The Missing Half of Enterprise AI: Why Language Models Need Business Context

Large language models have changed what people expect from enterprise AI.

They can understand natural language, generate content, identify patterns, and support increasingly capable agents. But inside a large organization, those capabilities only solve part of the problem.

For Dr. Ashwin Rao, Executive Vice President, Next-Gen AI & Technology at o9 Solutions, the gap is plain to see.

“Language is not complete cognition,” he said at aim10x Europe. “LLMs are powerful, but we need AI that goes well beyond the AI of languages.”

Enterprise decisions depend on rules, constraints, mathematics, domain knowledge, governance, and an understanding of how the business actually works.

That is the thinking behind neurosymbolic AI: combining the strengths of neural AI with the structured intelligence of symbolic AI.

Why an LLM is only part of an enterprise agent

Rao defines an agent simply: something you can tell what to do without specifying every step required to complete the task.

That sounds straightforward. In an enterprise, it is anything but.

An agent needs to understand the context of the business, interact with tools and systems, follow policies, use decision models, and check whether its output is correct.

According to Rao, an LLM is only one component.

“You need to tell the LLM what the enterprise context is, what are the details of your enterprise,” he said.

The surrounding capabilities include context engineering, access to tools, control loops, and evaluation.

Those tools may include databases, governance policies, forecasting models, optimization engines, and other capabilities companies have developed over many years.

Evaluation is particularly important. Enterprise AI needs guardrails and mechanisms for checking whether an answer respects the realities of the business.

Without that foundation, an agent can sound convincing without being reliable enough to support important decisions.

Combining two forms of intelligence

Rao describes neural and symbolic AI as highly complementary.

Neural AI, including large language models, is strong at pattern recognition. It can process large quantities of information, identify relationships, and adapt quickly.

Rao compares it with fast, intuitive thinking.

Symbolic AI is more structured.

It represents business rules, constraints, mathematical logic, and domain-specific decision models. It can capture how inventory, pricing, marketing, supply, and other decisions are supposed to work.

“The best part of it is that it’s traceable, it’s auditable, and it can explain things to you in very clear business terms as to why it made the decision,” Rao said.

That exactness matters in enterprise environments.

Language models operate through approximation. That helps them scale, but many business decisions require precision.

A planning system may need to respect capacity constraints, regulatory rules, decision rights, or contractual conditions. An approximate answer may not be sufficient.

Neurosymbolic AI combines these capabilities.

Giving AI a model of the business

At the foundation of o9’s symbolic AI approach is the Enterprise Knowledge Graph.

Rao describes its core layer as a digital representation of the value chain.

Customers, products, locations, forecasts, plans, and business events can be represented within the same connected structure.

“This is like the digital twin of your business,” he said.

A second layer captures decision context.

That means recording more than the decision itself. The system can capture who made it, when it was made, which alternatives were considered, and which facts and assumptions informed it.

Those decisions can then be connected with the outcomes that followed.

Over time, the organization can begin learning which decisions worked well under particular circumstances.

Turning decisions into institutional knowledge

Rao sees this as one of the most important capabilities of enterprise AI.

Organizations make thousands of decisions, but much of the reasoning behind them disappears after the fact.

By recording decisions and connecting them to outcomes, companies can begin building a clearer picture of what works.

A learning layer can identify patterns in successful and unsuccessful decisions and translate those lessons into rules and policies.

“That is not a one-and-done thing,” Rao said. “You’re constantly going to learn what decisions are working well because, as circumstances change, what decision may have worked in the past may not work now.”

The result is a body of institutional knowledge that can evolve with the business.

This gives future agents more than access to historical data. It gives them context about how decisions have been made and what those decisions produced.

Connecting decisions that are usually separated

The final layer Rao described connects decision models across functions.

Inventory decisions influence pricing. Pricing can affect demand. Marketing decisions can alter both.

Yet in many companies, those models remain separate.

“I don’t think most enterprises have them connected,” Rao said.

A connected compute graph can represent those dependencies.

This creates the foundation for more integrated decision-making, where an action can be assessed for its broader impact rather than evaluated within one function.

The same principle applies across strategic, tactical, and operational decisions.

Bringing unstructured information into the picture

Traditional enterprise systems are built largely around structured data.

ERP records, master data, planning systems, and transactional tables provide an important foundation, but a significant amount of business knowledge exists elsewhere.

Emails, Slack conversations, documents, images, public information, and the knowledge held by employees can contain signals that never reach traditional planning models.

Rao believes large language models create new possibilities for incorporating this information.

“With our neurosymbolic agents, we can extract that signal out of the noise, structure it, and put it in your knowledge graph,” he said.

External information can also be combined with private enterprise data.

At the same time, real-time data can complement the more stable reference information companies already maintain.

The result is what Rao describes as a “living, breathing, evolving” Enterprise Knowledge Graph.

Moving from code to business intent

The combination of enterprise knowledge and generative AI could also change how decision applications are created.

Rao pointed to the growing ability of coding agents to generate software from descriptions of the problem that needs to be solved.

In an enterprise setting, those descriptions can be grounded in the Enterprise Knowledge Graph rather than expressed only in general language.

A user could describe a need for forecasting, causal analysis, scenario simulation, optimization, or market-share analysis using the concepts and relationships already represented in the business model.

The underlying code, logic, and decision models could then be generated.

“This opens up so many possibilities,” Rao said.

This could broaden enterprise planning beyond fixed applications for individual processes.

Answering the four questions behind every decision

Rao framed many enterprise problems around four questions:

What happened?

Why did it happen?

What is likely to happen next?

What action should be taken?

The first two help organizations understand the past. The next two are focused on the future.

Neurosymbolic AI is intended to support all four by combining pattern recognition with structured business knowledge and decision models.

That combination becomes especially important as companies move from analytical AI toward agents that can recommend or take action.

Towards faster enterprise AI deployment

Rao sees data integration as one of the areas where the combination of neural and symbolic AI could have an immediate impact.

Work that historically took six to twelve months could, he said, increasingly be completed in weeks.

The longer-term goal is even more ambitious.

“You come in and data is just coalesced together. You’ve got an EKG. You tell it what problem you want solved, what decisions you want to make. It will be done,” Rao said.

That future depends on more than increasingly powerful language models.

For Rao, enterprise AI needs a detailed understanding of the business it is operating within, including its rules, relationships, decisions, and accumulated knowledge.

LLMs provide one form of intelligence.

The missing half is the enterprise itself.

Watch