerts built into every learning path. One dashboard. Real numbers. No theatre.
See day-180 data in action → or explore the platform