AI agents are being deployed across enterprise operations, but many organizations are discovering that the models behind those agents don't understand their business. Impetus Technologies Inc. has built an operational framework to close what it calls the "context gap" - the space between what an AI model knows and the specific rules, data, and relationships that define how a company actually works.
"We were in this world, and we knew what data is and where it sits," said Deepak Khosla, chief growth officer and head of AI at Impetus. "We figured out the gap is not the large language models. The gap is the context and that's why we want to fill that gap."
The context engineering problem
Impetus treats organizational context as managed infrastructure rather than something that gets handled with better prompts. The company has developed a methodology called CEDL - Context Engineering Delivery Lifecycle - that structures how agents receive, learn from, and act on company knowledge.
"The old methodology of building IT solutions, AI solutions and agentic solutions is no longer going to work over here," Khosla said. "You create the context for 'C,' you engineer the context, that's for 'E,' then you have this whole learning model, which is basically you would learn what's happening, what's not. Then you bring back those signals and make those agents work again."
The approach involves three layers: modernizing legacy data systems, organizing semantic meaning through knowledge graphs and ontology layers, and orchestrating agents for deployment. Memory management is also a core piece, because agents need short-term and long-term memory to function reliably.
"Memory is very important because you need to have short-term memory, long-term memory," Khosla said. "Memory is even more important because of the way the LLM works. If they learn the wrong thing by mistake … that's going to stick with you, and that's going to stay with all the actions that's going to happen."
From data platforms to production agents
Impetus wraps this work into its Leap AI suite of products, which moves agents from pilots into production. The company starts with data platform modernization and builds the knowledge and semantic layers as part of that process. The mental model is how you'd onboard a smart new hire.
"It's like a new hire, a new hire in the company, a smart person, and you don't tell them anything about the company, and the smart person is not going to do anything well," Khosla said. "But if you tell the smart hire about your business, your rules and everything, the person will do good. We make those agents smarter by bringing the context there."
For operations leaders, the practical takeaway is: the bottleneck isn't model capability - it's whether agents understand your workflows, your exception handling, and your tolerance for error. An agent needs short-term memory for the task at hand and long-term memory for the organization's rules and conventions. Teams need a feedback mechanism to correct agents when they get it wrong. And the context layer must be built alongside the data platform, not treated as an afterthought. That framing aligns closely with how operations teams are already approaching AI for Operations, where the focus has shifted from getting a model to working reliably within an existing operational structure.
Operations managers who are evaluating AI agent deployments can apply this logic directly to their own planning. Before pushing agents into critical workflows, identify where the context gaps are. Which business rules are documented? Which ones live inside a senior employee's head? How does the agent learn when it is wrong, and how long does that correction persist? Those answers determine whether a pilot stays a pilot or becomes something that actually changes the way work gets done.
Those planning their first enterprise agent deployments may want a structured approach of their own. An operations manager learning path focused on AI covers the practical questions of governance, rollout, and evaluation that come up long before the first agent runs.
Why this matters for operations
Operations teams carry the risk when agents fail. A pilot that handles data incorrectly can corrupt downstream processes, cost money, or introduce regulatory exposure.
The Impetus framework offers a way to think about agent readiness: context, feedback, and memory. If an agent doesn't have organization-specific context, know how it memorizes what worked, and have a mechanism for correction, it isn't ready for production. For operations professionals planning AI projects, the practical question to ask vendors and internal stakeholders isn't "what can the model do?" It's "what does the agent know about us?"
Your membership also unlocks:
