AI & automation
AI agents in business — 8 use cases and rules for safe implementation
Maciej Zaleski · Published · Updated · 7 min read
An AI agent in a company is a solution that can interpret a task, use the tools made available to it, and carry out successive steps within defined boundaries. Unlike a simple chatbot, its work does not have to end with a text response: it can retrieve data from an API, prepare a change in a system, and submit it for approval. Its usefulness depends on process quality and permission controls, not freedom to act alone.
For a company, the safest starting point is usually a narrow scope: an identifiable goal, a limited set of tools, and a clear escalation path. Not every stage has to be autonomous. An agent that prepares a well-documented proposal often delivers a better organizational outcome than a system making decisions independently without sufficient context.
AI agents in business versus chatbots and automation
A chatbot primarily describes a way of communicating. It can answer questions based on instructions or a company knowledge base. An agent describes how a task is performed: it selects available tools and uses their results to proceed to the next step. A conversational interface can be part of either solution.
Traditional automation follows a predefined sequence of rules. An agent uses a model to interpret information or select an action within the permitted scope. The distinction is not always clear-cut, so specifications should describe concrete capabilities: reading data, preparing a proposal, writing a change, and the approvals required.
There is no point adding independent planning where a predictable workflow is sufficient. Structured rules are easier to test, and the model can be used only to interpret an unstructured fragment. This architecture limits the number of decisions whose correctness must subsequently be assessed.
Eight practical use cases
The following scenarios are proposed use cases, not accounts of results from specific implementations. Each requires defined inputs, correctness criteria, and actions requiring human consent. The same idea may be useful in one organization and uneconomical in another.
Customer service — gathering ticket context
An agent can identify a message's subject, retrieve contact history, and prepare a summary for a support representative. It can also identify missing information needed for further handling. It should use only records available to the particular user and distinguish customer information from binding system instructions.
Initially, it is worth leaving response sending to the representative. The employee should see the sources behind the summary, not just the finished text. If the customer cannot be identified unambiguously, the system should request a check rather than link records independently.
Customer service — checking case status
Following appropriate authentication, an agent can read the status of an order, reservation, or complaint and present it in understandable language. The answer should come from the current system, not the model's memory. Inability to access the source must be clearly communicated.
Permission to read a status does not have to include changing a date, canceling an order, or granting compensation. These operations require a separate risk assessment. Separating read and write tools makes it easier to introduce further capabilities without opening up the entire system at once.
Sales — preparing a salesperson for a conversation
An agent can gather the relationship history, recent agreements, and open tasks from the CRM. The result is a working note with source links and a list of questions. It should not fill in missing information about the customer's budget or present assumptions as confirmed statements.
An important prerequisite is organized access to sales data. The agent should not disclose information from proposals the user is not entitled to view. If sources conflict, it is better to highlight the discrepancy than choose a convenient version without justification.
Sales — drafting a proposal and checking completeness
Based on an approved service catalog, an agent can prepare a proposal structure and identify missing data. Pricing rules, discounts, and terms of cooperation should come from a controlled source. The model should not independently promise deadlines or invent solution parameters.
Approval from the person responsible for the proposal is needed before sending it. The interface should distinguish data retrieved from the system, generated content, and fields still requiring a decision. This tells the reviewer which elements need particularly careful checking.
Back office — initial document handling
An agent can identify a document type, read fields, and check their completeness before forwarding them for further validation. For an illegible scan, it should pause the case or route it to an operator. Reading uncertainty must not be concealed by filling in data based on probability.
A good starting point is preparing a proposed entry without automatic posting. We discuss field checks, exceptions, and integration in the article on automated document processing. Text recognition and correct document understanding are separate stages requiring separate evaluation.
Back office — finding procedures and preparing a task
An agent can find the current procedure, identify required attachments, and prepare a task for the appropriate department. It should provide the source and document version. If the procedure has expired or materials conflict, it should not create a new policy based on the model's general answers.
This use case requires content owners. Without knowledge-base updates, an agent can efficiently spread outdated rules. Good practice is to let users report a problem directly alongside the answer and forward that report to the person responsible for the source.
IT operations — supporting incident analysis
An agent can collect alerts, specified logs, and information about recent deployments, then prepare a timeline. Conclusions should be presented as hypotheses requiring confirmation. A deployment followed by a failure does not, by itself, prove causation.
Initially, tools should operate in read-only mode. Restarting services, changing configuration, or rolling back a deployment requires explicit approval rules and a rollback procedure. External log content must not become instructions to execute administrative commands.
IT operations — preparing a technical change
An agent can prepare a proposed configuration update, a change description, and a list of tests. The result goes through code review and the standard deployment process. An agent should not bypass controls simply because it works faster than a human.
At Flex Funding, the frontend migration from Nuxt 2 to Nuxt 4 used AI support, while Go2ops was responsible for code review, QA oversight, and maintaining the environment in line with compliance requirements. This illustrates the importance of quality control in AI-assisted work, not proof of an agent acting autonomously.
Connecting to APIs and data
An agent should receive narrow tools corresponding to business actions rather than general database access. Reading the status of a specific order, for example, is easier to control than the ability to run arbitrary queries. Permissions must be checked on the application side, independently of model instructions.
Write operations need parameter validation, an operation identifier, and safeguards against repeated effects. An unavailable API should not trigger an unlimited retry loop. Limits on time, step count, and resource usage help stop a task that is not progressing toward a solution.
Risks and practical safeguards
The main risks are incorrect conclusions, information disclosure, and unauthorized actions. The content of a message, document, or page may also try to persuade the model to ignore rules. External materials must therefore be treated as data, with access control implemented outside the model.
Before launch, prepare:
- a list of permitted tools and the scope of access to each;
- human approval thresholds for actions with significant consequences;
- tests involving incorrect data and attempts to force unwanted actions;
- an operation log without unnecessary disclosure of confidential data;
- a shutdown procedure and a return to manual handling.
Current GDPR, AI Act, and industry regulatory requirements should be verified with a specialist. The assessment depends on the specific use case, data, and system role. Merely calling a solution an assistant does not determine the organization's obligations.
FAQ
Does an agent have to work entirely independently?
No. It can simply gather data and prepare a proposal. The level of autonomy should follow from risk and test results, not the product's name. Human approval is a legitimate part of the design.
Does an agent need access to the entire ERP?
Usually, you should begin with limited operations exposed through an API. The scope should match the task and user permissions. Broad access makes the consequences of an error harder to assess and is not a prerequisite for usefulness.
How do you measure an agent's work quality?
Assess the correctness of the final result, the number of interventions, total task time, and incorrect or unnecessary actions. Also check cases where the agent should refuse a task. Fluent answers alone are not enough.
When is conventional automation better?
When steps are predictable, data has a fixed format, and decisions can be expressed as rules. An agent does not have to replace the entire workflow. It can handle only the part requiring text interpretation.
Summary
An agent's value depends on access to reliable data, limited tools, and predictable exception handling. Start with a use case whose result can be assessed without giving the system broad permissions. We describe the first pilot plan in our guide to implementing AI in a company.
If you are considering a specific process, let's discuss AI & automation. Together, we can define the boundaries of the agent's responsibility and check whether this way of working meets the team's needs.
Maciej Zaleski
Managing Director, Go2ops