The Value of Being Wrong
At the start of 2025, we published our predictions for the year in enterprise AI. Some were right. Some were partially right. And some were flat-out wrong. The wrong ones are more instructive than the right ones, so let us start there. Any advisory firm that only highlights its correct predictions is selling you confidence, not insight.
Where We Were Wrong
- We overestimated enterprise AI deployment speed. We predicted that by end of 2025, a majority of Fortune 500 companies would have production AI generating measurable business value. The reality: a significant minority have achieved this, but the majority are still in various stages of experimentation and piloting. The organizational change required to deploy AI in production is harder and slower than we, or frankly anyone in the market, anticipated. Technology was never the bottleneck. People and processes were.
- We underestimated the open-source model ecosystem. We expected proprietary models from the major labs to maintain a wide capability lead over open-source alternatives throughout 2025. The gap narrowed faster than we projected, and in some domains it effectively closed. Open-source models are now viable for a broader range of enterprise use cases than we predicted, which fundamentally changes the build-versus-buy calculus we advise clients on.
- We overweighted the regulatory headwind. We predicted that AI regulation, particularly the EU AI Act, would meaningfully slow enterprise AI adoption in Europe. While compliance costs are real and non-trivial, most enterprises are treating regulation as a design constraint rather than a reason to delay. The expected chilling effect did not materialize at the scale we predicted. Companies adapted faster than regulators moved.
Where We Were Right
We predicted that AI infrastructure would see massive investment that would outpace near-term demand. We predicted that the talent constraint would be more binding than the technology constraint for enterprise adoption. And we predicted that vertical, domain-specific AI applications would generate more enterprise value than horizontal AI platforms. These played out largely as expected.
What We Learned
The biggest lesson: organizational readiness is the rate limiter for enterprise AI, not technology capability. Every prediction we made that assumed technology adoption would drive organizational change was too optimistic. Every prediction that assumed organizational factors would constrain technology adoption was roughly right.
This has direct implications for how we advise clients going forward. Technology evaluation and selection is maybe 20% of the work of AI transformation. The remaining 80% is organizational: workflow redesign, talent development, change management, governance, and cultural shift. That is where we are concentrating our energy, and our clients' budgets, in 2026.