Six decisions before you approve your first agent

Before the first agent

A company that uses AI well is not the one with the most models. It is the one that organizes its operation so people decide better, with agents working within clear limits. Adoption is no longer a distant bet: according to McKinsey, 65% of the organizations surveyed were regularly using generative AI in early 2024, nearly double its previous survey (McKinsey, “The State of AI”, 2024).

This guide sums up, in six decisions, what management at a mid-sized company should be clear on before signing.

1. Where the value is

The most common failure does not come from picking the wrong technology but from not knowing what you want to achieve. BCG found that 74% of companies have yet to show tangible value from their use of AI (BCG, “Where’s the Value in AI?”, 2024). Before approving, answer three questions:

• Which process loses money today? Not the most visible one: the one that costs the most hours or errors.
• What is fixing it worth? In business units: requests, documents, reconciliations.
• What changes if it works? Margin, service capacity, team time freed up.

If you cannot answer them on one page, you are not ready to approve an agent yet: you are ready for a diagnostic. This article explains what to review.

2. Build or work with a partner

For a mid-sized company, building an in-house AI team is usually the most expensive and slowest path: what it needs is agents that work, not a lab. Working with a partner that designs, implements and operates, while the internal team keeps control, is usually the balance point.

What matters is being clear on who answers for the agent day to day: the company decides, the provider operates, and both review results on an agreed schedule.

3. Capabilities, not departments

Organizing the plan by department (one pilot for sales, another for collections, another for service) multiplies projects without reusing what was built. It works better to think in three capabilities that serve the whole company:

• Conversation. Agents that serve customers on WhatsApp, web, voice and email, and hand the case to a person with full context when needed.
• Operations. Agents that carry out tasks on the company’s documents and systems, with a record of every step.
• Data. Dashboards that show what is happening across conversations and processes, with a view for each role.

4. Controls by design

Controls are not a committee: they are part of the design. The EU AI Act and the OECD principles point to requirements like these, and each one has a concrete answer:

• Traceability. Every action the agent takes is recorded and can be reviewed.
• Explanation. Answers cite their source; no source, no answer.
• Bias. Testing before going live and monitoring afterwards.
• Personal data. Each client’s information kept separate, encrypted and with defined retention periods.
• Continuity. If the agent fails, a person takes the case.

In financial services, these controls carry their own requirements. Leaving them for after the first rollout costs more than designing them in from the start.

5. Metrics the board understands

Useful metrics are not the technical ones but the ones the board can read without a translator. Four to ask for from the first month:

• Resolution without human intervention. The share of cases the agent closes on its own. At Cootradecun, 85% of requests.
• Cycle time. From request to answer, compared with the previous period.
• Unsupported answers. Cases where the agent answered without a source and the review caught it. It should drop month after month.
• Cost per case. Technology plus human supervision, divided by cases handled.

6. The order

The expensive mistake is starting with the most visible process. The order that works best:

1. Diagnostic. Understand the operation and choose where to start.
2. First agent in production. One area, one channel and clear decision limits. From diagnostic to production in 30 days.
3. Extension. Take it to the internal processes that depend on documents.
4. Dashboards. Measure the operation with what the agent already records.
5. Ongoing operation. Adjustments month after month, with the metrics in view.

What comes next

The World Economic Forum expects 39% of the skills the job market asks for today to change by 2030 (WEF, “The Future of Jobs Report”, 2025). The question for management is not whether there will be AI agents in the company, but who will run them, within which limits and with which controls.

A company that uses AI well is not measured by how many agents it has, but by how many of its operating decisions are recorded and can be defended.