Enterprise knowledge management is moving deeper into the operational architecture supporting artificial intelligence (AI) as it becomes clear that, beyond access to corporate documents and databases, agents require enough context to determine which information applies to a particular user, system, location and business process.
This requirement is changing how companies approach knowledge management. Systems traditionally designed to document procedures and help employees resolve problems are increasingly being used to supply AI applications with curated knowledge, operational telemetry and metadata that can improve the accuracy and consistency of automated decisions.
To this point, Gartner reported in April that context, including semantics and metadata, is becoming critical infrastructure for enterprise AI because autonomous agents require governed access to trusted information. The firm found that organizations with the most mature AI-ready data and analytics capabilities were achieving up to 65% greater business outcomes, including revenue growth and cost optimization. ISG has similarly found that enterprises scaling AI are increasingly coordinating AI and data strategies as applications demand continuous access to accurate, contextually relevant operational information.
The shift also has economic implications. Large AI models can process increasingly expansive amounts of information, but sending more enterprise data through those models can increase token consumption without guaranteeing more consistent answers. Companies are consequently examining ways to narrow the information supplied to AI systems before a query reaches a model.
During a recent BizTechReports executive vidcast interview in the wake of the 2026 CIO 100 Awards & Conference, Joel Raper, chief commercial officer at Unisys, suggested that this evolution is forcing companies to reconsider the relationships across knowledge management, data preparation and AI deployment.
Raper's perspective draws partly from his previous responsibility for digital workplace services, where hundreds of employees were involved in creating knowledge used to support IT operations. That model depended heavily on people documenting how problems were resolved so other employees could reuse the information. Agentic systems create an opportunity to make that process considerably more dynamic.
Knowledge Management Moves Into the AI Architecture
Traditional IT knowledge management often revolved around articles describing how particular problems should be diagnosed and resolved. The effectiveness of those repositories depended on the quality of the documentation and whether employees actually used it.
It is a limitation that has taken on greater significance as AI systems become dependent on enterprise knowledge. ISG, for instance, has reported that making data usable for AI was the most frequently cited data challenge companies expected to face through the end of 2026, while nearly one-third identified data quality, accuracy and consistency among their biggest AI challenges. Meanwhile, Gartner analysts have elevated “context” itself to an architectural priority, arguing that AI agents require trusted data, metadata and semantic information to operate effectively.
Raper said organizations can now capture information generated while work is being performed, including voice conversations, instant messages and the steps employees take while resolving problems. AI can use those interactions to help create and update knowledge and establish context without depending entirely on employees to document processes afterward.
The next stage involves connecting that knowledge with operational telemetry. An AI system troubleshooting an application, for example, may need to know the operating system being used, the version of the application, recent patches and when a device was last restarted. Those data points can help distinguish among multiple knowledge articles that might otherwise appear relevant.
The same principle grows more important as agentic systems move beyond IT. For example, an employee asking about parental leave may require an answer based on country, jurisdiction, employment status and company policy. A finance employee asking about an internal process may require access to entirely different information than another employee entering a similar prompt.
Knowledge management therefore begins to function as a contextual layer between enterprise information and AI systems. The objective is to determine which sources should carry the most weight before the system generates an answer or initiates an action.
Achieving this objective is particularly important as AI agents move from providing information to performing work. Raper described an evolution where organizations can monitor how reliably particular knowledge resolves a problem. Once confidence reaches an acceptable level, the same knowledge used to advise an employee could potentially support automated remediation. The progression, he stated, creates a path from knowledge capture to decision support and eventually automation.
AI Economics Put Greater Pressure on Context
The cost of operating AI systems is adding urgency to the issue. Enterprise AI strategies frequently emphasize model capabilities and access to large amounts of corporate information.
The economics surrounding those choices are receiving greater scrutiny. Gartner reported in August that growing token consumption and consumption-based licensing could push AI coding costs above the average developer's salary by 2028. Recent BizTechReports coverage of enterprise AI deployments has also highlighted the widening range of foundation, specialized and open-source models available to enterprises, a trend that is pressuring companies to align model capabilities and costs more closely with individual workloads.
Raper argues that the amount of information placed into a model can turn into an economic liability when organizations repeatedly consume large numbers of tokens to answer relatively narrow business questions.
A more targeted approach preprocesses requests before they reach the underlying AI model. Instead of sending a broad collection of corporate information with every prompt and asking the model to determine what matters, the surrounding architecture first evaluates the request and identifies the information most likely to be relevant. It can consider factors such as the identity and role of the user, geography, the business process involved and operational data associated with the request.
For a human resources question, for example, geography, employment status and applicable corporate policies might determine which information is retrieved. An IT support request instead could trigger retrieval of the device’s configuration, operating system, application versions, recent patches and previous incidents. The model receives a smaller and more relevant body of information from which to generate its response.
That distinction has both operational and financial consequences. Reducing irrelevant context can lower token consumption while limiting the number of competing information sources the model must interpret. It can also allow organizations to direct relatively predictable requests toward smaller or less expensive models while reserving more capable models for problems requiring broader reasoning or additional information.
Raper cited an enterprise environment containing roughly 50,000 knowledge articles and associated telemetry where the system's token costs were lower than his own AI consumption over the same period. The example illustrates a broader economic issue emerging around enterprise AI. Model efficiency increasingly depends on the architecture surrounding the model as much as the capabilities of the model itself.
Companies can potentially use smaller models for many routine requests when the relevant context has already been identified. Larger models can then be reserved for circumstances where the available enterprise knowledge is insufficient.
This approach also addresses another persistent problem with generative AI. Giving a model access to more information does not necessarily ensure that it will provide the same answer consistently. Controlling sources and establishing contextual priorities can reduce the range of information the model must interpret. That could become increasingly significant as AI moves into business processes where inconsistent answers carry operational or regulatory consequences.
AI Deployment Creates a New Path to Better Data
The changing role of knowledge management also challenges the assumption that companies must substantially clean and restructure their data before pursuing enterprise AI.
Raper said companies can deploy AI against existing information and use the results to identify where their most consequential knowledge gaps reside. Failed or incomplete responses can serve as signals showing where additional documentation, cleaner data or connections to other systems are required.
He pointed to a customer deployment spanning IT, human resources and finance that resolved approximately 35% of incoming requests during its first 30 days. The unresolved requests provided information that could then be used to prioritize improvements.
Under that model, AI deployment and data improvement can proceed as parallel activities. Instead of attempting to clean every repository before implementation, companies can identify which deficiencies materially affect AI performance and concentrate resources on those. Knowledge environments can then improve iteratively as new questions, processes and business requirements emerge.
That said, Raper stressed that this does not eliminate governance requirements. AI systems combining employee information, operational telemetry and corporate data must still enforce access controls and comply with privacy and data sovereignty requirements. Information appropriate for processing inside a company's infrastructure may be inappropriate for submission to an external model.
Those considerations grow more complex when AI agents can act on information rather than simply retrieve it. Governance must cover the knowledge an agent can access, the contextual information it is permitted to take and the actions it is authorized to perform.
Indeed, Gartner has argued that agentic execution requires identity, permissions and audit controls to be enforced through the enterprise control plane. Separately, recent BizTechReports coverage of autonomous AI governance has highlighted the need for organizations to understand which agents are operating, what permissions they hold and which tasks they are authorized to perform.
The broader development is turning knowledge management into a more active component of enterprise AI architecture. Corporate knowledge repositories historically helped preserve institutional expertise and accelerate human problem-solving. Agentic systems extend that role by using the same knowledge to interpret requests, select information, make decisions and potentially initiate actions.
That raises the strategic value of the context surrounding enterprise data. Companies pursuing agentic AI will increasingly have to determine which information their systems should trust, how that information should be weighted and how quickly the knowledge environment can adapt when circumstances change.
In that environment, concluded Raper, the quality of an AI system may depend increasingly on how effectively an organization structures and continuously improves the knowledge management surrounding it.
###