Antoni Kozelski, founder of Vstorm, an agentic AI company that helps companies ethically implement agentic AI and transform their business.
MIT’s Project NANDA found that 95% of generative AI pilots deliver no measurable impact on the profit and loss statement. However, BCG researchers report that the small group of companies leading on AI achieve 1.7 times the revenue growth of what they refer to as the “laggards.” And yet the same technology produces both outcomes.
In our own work, we find the difference is rarely the model. It’s the strategy: which process to solve, in what order, and who owns the result once it’s live. About 70% of the companies that came to us last year, by our own estimate, arrived after a previous failed attempt, either internal or outsourced. Almost none of those were technology failures.
The Structural Case For Leading With Strategy
The instinct to run an agentic project like any other software delivery is the first and most expensive mistake. Conventional software is predictable: The same input returns the same output, scope can be fixed up front, and an established framework carries the work to completion. Agentic systems behave differently. They’re non-deterministic, they reason across several steps, and they draw on multiple live systems at once. The certainty that ordinary delivery assumes as default is absent.
No one has yet written the agentic playbook. In web or mobile development, a team can reach for a settled reference architecture; agentic AI has no equivalent. Open standards are still being drafted. The Agentic AI Foundation, hosted by the Linux Foundation and launched in late 2025, exists to coordinate them, and we contribute to that work.
When no reference architecture exists, every implementation decision is a strategic one, whether a team recognizes it or not.
Where Most Projects Stall
We’ve worked with many organizations caught between a convincing demonstration and a system that never reaches production. The reasons are strikingly similar each time.
1. They run agentic delivery like a software project.
When teams apply the assumptions of deterministic software, such as a fixed scope, a linear plan and a demo that stands in for proof, to a non-deterministic system, the plan meets reality and breaks. The method has to match the technology.
2. They leave the strategy implicit.
The decisions that matter most—which process, in what order, how it integrates and who owns it—are strategic. Left unmade, they’re settled by default during the build, and the gap left between a working demonstration and a production system is exactly where the project stalls.
3. They optimize for breadth instead of depth.
Reaching for breadth too early is how momentum is lost. To reach production, carefully choose a small number of well-scoped use cases, not an ambitious sweep of all.
4. They treat model access as capability.
The models are available to everyone. Buying access to them, or a partner who can confirm that something is buildable, isn’t the same as knowing what to build. That knowledge only comes from having carried production systems through to operation before.
Where The Returns Live
Most organizations still run their most complex processes by hand: work passed between people, tracked in spreadsheets and stitched across tools that don’t talk to each other. It functions, but it consumes time and headcount every day it runs, and it’s precisely the kind of work a well-scoped agent can absorb.
The returns come from sequencing, not scale. A client of ours in the print industry had been quoting and handling order enquiries manually, with staff assembling quotes by hand. Rather than automate everything at once, our work was staged: first an agent that could look up orders and report their status, then one that could quote products and options a customer didn’t know existed. Each step earned the next. The sequence is simple:
1. Start where the manual cost is highest and the scope is clearest.
2. Prove the result in production, not in a demonstration.
3. Widen the scope only once the first step has earned it.
What The Evidence Shows
In our work delivering agentic AI systems into production, we’ve found that those that succeeded shared three traits:
1. They started from a clearly defined operational problem.
2. They demanded a system that worked in production.
3. They chose a partner for what it had reliably shipped.
The reward for getting this right can be significant, which is why hesitation carries its own cost. BCG’s “future-built” companies, around 5% of the total, post 3.6 times the three-year shareholder return and 1.6 times the EBIT margin of the “laggards,” with agentic AI as the widening force. McKinsey researchers finds growth leaders are three times more likely to raise AI investment sharply. And the window is still open: Gartner analysts puts deployment at 17% against more than 60% of organizations that intend to adopt within two years.
The technology is shared. The models are a commodity. What decides the outcome is the quality of the harness built. In a market this young, strategy isn’t the soft part of the work. It’s what decides the terms of success.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
