Where We Actually Are
Halfway through 2025, the enterprise AI landscape has matured enough to see clear patterns. The hype cycle has not ended, but it has differentiated: some companies are generating real returns from AI while others are stuck in an expanding cycle of pilots and evaluations. Here is what we are seeing across our client base and the broader market.
What Has Changed Since January
The model capability gap is closing. The difference between frontier proprietary models and top open-source alternatives has narrowed meaningfully. Llama 3 variants, fine-tuned for specific tasks, are competitive with GPT-4o and Claude 3.5 Sonnet in many enterprise use cases. This has shifted the build vs. buy conversation: more companies can seriously consider self-hosted models for production use cases, whereas six months ago this was a realistic option only for the largest engineering organizations.
Agentic AI has hit the trough of disillusionment. The early enthusiasm for autonomous AI agents has collided with production reality. Companies that deployed agents broadly are discovering the failure modes we have documented: cascading errors, cost explosion, and debugging difficulty. The survivors are those who scoped their agents narrowly and maintained strong human oversight. The industry is converging on a more realistic view: agents are powerful for specific, well-defined workflows, not for open-ended autonomy.
The compliance conversation has become urgent. With the EU AI Act moving toward enforcement and US regulatory activity accelerating at the state level, AI compliance has shifted from "future concern" to "current requirement" for many enterprises. Companies that treated compliance as a later problem are now scrambling.
What Has Not Changed
Data infrastructure remains the bottleneck. The most common reason AI projects fail is still not model capability or engineering talent. It is messy, fragmented, ungoverned data. This has not improved because it requires sustained, unglamorous investment that does not make for impressive board presentations.
The pilot-to-production gap persists. Despite growing awareness of this problem, most enterprises are still running more pilots than they are shipping to production. The organizational dynamics that create this gap, misaligned incentives, unclear ownership, insufficient production planning, remain largely unaddressed.
Talent is still the constraint. The shortage of engineers who can build production AI systems has not eased. If anything, it has worsened as demand has grown faster than the pipeline of experienced practitioners. Companies that invested in training their existing engineers are starting to see returns.
The Emerging Divide
We are seeing a clear bifurcation in the market:
Companies that ship share common traits: they start with specific business problems, invest in data infrastructure before AI capabilities, enforce hard timelines on pilot decisions, and measure AI impact in business terms.
Companies that pilot share different traits: they start with technology excitement, skip data infrastructure investment, allow pilots to run indefinitely, and measure AI by technical metrics that do not connect to business outcomes.
The gap between these two groups is widening. Companies in the first group are building organizational muscle for AI delivery that will compound over time. Companies in the second group are spending more and learning less with each passing quarter.
What to Focus on for the Rest of 2025
- If you have pilots running, enforce 90-day decisions. Ship or kill.
- If you have not invested in data infrastructure, start now. Every month of delay increases the cost.
- If you are not tracking AI ROI in business terms, start measuring before your next board meeting.
- If you are relying on external hires for AI capability, invest in internal training as a parallel strategy.
The second half of 2025 will reward companies that have built disciplined AI delivery practices. The window for catching up is still open, but it is narrowing.