Against the backdrop of today’s generalized freakout about the potential existential threat posed by artificial intelligence (AI), business software providers are rapidly building out AI agents that can safely and reliably improve human productivity and compress operational and decision-making cycles to boost responsiveness and resiliency as well as cut costs. In addition to embedding already-built agents, they are offering agent studios that extend agentic capabilities that can be used to capture an enterprise’s unique needs or used by implementation partners to monetize their industry-specific intellectual property.
ISG Research asserts that by 2028, almost all business software providers will have augmented the capabilities of their applications with some agentic AI capabilities to lower costs, upskill users and improve customer service and agility. Enterprises, implementation partners and service providers must recognize where application-embedded agents are heading before making major investments that can become obsolete within the next one to two years.
Agents Managing Deterministic, Quasi-Deterministic and Probabilistic Processes
Currently, the application-embedded agents in general availability, and those that are on roadmaps, are almost exclusively focused on deterministic and quasi-deterministic processes. Deterministic processes apply defined rules to identical inputs, producing repeatable, auditable and controllable results that diminish the risk of errors and hallucinations. They offer consistency, auditability and ease of control. At the same time, they are limited when inputs, business conditions or exceptions do not fit a defined rule set. Quasi-deterministic processes combine rules with bounded judgment and exception handling while tolerating variable data quality. While balancing control with flexibility, they require clear escalation paths and human-in-the-loop (HITL) interventions to prevent inconsistent decisions. ISG’s 2026 AI Value Study finds that a majority of enterprises rely on HITL rather than autonomous agents for AI-enabled processes. For task automation and workflow execution, 62% have HITL.
The third broad type of processes, probabilistic ones, use models to infer, predict or generate likely outcomes. They can be effective for executing less defined and more open-ended processes, ones where context, patterns and unstructured information matter. However, their outputs are not guaranteed to be repeatable or correct.
The Value of Hybrid Agents
Hybrid agents that manage deterministic and quasi-deterministic tasks are typically the most effective in business applications. Deterministic systems execute the formal calculation, control or validation, while AI and machine learning (ML) support preparation, interpretation, monitoring and exception management. This hybrid approach allows enterprises to increase throughput and improve user experience while retaining the governance needed for risk management of critical business processes.
Generative AI (GenAI) and ML are emerging as important enablers of deterministic business processes, particularly where organizations seek productivity gains without compromising control, auditability or accountability. These features are designed to improve the quality and speed of the work performed while following established rules, calculations and approval logic. AI can translate business requirements into workflows and rules, extract and normalize information from unstructured documents, classify incoming data and identify anomalies or exceptions requiring attention. Machine learning can improve matching, prioritization and exception handling by recognizing patterns in historical outcomes and user corrections. GenAI can make processes more safely actionable by explaining exceptions in plain language, summarizing supporting evidence and, wherever necessary, creating explanatory or guidance text for review. In finance, for example, AI can help reconcile incomplete records, flag unusual journal entries and explain a failed tie-out, while approved accounting logic remains the authoritative basis for the result.
Built-in Automated Testing is a Requirement
Software application providers must offer built-in tests for customers so they can periodically assess the performance, quality, cost and safety of their agents. Agents in deterministic and quasi-deterministic end-to-end workflows must be regularly evaluated for their output quality, reliability, control effectiveness and auditability. For deterministic use cases, testing should focus on exact reproducibility through golden data sets, rule and regression testing, boundary conditions and negative cases as well as the validation of data lineage and access controls. Any change to rules, mappings, thresholds or workflow logic should trigger retesting against approved expected outcomes. Quasi-deterministic agents require these same disciplines, supplemented by scenario testing for ambiguous, incomplete and conflicting inputs. Testing must confirm that each agent operates within defined decision boundaries, handles routine cases consistently and escalates uncertain situations rather than making unsupported judgments.
The boundary conditions used in agents are especially vulnerable to obsolescence since business conditions evolve and thresholds for HITL interventions depend on context. So, where numerical limits and ranges are used, they must be periodically reassessed to ensure they optimize tradeoffs between effectiveness and workloads. Software providers and enterprises can use AI to test and refine non-absolute boundary conditions, converting subjective concepts (such as completeness, acceptable similarity or materiality) into measurable and reviewable criteria.
For example, rather than treating a document as simply complete or incomplete, AI can assess whether required information is present, usable, current, consistent and supported by acceptable evidence. Similarly, instead of requiring an exact record match, AI can evaluate multiple attributes, such as legal name, tax identification, address, bank information, historical aliases and transaction context, and then provide a confidence assessment with supporting evidence. In turn, tests should be designed to flip these capabilities to generate ambiguous and adversarial cases, compare results with expert decisions and analyze disagreements to identify unclear policies, missing data or poorly calibrated thresholds. Tests should be able to assess agents in shadow mode, where they run in tandem with production states without affecting business outcomes.
Insights derived from testing can be used to approve changes to decision rules and escalation criteria. However, AI should not autonomously redefine a material boundary condition. In critical business processes, the organization must retain ownership of policy, thresholds and decision rights, while AI provides evidence, pattern recognition and recommendations that improve the quality and consistency of those controls.
Shadow-mode deployment is a particularly effective approach to ongoing testing because organizations can compare agent outputs, recommendations and exceptions with existing processes and expert decisions before production use. Ongoing monitoring is essential to identify changes in data quality, policy, process conditions or agent behavior. Testing must confirm that routine cases are processed correctly, uncertain cases are escalated appropriately and each outcome can be reconstructed from its inputs, rules, evidence and approvals.
Cost and Operational Advantages of Embedded Agents
Embedded agents in business applications have inherent advantages over generalized agents. Embedded agents natively access the application’s data, workflow state, metadata and permissions, minimizing compute costs. Many business application platforms include a data fabric that includes information from all necessary sources of record, which I have (tongue in cheek) called a data pantry. For example, the data component of the platform would include a persistent store of relevant data from all of the enterprise’s ERP systems needed for a supply chain planning application with a common semantic layer designed for the users and use cases using that application. This limits compute costs and reduces application programming interface (API) calls necessary for creating and maintaining agents. Agents embedded in an application also do not need to convert operational information into prompt text.
A generalized agent needs to repeatedly retrieve data and documents that are external to it, including invoices, purchase orders, policies, approval matrices and correspondence. Embedded agents require fewer tokens, API calls and almost never need token-eating disambiguation routines. Embedded agents can be designed to create compact, context-driven, structured outputs (such as exception codes, a confidence score, evidence links and the next best action) rather than producing narrative responses. They have persistent access to prior decisions, approvals and lineage, which repeatedly reduces the need to reconstruct context each time the agent runs.
That noted, embedded agents do not guarantee savings if they are poorly constructed. They should be built applying deterministic application logic and structured queries for standard processing, reserving model inference for ambiguity, anomaly detection, interpretation and exception management.
Agent Studios Will Be the Norm
Agent studios are designed to let organizations create, extend, validate, deploy and manage agents within an application platform. This is somewhat like an extension of the software development kit, which has enabled enterprises to extend and add customizations to multi-tenant software-as-a-service (SaaS) offerings. The availability of no-code, low-code and pro-code options broadens the potential user base from functional administrators to developers. Most software providers will adopt a marketplace strategy to support partner-built agents, enabling domain experts to monetize their unique intellectual property. Enterprises will want business applications that offer a studio to rapidly build a broader agent ecosystem, addressing the nuances of even narrow niches of an industry with distinct business processes and terminology.
A Hybrid Approach, Not an Either/Or
As the use of agents spreads and becomes a routine element of business computing, enterprises will adopt a hybrid approach to designing, deploying and maintaining agentic systems. They will aim to design agentic systems that are the most cost-effective within the constraints of governance, risk and maintainability. Where an end-to-end process touches one or more agent-embedded application, the generalized agent will act as an orchestrator, leaving the work that can be performed within the application to the native agents. I strongly recommend that enterprises, implementation partners and service providers recognize where the market is going before making major investments that will be obsolete within the next one to two years.
Regards,
Robert Kugel