# AI agents in operations: responding is not operating > AI agents in operations, separated from demo-ware: the difference between an agent that answers and one that operates with permissions and traceability. - Canonical: https://arkatai.com/en/ai-agents-in-operations/ - Site: Arkatai (https://arkatai.com) — agentic operations as a service - Language: en - Published: 2026-07-18 --- Executives have seen agents complete a task in a demonstration and pilots that never entered a production process. The difference is not explained only by model quality. It depends on whether the agent produces an answer for a person to review or has authority to act in company systems. As a Chief Product & Technology Officer, I use that distinction when designing agent operations. ## Responding and operating require different controls An agent that responds summarizes a contract, drafts a reply or retrieves a refund policy. A person remains between the output and an external consequence. This category is widely available through models and products that competitors can also buy. An agent that operates issues the credit note, updates the order, blocks the shipment or sends the email. A person does not approve every step, so the system needs permissions, limits and traceability defined against the process it executes. **Permissions:** exactly which systems the agent can touch and which verbs it can use in each. Read the order history, yes; change a price, no. Not as policy on a slide, as enforced access. **Limits:** the quantitative boundaries of autonomous action. Approve refunds up to a defined amount, handle up to a defined percentage of daily volume, act only on cases matching defined criteria, and escalate everything else to a named human with the case file and its reasoning attached. **Traceability:** every action logged with what the agent saw, what it decided, and why, in a form a supervisor, an auditor, or a regulator can review after the fact. When something goes wrong, and something eventually will, the question "what happened and why" must have a fast, complete answer. Ask what happens when the agent is wrong. A system intended for production should have an escalation path, a maximum scope for one error and a recovery procedure. If those controls are undefined, the deployment is not ready to operate. ## What a demo leaves out A demo shows the standard flow of a task on prepared inputs: process an invoice or answer a ticket. Production also includes accumulated exceptions: the customer who pays by monthly transfer against a spreadsheet, the product code with different meanings in two warehouses or the approval that follows a different path in practice. Staff may apply those rules without having documented them. The standard flow is becoming product. The exception layer requires work specific to the company. [The custom phase](/en/the-custom-phase/) separates the components that are utility, product and custom. A licence leaves that deployment work with the client. A managed operator can perform it inside a shared system and remain responsible for maintaining the agents. ## What a production deployment requires A production deployment follows a sequence that starts before technology selection. It starts with one process and one accountable executive. The team documents how the process runs, including exceptions and knowledge held by its operators, in a form precise enough to execute. The resulting rules, tolerances and escalation paths form [the operating model as code](/en/operating-model-as-code/). The agent can begin in shadow mode, proposing actions for human approval. This produces an error record without granting write authority. Autonomy is expanded by class of case: a refund limit rises after an agreed period meets its error target, and case criteria widen after escalations show how the system handles exceptions. I cover the permission and review structure in [AI agent governance](/en/ai-agent-governance/). The model responds in seconds, but deployment can take months. The time goes into documenting the process, agreeing limits and connecting systems that were not designed for software operators. A plan based on response latency ignores that work. I lay out the preparation sequence in [how to prepare your company for agents](/en/prepare-your-company-for-agents/). This does not require the client to create an AI department. The client needs a process owner who can decide rules and priorities. In a managed model, a forward-deployed engineer connects and adapts the shared system while the provider owns integration, reliability and the technical lifecycle. ## Where agents operate today Agents that answer over company documents and data are broadly available. Agents that operate are being used in high-volume processes with describable rules and an error budget under supervision: order handling within tolerances, first-tier customer operations, reconciliations and internal service requests. The recurring conditions are bounded scope, explicit limits, expansion based on records and a human owner. An agent cannot take over an undocumented process end to end without people defining the process, connections and controls. A demo that implies otherwise omits the implementation work. Companies must either plan that work or discover it through failures during deployment. ## Questions boards ask me ### Which process should we start with? Pick one with volume, describable rules, a tolerable cost per error and an owner who wants it. Ownership can matter more than the potential value because the first deployment requires frequent process decisions. It also builds controls and experience that can be reused. ### Can we skip the shadow-mode phase to move faster? You can, but the first visible error will occur without a prior record of how the system performs. Shadow mode supplies the evidence needed to grant write permissions by class of case, and high volume can produce that evidence quickly. ### How do we evaluate vendor claims in this space? Ask four questions: what happens when the agent is wrong, who encodes processes and exceptions, whether the client can inspect the full trace for any decision, and who keeps the agents current after deployment. The answers distinguish a software licence from an operated service. ### Is this an IT project or a business project? The business owner defines how the process runs and what the agent may do. Integration, reliability and security can belong to an internal technology team or to the managed-service provider. A company without that team can still deploy agents, but it cannot outsource the business decisions.