A 2025 MIT study noted that 95% of enterprise AI pilots fail to move beyond an initial demonstration. The reasons are familiar: messy data, resistant organisations, or technology oversold ahead of its capability.
However, Ashwin Rao, Chief Technology Officer at o9 Solutions, a supply chain planning company, argues that these explanations miss a more basic problem with architecture itself.
"I don't think those are separate problems. In my experience, the hardest problem in enterprise AI is organisational change management. Data quality matters as well. But both of those issues are closely connected to architecture,” Rao tells AIM.
Most enterprise AI products built since 2022 lean heavily on large language models, using them to interpret queries, generate analyses, and produce recommendations.
"A standard LLM-based agent can often generate a plausible answer to each question individually," he says. "The challenge is that it does not inherently understand how those decisions connect across the business."
He points to a recurring failure pattern: "A system might recommend a replenishment quantity that violates a contractual minimum order, or identify a relationship in the data and mistake correlation for causation."
This is the gap o9 says its hybrid model is designed to close. The company calls the approach neurosymbolic AI, combining neural AI, which learns patterns from data, with symbolic AI, which operates from defined rules and structured relationships. Neural AI resembles learning through repetition and experience. Symbolic AI resembles coaching and technique.
The company has built this into a framework it describes as ‘10-80-10’.
The first 10% of the process involves the language model interpreting intent: which products, locations, and planning periods are in question. The middle 80% is enterprise reasoning, where the system traces plans, forecasts, inventory positions, and decisions across the business to identify the actual cause. The final 10% returns to the language model, which communicates the findings in business terms.
"In many cases, the inventory issue that appears in a report is not the actual cause," Rao says. "Excess inventory may be linked to a forecasting assumption, a production policy, a procurement decision, or a combination of factors."
The architecture predates the current generative AI boom by years. o9 had been developing the Enterprise Knowledge Graph and its underlying planning models for 15 years before large language models became widely available.
"What changed in 2022 was not the need for the Enterprise Knowledge Graph," he notes. "What changed was who could use it. The capability existed already, but required specialist users. Language models opened access to it and also made it possible to work with unstructured sources such as emails and supplier communications at scale”
This puts o9 at odds with prevailing market sentiment, where boards and executives have made substantial commitments to generative AI and, in some cases, expect to see more of it rather than less. Rao says this produces friction in early conversations, though it tends to resolve quickly.
"What we typically find, however, is that the conversation quickly shifts from language capability to decision-making," he explains. "The question becomes less about how much of the system is powered by an LLM and more about whether the system can be trusted when decisions have real operational and financial consequences."
That question of trust extends to how enterprises handle disagreement between functions. Finance, supply chain, and commercial teams frequently work from different assumptions and different numbers. Rao does not present the Enterprise Knowledge Graph as a tool for forcing consensus.
"When teams disagree, the system helps make the disagreement explicit," he says. "Is it a different demand assumption? A supply constraint? A financial target? A policy decision?" The aim, he notes, is a shared context for decision-making rather than agreement on outcomes.
On sudden disruptions, such as port strikes or geopolitical shocks, Rao is measured about what the system can offer. "No system can reliably predict every black swan event," he adds. "What matters is how quickly the organisation can understand what changed and decide what to do next."
In these situations, neural AI interprets incoming signals while the symbolic layer connects them to inventory, capacity, and financial outcomes that already exist within the business.
Asked whether o9's investment in India reflects faster adoption of complex AI architectures among Indian enterprises compared to Western markets, Rao declines to frame it in geographic terms. "What tends to separate successful enterprise AI initiatives from unsuccessful ones is not geography. It is whether the organisation has connected its data, decisions, and processes in a way that allows AI to operate within the realities of the business," he shares.
The broader argument is that the industry's pilot failures are not primarily a language problem. They are a structural one, rooted in whether systems can connect to the actual mechanics of how a business runs, and whether their conclusions can be traced and trusted when the financial stakes are real.
Your reaction