Why CIOs are right to move carefully on AI

Based on Benedict Evans' 'AI Eats the World' presentation, November 2025

The return of Benedict Evans’ brilliant twice-yearly tech trends presentation, this time titled “AI Eats the World” was a moment to reflect on our experience at Valliance working with CIOs to build AI strategies in an environment where, as Evans’ analysis reveals, we’re still collectively defining what “AI work” actually means.

As AI capability, functionality, buzz, and FOMO explode across the enterprise, “What’s our AI strategy?” has become a fixture of every boardroom agenda. But the gap between discussion and deployment remains stark. Evans cites Morgan Stanley data showing that only a quarter of CIOs have launched at least one LLM led project, with 40% not planning deployment until 2026 or later (Slide 57).

This isn’t caution; it’s wisdom. But what’s creating this hesitation? And why are traditional consulting approaches struggling to bridge the gap? In this post, I’ll address both questions.

The Three Fundamental Uncertainties

Drawing on Evans’ framework, we see three critical uncertainties that any serious AI business case must address simultaneously:

1. How far will it scale? (The Science Question)

Will LLMs continue improving with more compute and data, or are we approaching diminishing returns? Evans highlights that even the teams building these systems don’t know whether models will get “10x and then 100x and then 1000x better” or whether we’ll “flatten out” sooner than expected.

The implication: Don’t bet your infrastructure strategy on certainty about model capabilities in 2026. Any architecture must account for this fundamental uncertainty.

2. How is it useful? (The Product Question)

LLMs are what Evans calls “infinite interns”, they can do many things but lack judgment about when they should be used. As Evans observes, ChatGPT went from science project to 100 million users at unprecedented speed, but “the magic might not be useful, in that form, and it might be wrong.”

The implication: Start with use cases where errors are visible and consequences are manageable. This is discovery work, not deployment at scale.

3. How do we deploy it? (The Execution Question)

Are LLMs infrastructure (like databases) or platforms (like iOS)? Evans describes two extremes: either “a handful of giant, capable, capital-intensive, expensive and very general models” that everything plugs into, or a proliferation of specialised models distributed across the stack. The architecture question remains genuinely open.

The implication: Cloud adoption took 15 years to reach 30% of enterprise workloads. AI deployment will be measured in years, not quarters. This demands architectural flexibility over premature optimisation.

Why Traditional Consulting Models Face Structural Limitations

McKinsey, Bain, BCG, and the major system integrators excel at synthesis and frameworks. But Evans’ analysis of this “no one really knows” phase highlights why their traditional operating models face structural limitations in this environment.

Consider three core tensions:

Problem 1: Recommendations vs. Implementation Partnership

The traditional model separates strategy from execution: consultants write recommendations, you implement them. But when no one knows what works, recommendations are necessarily theoretical. Evans’ observation that Accenture delivered 300 projects for $300 million, averaging $1 million per project, suggests “a lot of pilots, not deployment.”

In a discovery phase, you need partners who will implement and iterate with you, not document a strategy for you to execute alone.

Problem 2: Billable Hours vs. Value Delivery

Time based pricing rewards hours spent, not outcomes achieved. When AI pilots fail regularly (and they do), hourly billing penalises the experimentation necessary for learning. You’re paying for discovery time, not for discovered value.

The model assumes the path is known and you’re paying to walk it. So clients are paying to find the path which requires a different commercial structure entirely.

Problem 3: Junior Leverage vs. Experienced Practitioners

Traditional consulting economics depend on junior leverage: partners sell, analysts deliver, and the model scales through training junior talent on client engagements. But in a genuine discovery phase where even the platform builders don’t know how their technology will evolve, you need experienced practitioners who’ve built and deployed systems in production, not analysts developing their first AI strategy.

Evans notes that “the single biggest business from this in 2024 might be for consultants explaining what it is.” There’s nothing wrong with education except when you need implementation.

What This Means for CIOs

If you’re a CIO moving carefully on AI deployment, Evans’ framework validates your caution. The uncertainties are real, the technology is still evolving, and the use cases are still being discovered.

But moving carefully doesn’t mean moving slowly, it means moving deliberately with the right partners.

Traditional consulting models were built for a different kind of problem: known solutions deployed at scale, where the challenge is execution, not discovery. AI in 2025 and beyond is the opposite: unknown solutions requiring discovery, where strategy and execution must happen together.

Valliance’s model is purpose built for exactly this moment.

• Strategic advisory on competitive positioning

• Architecture designed for model evolution, not locked into today's leaders

• Help identifying which AI capabilities are competitive differentiators vs. commodities

We help you move fast without moving blindly. We take strategic risk with you.