AI agent in the company: how to move from initial tests to everyday use

In a company, people may start using artificial intelligence on their own initiative: someone asks for help writing a text, summarizing a document, or analyzing a spreadsheet. The person runs the task, chooses what to keep, and checks the result.

Another context appears when the company incorporates AI agents into a sales, purchasing, or service process. In that case, the agent can look up information, choose tools, and execute steps within defined conditions. Its work becomes part of a routine on which other people depend.

These uses can be combined. The difference that matters here is responsibility for the process: when an agent starts participating in the company’s work, it is necessary to define what it can do and how its actions will be monitored.

That monitoring is where ARGUS can help, with coordination, traceability, and visibility into the consumption of integrated agents. At the end of the article, we show how that support works in practice, using a comparison with a soccer team.

The global McKinsey research published in August 2026 helps size this transition. While 80% of participants reported improved individual productivity with AI, 37% attributed some positive impact to their organizations’ operating profit. These are different measures, but the contrast shows why personal gains and business results need to be assessed separately.[1]

The survey heard 1,719 participants in 97 countries, between May and June 2026. The numbers represent the responses from that sample, not a specific estimate for Brazilian companies or midsize organizations.[1]

For a midsize company, the path may begin with a bounded flow and a lean team. The essentials are being able to verify delivery and handle problems.

Define the work before releasing the agent

Imagine a fictitious company that tested an agent to prepare commercial proposals. It consults documents, organizes information, and delivers a draft that reduces the preparation work for the team.

After the first tests, the idea emerges to incorporate it into the routine. But some situations still need to be resolved. Which price table should be consulted? Can the agent access contracts for any client? And if information is missing, should it stop the task or prepare a partial proposal?

To move from tests to everyday use, the first step is to delimit its task.

It will be able to consult an authorized catalog and prepare the document for review. Granting discounts, assuming contractual commitments, or sending the proposal to the customer may still depend on human approval.

It is also necessary to define what will count as an acceptable deliverable. “Generate a proposal” says little if the document requires so many corrections that the team prefers to start over. It is worth assessing the consistency of the items, the use of current conditions, and the effort of review.

Microsoft’s guidance for agent adoption recommends identifying a business owner and establishing measurable outcomes for each initiative.[2]

In the example, someone in the sales area must answer for accepting the deliverable and updating the rules. Whoever maintains the integration answers for technical functioning. In a midsize company, the same person may accumulate several responsibilities, but they must be clear, including when there is disagreement about the quality of a proposal.

Having an agent working is the beginning of evaluation. The result needs to show up in the process it helps execute.

Grant the access required for the task

After defining the work, it is necessary to verify whether the access corresponds to what was authorized.

The agent may need to consult the service catalog and a current price table. That does not mean it should access all previous proposals or special conditions negotiated by the company.

Microsoft’s agent governance documentation recommends restricting access to the sources needed for the function and applying identity and permission controls. It also addresses care for data stored in memory and execution logs.[3]

In the commercial flow, this translates into concrete decisions: which documents can be consulted, who maintains their versions, what information can be sent to the model, and where the generated proposal will be stored.

The restrictions must be applied by the systems, tools, and integrations responsible. A written instruction in the agent’s configuration does not replace access control.

Just as a player needs to know their role and respect the rules of the game, each agent needs to operate within the permissions received. ARGUS supports applying those limits to agents and integrated flows, according to the policies and controls configured by the company.

Before releasing use, it is also worth testing the boundary: if the agent tries to consult a document outside its scope, will access be blocked?

Test the situations that appear in everyday life

The first tests usually use organized requests and selected documents. In routine work, incomplete information, old files, and contradictory guidance arrive.

The evaluation needs to represent those conditions. The NIST AI Risk Management Framework recommends documenting tests and metrics and demonstrating performance in conditions similar to the deployment environment.[4]

For the proposal agent, it is worth checking how it behaves when faced with a service outside the catalog, an expired table, or a request that requires a condition not yet approved.

In some cases, the acceptable deliverable will be a partial proposal indicating what is missing. In others, the expected behavior will be to forward the request to a person.

These criteria should be aligned before evaluation. That way, the team can distinguish a failure from a correct interruption: an agent that recognizes the lack of authorization may be doing its job.

It is also advisable to record which versions of the model, instructions, and sources were evaluated. A relevant change in any of these components may require new tests before continuing use.

Prepare the team to review and intervene

Human review needs a defined place in the process. Whoever receives the draft must know what to check, how to reject the deliverable, and where to forward an inconsistency.

If that person is unavailable, the flow needs a planned destination. The absence of review must not accidentally become authorization to send the proposal.

It is also necessary to agree on who can suspend the agent, how the team will continue the work, and which records will make it possible to identify affected deliveries.

NIST includes in AI risk management responsibilities and mechanisms to deactivate systems with outcomes incompatible with the intended use, in addition to monitoring, incident response, and recovery.[4]

Suspending new executions does not automatically correct information already sent. Therefore, the response must consider both the agent and the effects of its actions.

Monitoring begins together with real use. Microsoft recommends that the first agents go into production with active monitoring from the start.[2] In this example, the team can track accepted proposals, corrections needed, query failures, execution time, and consumption.

In that monitoring, ARGUS helps gather evidence of executions, tool calls, and consumption of integrated models. That information supports investigating failures and guiding adjustments while the team assesses delivery quality and business results.

Consumption deserves attention: in the 2026 McKinsey survey, about 20% of participants said that AI operating costs, including tokens, limited their use in organizations.[1] That reinforces the importance of observing consumption and quality together. A fast deliverable can lose usefulness if it requires successive attempts and a lot of rework.

Decide what is ready to enter the routine

Before releasing the agent, the company should be able to bring together a few answers:

• Which task is authorized and which actions require approval?
• Who answers for the deliverable and for technical functioning?
• Which sources, tools, and permissions were released?
• What was tested and what limitations remain?
• How to handle exceptions, suspend the flow, and continue the work?
• Who will monitor results, quality, and consumption?

That record can be simple and proportional to the process. Its value is to let everyone involved understand what was authorized and have the conditions to act.

Attention to these criteria also appears in market projections. In June 2025, Gartner predicted that more than 40% of agentic AI projects would be canceled by the end of 2027, citing rising costs, unclear business value, or inadequate risk controls as reasons.[5] This is a forecast, not an observed cancellation rate.

For those starting out, the practical implication is to assess the agent’s usefulness together with the conditions to keep it running.

If tests show that it prepares a useful draft but still does not meet the conditions for sending, authorization may remain limited to preparation. The work released must correspond to what the company has been able to evaluate and is prepared to monitor.

This is a concrete application of agentic AI governance: defining under which conditions a task can be delegated and keeping the capacity to intervene.

The move from tests to everyday use happens when the company can use the deliverable with clear criteria, recognize its limitations, and answer for how the process works.

From vision to practice: ARGUS

Think of a soccer team. Goalkeeper, defenders, and forwards have different roles and skills. The coach coordinates the work, and each player needs to know their role and the limits of their actions. In a process with several AI agents, there can also be specialists and a coordinating agent.

In this analogy, ARGUS helps organize the environment in which the team operates: it connects agents and tools, supports applying the rules defined by the company, and provides visibility into what happens during executions. The company sets the objectives and authorizations; the platform offers resources to coordinate and monitor integrated work.

In preparing commercial proposals, for example, one agent could consult the catalog and another prepare the draft, with coordination across stages and human review before sending. This division is illustrative: a simpler flow can work with a single agent. The design depends on the process need.

ARGUS supports monitoring and traceability of agents and integrated flows. The team can follow executions, tool calls, errors, and response time. If a proposal comes out inadequate, these records help investigate the path taken and locate steps that need adjustment.

The AI Gateway in ARGUS adds visibility into calls, tokens, and costs of integrated models, in addition to routing policies. That helps relate consumption to the work performed and provides information to review model use. It does not represent a guarantee of savings nor automatic control of the entire company AI bill.

Policies, permissions, and human oversight support applying the limits of action, according to the integrations and configuration adopted. The benefit for management is to bring together elements to decide with more clarity: where the agent is useful, which failures repeat, and what needs to change before expanding its participation.

This coordination, monitoring, and governance work is called AgentOps, or AI agent operations. AIOps, applying artificial intelligence to IT operations, can add operational context as a complementary resource. Using ARGUS also reaches processes in other business areas, respecting the integrated scope and the company’s responsibilities.

To assess this transition in your company, talk with the team about a specific flow and the controls needed to incorporate it into everyday work.

About Target Solutions

Target Solutions has worked for more than 15 years supporting organizations in evolving their infrastructure, monitoring, observability, automation, and operational intelligence operations. Over that journey, it has participated in projects in complex environments across telecom, technology, government, energy, and large enterprises, contributing to higher operational efficiency and service reliability.

As part of that experience, it developed ARGUS, its Operational and Agentic Intelligence platform, which combines AgentOps, AI Gateway, and AIOps capabilities to integrate operational context with AI agents’ actions, expanding visibility, governance, traceability, and control as people, automations, systems, and agents begin to share the operational environment.

About the author

Paulo Florêncio de Menezes is a partner and Co-Founder of Target Solutions and has worked for more than 15 years on projects related to infrastructure, monitoring, observability, automation, and IT and telecom operations.

He holds a degree in Data Processing from the Federal University of Amazonas (UFAM), an MBA in Business Management from Fundação Dom Cabral (FDC), and a Professional Master’s in Economics from Fundação Getulio Vargas (FGV).

In his Master’s research, he developed the IBEOP-IA (Brazilian Index of Occupational Potential Exposure to Artificial Intelligence), investigating the potential exposure of Brazilian occupations to AI capabilities by combining occupational data, artificial intelligence patents, and natural language processing techniques.

References

[1] McKinsey & Company. The state of AI in 2026: On the road to ROI. Published on 25/08/2026. Survey with 1,719 participants in 97 countries, conducted from 04/05 to 08/06/2026.

[2] Microsoft Learn. Your next steps for agentic transformation. Consulted on 29/10/2026.

[3] Microsoft Learn. Govern and secure AI agents. Consulted on 29/10/2026.

[4] National Institute of Standards and Technology (NIST). AI Risk Management Framework: AI RMF Core. Measure and Manage sections. Consulted on 29/10/2026.

[5] Gartner. Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027. Press release of 25/06/2025.