An enterprise AI agent uses a language model to carry out tasks in business software. Given an emailed order, for example, it might look up the customer, check the product codes and create a draft in the order management system. The model chooses which tools to use as it works through the request. The company sets the permissions and approval requirements.
Anthropic’s engineering guidance on agents calls workflows “systems where LLMs and tools are orchestrated through predefined code paths,” and agents systems “where LLMs dynamically direct their own processes and tool usage.” An agent can therefore handle requests that require different sequences of steps, provided it has the tools and instructions to handle those differences.
How an enterprise agent works
An agent reads the result of each action before choosing the next one. A failed customer lookup, for example, may require another search or a request for help. It should not proceed as though the lookup succeeded.
- TriggerAn email, a ticket, a form, a schedule or a record change
- Gather contextRead the records, documents and policies the agent is allowed to see
- Choose an actionThe model picks a tool and fills in its inputs
- ActCall an API, a connector, a robot or a screen-level action
- Check and repeatRead the result, then continue, retry or escalate
Around the loop: an identity with scoped permissions, approval gates on sensitive actions, a log of every step, and a limit on how long the loop may run.
Instructions. The agent receives a description of its job in natural language. In Salesforce Agentforce, for example, builders define subagents, write instructions for each, and attach a library of actions the agent can choose from.
Tools. Tools are the actions the agent may take. ServiceNow agents use flow actions, subflows, scripts and skills as tools. In UiPath, an agent’s tools can include existing automations, API workflows and MCP servers, plus escalation paths such as Action Center for human decisions. Tool descriptions tell the model when to use each tool and which inputs it needs.
Deterministic steps. Several platforms let builders mix model reasoning with fixed logic. Salesforce’s Agent Script combines deterministic workflow logic with the model’s reasoning, and Skan says its agents run repeatable work deterministically and apply model reasoning only where judgment is needed. For example, a model can interpret a supplier email while code checks the approval threshold for the resulting purchase order.
Approvals. Approval behavior depends on the product and its configuration. In Microsoft’s Agent Framework, for example, “by default, all tools provided to an agent are invoked without user approval”, and developers add a tool approval mechanism for high-risk operations. Check those defaults before connecting an agent to systems where it can make changes.
Published examples in production
C.H. Robinson and Octopus Energy have published results from agents handling orders and customer email. Both report their own figures.
Freight order intake at C.H. Robinson. The logistics company reports more than 30 agents across quoting, order processing, freight classification, appointments and tracking. Its Orders Agent reads emailed tenders and their attachments and builds orders in about 90 seconds, at 5,500 truckload orders a day. The Appointments Agent sets more than 3,000 appointments daily.
Customer email at Octopus Energy. In a three-month trial reported in July 2026, an assistant called Arlo answered straightforward emails about tariff renewals, payment dates and account details: about 8,000 emails a week, or 4% of the supplier’s UK customer email. The company reports 76% customer satisfaction against 72% for comparable human replies. Octopus routes vulnerable customers, sensitive cases and complex complaints directly to experienced staff rather than Arlo.
Agents compared with chatbots, workflows and RPA
| Approach | Who decides the next step | Actions on business systems | Typical example | Where it fails |
|---|---|---|---|---|
| Knowledge assistant | The model chooses the answer; any actions follow fixed triggers | Mostly reads; can invoke predefined actions | Explain the travel policy | Gives a plausible answer that is wrong |
| Workflow automation | Rules written in advance | Reads and changes records along set paths | Route an invoice over a threshold to a manager | Inputs that do not match the rules |
| RPA | A recorded or coded script | Reads and changes records through the screen, step by step | Copy order data from a portal into an ERP | Screen changes and unexpected states |
| AI agent | The model, choosing among permitted tools and their order | Can read or change records within permitted tools; some agents are read-only | Read an emailed tender and create the order | Wrong reasoning, or more authority than it needs |
The distinguishing feature of an agent is that the model chooses which tools to call and in what sequence. Whether it may change records is a separate permission decision.
The approaches are often combined. UiPath’s Maestro runs agents, robots and people inside one BPMN process with decision tables for business rules. UiPath’s own agent documentation also says that “tasks with high accuracy, legal, financial, or regulatory constraints should continue to rely on deterministic automation.” A common design is an agent that interprets a messy request, followed by a fixed workflow that writes the transaction and routes the approval.
The choice also affects cost. In Microsoft Copilot Studio’s standard billing rates, one agent action costs 5 Copilot Credits, while 100 actions in a predefined agent flow cost 13. At those rates, using a predefined flow for fixed steps can substantially reduce credit consumption.
What an agent needs before production
A specific task definition. The agent needs the trigger, the records employees consult, the decisions they make, what counts as done, and the cases that go to a person. Check the written procedure against actual cases. If employees consult an extra spreadsheet or seek approval for an exception, the instructions need to account for that. Process intelligence and task mining can help identify steps missing from the documentation.
Scoped identity and access. Give the agent only the access needed for the task, and make its actions attributable to an agent or an authenticated user. Workato, for example, says its agents act on behalf of verified users rather than shared service accounts. Microsoft’s Agent 365, generally available since May 2026, extends Entra access controls to agents and keeps a registry of agents across the tenant.
Approval gates and escalation routes. Decide which actions need a signature, who signs, and where a rejected or unclear case goes. Payment changes, customer commitments and access grants are the usual candidates.
Logs and a stop control. Every tool call, input and output should be recorded so a reviewer can reconstruct a decision, and someone should be able to stop an agent quickly. ServiceNow’s governance product, for example, includes a kill switch for individual agents.
Evaluation on complete cases. Test the agent on historical cases, including missing documents, conflicting records, tool failures and inputs that try to override its instructions. Check whether it changed the right record, used the correct values and obtained any required approval. Salesforce’s builder, for instance, supports batch tests generated from real interactions and traces showing which actions fired.
A cost model. Agents are billed by consumption on most platforms: Copilot Credits at Microsoft, Flex Credits at Salesforce ($500 per 100,000, with a standard action at 20 credits), and usage tiers elsewhere. Model the volume of cases, the steps per case and the retry rate before committing.
Choosing the first workflow
Order intake, service-request triage and supplier-document checks are reasonable candidates when the team can define which cases the agent may handle. Start with historical cases and compare its decisions with the accepted outcomes. Include missing documents and exceptions as well as routine requests.
Before allowing live actions, decide who will review failures and how to reverse a mistake. Measure completed cases, errors, review time and cost. An agent that finishes quickly but sends every case back for checking may save little time overall.
For the software choice, see our comparison of enterprise AI agent platforms. WIQ’s platform helps teams document the process in an agent blueprint, review it and deploy the agent.
Common questions
Do enterprise AI agents replace RPA?
No. Robots remain the cheaper and more predictable choice for fixed, high-volume screen work. Agents handle the interpretation that robots cannot, and many platforms let an agent call an existing robot as a tool.
Does an agent need full autonomy to be useful?
No. A common design has the agent gather evidence and prepare an action while a person approves the final change. The team can allow more actions once testing and live review show which cases the agent handles reliably.
What is the difference between an AI agent and agentic AI?
Agentic AI is the broader class of systems that plan and act toward goals. An enterprise AI agent is a specific application of that approach, configured to carry out a task in a company’s systems.
Where does an agent’s knowledge of the process come from?
From its instructions, the tools it is given, and any documents it can retrieve. An agent cannot reliably follow an exception that nobody has explained to it. Checking instructions against screen recordings, system logs and employee input helps uncover those omissions.