Your AI Programme Isn't Failing Because of the Technology. It's Failing for the Same Reason Your ERP Rollout Did.

The conclusion first: AI adoption is not struggling because the technology is immature or the risk is unprecedented. It's struggling because organisations are repeating the same root causes that also sank a majority of ERP rollouts in the 1990s and 2000s, and a large share of cloud migrations in the 2010s: unclear sponsorship, bolt-on thinking instead of redesign, underfunded change management, weak data and process readiness, and business cases built on vendor timelines instead of realistic ones. The technology changes every decade. The failure mechanics don't.

If you've lived through one of these three waves, none of this will surprise you. What should concern you is that we keep rediscovering it as if it's new.

The numbers, side by side

Start with what the data actually says, not the hype cycle.

ERP (1995–2010): Gartner's research finds that by 2027, more than 70% of recently implemented ERP initiatives will fail to fully meet their original business case goals, with as many as 25% failing so badly they're abandoned outright. Budget overruns of 300–400% and timeline extensions of roughly 30% are typical. Gartner also finds 75% of ERP strategies aren't strongly aligned with the business strategy they're meant to serve, a sponsorship and governance gap dressed up as a technology problem.

Cloud (2010–2020): After five years of decline, wasted cloud spend has climbed back to 29%, and 17% of organisations exceeded their cloud budgets last year, largely driven by AI workloads bolted onto infrastructure that wasn't built for them. Lift-and-shift migrations, moving workloads to the cloud without re-architecting them, complete three times faster than refactoring but run 40% more expensive on an ongoing basis. That mismatch is now driving a repatriation trend: a quarter of organisations have pulled at least one workload back on-premises, and 86% of CIOs surveyed by Barclays said they planned to move some workloads back.

AI (2023–2026): McKinsey's State of AI research finds 88% of organisations now use AI in at least one function, yet only a third are scaling it enterprise-wide and two-thirds remain stuck in pilot or experimentation. Only 39% report any enterprise-wide EBIT impact, and 51% say AI use has already had a negative impact somewhere in the business. MIT's widely cited "GenAI Divide" study found 95% of generative AI pilots deliver no measurable P&L impact, attributing this to what it calls a "learning gap": tools that don't adapt to how the organisation actually works. Gartner predicts over 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls, governance failures, not engineering ones.

Different technology. Same shape of failure.

Six failure patterns, three decades

Same causes, every time. Here's how each one shows up across ERP, cloud and AI:

1. Sponsorship gap. Waste Management sued SAP for $500m after the vendor sold a "ready-made" solution that didn't exist, with no business owner in the room able to challenge the pitch; Gartner separately finds 75% of ERP strategies aren't aligned with overall business strategy. Cloud programmes frequently ran as infrastructure projects rather than business transformations, leaving governance an afterthought. Today, Gartner links agentic AI cancellations to "unclear business value", a sponsorship and governance failure, not a model failure.

2. Bolt-on vs redesign. Hershey compressed a 48-month ERP rollout to 30 months and ran ERP, CRM and SCM live simultaneously without redesigning the underlying process, triggering a 19% profit decline. Lift-and-shift cloud migrations run 40% more expensive long-term because the process was moved, not rebuilt. MIT finds generic AI tools stall in the enterprise for the same reason: they're layered onto unchanged workflows.

3. Change management underinvestment. Gartner ties ERP failure directly to underestimating the scale of organisational change required, not the software configuration itself. Cloud skills gaps compounded cost overruns as teams managed cloud-native environments with on-prem habits. BCG finds only companies that pair AI with people investment and workflow redesign generate value at scale; most don't.

4. Data/process readiness. Poor data migration is a widely documented ERP failure cause across implementation retrospectives. Data quality is now named a top concern for cloud-based AI workloads specifically. Same root issue, one wave later: AI trained or deployed on the same siloed, messy data.

5. ROI/timeline mismatch. Hershey's "original sin" was a Y2K deadline that overrode a realistic schedule. Vendor cost projections routinely undershoot actual cloud spend, which is part of what's now driving repatriation. MIT finds over half of GenAI budgets go to sales and marketing tools while the highest-ROI use cases sit in back-office automation.

6. Pilot-to-scale gap. ERP failures tended to surface post-go-live rather than in a public "stuck in pilot" phase. Shadow IT and PoC sprawl were a known cloud pathology before governance caught up. Two-thirds of organisations are still in AI pilot mode; only 5–5.5% (McKinsey, BCG) are scaling and capturing value.

What's genuinely different this time

Forcing a false equivalence would undercut the argument, so three things are worth naming honestly. AI's pace of change is faster than either prior wave, shrinking the window to course-correct. Model non-determinism means the same prompt can produce different outputs, which breaks some traditional QA and change-control assumptions ERP and cloud didn't have to contend with. And data/IP risk and regulatory uncertainty are live in a way they weren't for ERP master-data projects. These are real differences. They change the risk profile, not the underlying failure mechanics.

Why this take is different

Plenty of commentary already draws the "AI is the new cloud" comparison. Fewer connect all three waves into one operating framework, and fewer still write it from inside delivery, where the failure isn't a statistic, it's a steering committee that won't name an accountable business owner, or a change budget that gets cut the moment the programme slips. The pattern-recognition value here isn't "history repeats." It's that the same diagnostic that would have flagged Hershey's or Waste Management's ERP programme as high-risk before go-live works just as well on an AI programme today, and it's cheaper to run that diagnostic in month one than to run a retrospective in year two.

Worth noting: the same analyst firm, Gartner, is the source for both the ERP and agentic AI failure figures above. That's one consistent benchmark measuring both waves the same way, not two cherry-picked sources stitched together to make the parallel look tidier than it is.

The takeaway for delivery leaders

Before your next AI investment case goes to committee, run it against the causes above: who owns this outside IT, what process is actually being redesigned, what's budgeted for change management (not just licences), how ready is the data, and does the ROI timeline reflect delivery reality or vendor slideware. The organisations that got ERP and cloud right didn't have better technology. They had better sponsorship, better-scoped redesign, and the discipline to fund the change, not just the deployment. That playbook is available again. The only question is whether we use it before or after the retrospective.

Assumption flagged: this piece treats "failure" consistently across all three waves as "did not achieve the stated business case," acknowledging that source definitions vary (abandoned vs. over-budget vs. under-delivered). Figures cited are the most recent publicly available from Gartner, Flexera, Barclays, McKinsey, MIT, and BCG as of July 2026.