The AI System in Your Business Cannot Learn. You Need to Engineer Around That.

Every AI deployment decision your organisation makes should start from one uncomfortable fact: the AI system you are investing in cannot learn from its work inside your business. It is trained on data compiled before it was deployed. It is then fixed. It does not adjust its understanding as your context changes, as your market shifts, or as new information accumulates. A human employee hired today will be a different and more capable employee in two years. The AI system you deploy today will not be, unless someone re-engineers it. That distinction changes almost everything about how these systems should be managed.

Babak Hodjat Cognizant’s Chief AI Officer

I discussed this directly with Babak Hodjat in the most recent episode of The AI Adoption Podcast. Hodjat is Chief AI Officer at Cognizant, a professional services and technology company with more than 300,000 employees worldwide, and he has been working in AI since the late 1980s. His perspective on the state of the technology is unusually grounded. He is neither dismissive of AI's potential nor willing to leave its limitations unexamined.

The Limitation Nobody Is Talking About Loudly Enough

The AI systems at the centre of today's enterprise investment are large language models: systems trained on vast quantities of text and code that can understand and generate natural language, reason across complex problems, and operate across multiple languages simultaneously. These are, as Hodjat acknowledges, genuinely significant capabilities. The ability to have a machine abstract the world in a way that resembles human abstraction is something researchers in the field spent decades believing might never arrive.

But these systems do not learn from lived experience. Hodjat's analogy is the film Memento, in which the protagonist loses his short-term memory and must re-read notes tattooed on his own body to remember who he is and what he is doing. A large language model is structured similarly. It can hold a conversation, reason within a session, and appear to learn. Clear its context and it starts again from nothing. It has no memory of the last time you used it. It cannot update its understanding based on what it learned in your deployment.

Chelsea Chamberlain https://open.spotify.com/episode/25drYaPW8HzhAUdeKcTKhE?si=J1jLgEOYRQ-o-9pOOvt-Sg

Hodjat is clear that this is not primarily a short-term technical gap. Some aspects of it may be addressable through the modularisation that agentic systems enable. But it is a structural characteristic of how these systems are built, and organisations that deploy AI without accounting for it will find themselves managing an asset that requires constant human intervention to remain relevant to its context.

Agentic AI Is Real, Deployed, and More Demanding Than You Think

The more immediate practical development Hodjat describes is the emergence of agentic AI: systems in which a large language model is combined with code that allows it to take actions in the world, not just generate text. At Cognizant, this is not a pilot. The company has deployed a multi-agent intranet system across its entire global workforce. Employees can express an intent, in natural language, and a network of agents working across HR, finance, legal, IT, and other functions will collaborate to address it. The system has handled more than 11 million such commands. When an employee's child turns 26 and loses insurance eligibility under US rules, the system does not ask the employee to log into three different applications. It works across them simultaneously.

The productivity and support cost implications are real and measurable. But Hodjat is also direct about the demands these systems place on the organisations that run them. Agentic systems are not, in his assessment, a solution to the fundamental limitations of large language models. They require constant human re-engineering. As context changes, as new agents are added, as the boundaries of what the system is asked to do expand, the humans responsible for these systems must intervene, adjust, and test. The expectation that AI can be deployed and left to improve on its own is not consistent with how these systems work.

Chris Blatchford https://open.spotify.com/episode/7tMT1OgDGrvT3QfTXC9GNA?si=5bhGIroMQ0O3MdgKsrupWA

The Question Leaders Need to Ask Before the Next Deployment Decision

There is a practical implication in Hodjat's analysis that bears directly on how senior leaders should approach AI investment. The reliability of AI reasoning chains is not a given. A single large language model begins to fail catastrophically at around 300 to 400 reasoning steps. For an enterprise system making thousands of decisions, that failure rate is not acceptable. Cognizant's research addressed this by having agents check and vote on each other's reasoning steps, achieving one million error-free steps in a published paper. The engineering solution exists. But it requires deliberate design and ongoing maintenance, not an assumption of reliability.

The same discipline applies to the concentration of AI capability. Open source models are currently six to nine months behind the most capable commercial systems, and the gap is closing. For organisations with concerns about dependence on a small number of commercial providers, this is a relevant consideration in supplier and technology strategy.

The speed of change is Hodjat's third and perhaps most personally urgent point. Previous technological disruptions took decades. The automobile displaced the horse and cart over a generation. AI is disrupting knowledge work in months. The workforce planning, retraining, and organisational design questions that previous generations of leaders had years to work through are arriving simultaneously. Leaders who are still treating AI adoption as a future concern rather than a present operational reality are already behind.

The question for leaders is not whether AI is transformative. It is whether the organisations they lead are engineered for its actual characteristics, rather than the simplified version described in the popular press.

Listen to the full conversation on Spotify: open.spotify.com/show/296zibtjU3w4ANuUsPSu2D

Listen on Apple Podcasts: podcasts.apple.com/gb/podcast/the-ai-adoption-podcast/id1811897501