Agentic AI: from experimentation to industrialization

For several years, AI remained confined to POCs:

• isolated use cases

• difficult to demonstrate ROI

• difficulty scaling up

Today, a milestone has been reached:

• 30.4% of companies are in active development

• 57.3% already have agents in production

The subject is no longer feasibility but industrialization.

Not all companies are moving at the same pace:

• 67% of large groups are in production

• 50% of SMEs

This gap is explained by:

• existing infrastructures

• dedicated teams

• investment capacity

Industrialization is becoming a structural advantage

AI agents are already deployed in concrete cases:

• Customer service (26.5%)

• Data analysis (24.4%)

• Workflow automation (18%)

These use cases share specific characteristics:

• high repeatability

• high operational intensity

• direct business impact

Before: the problem was cost. Today, priorities have shifted:

• Quality → 32%

• Security → 24.9%

• Latency → 20%

The challenge is no longer making AI work, but making it reliable at scale.

Difficulties appear in production:

• hallucinations

• inconsistencies

• execution errors

But above all:

• poor context management

• instability of outputs

• drift in complex processes

Agentic AI introduces systemic complexity.

At scale, it is essential to understand what the AI is doing:

• 89% of companies have observability

• 94% of teams in production use it

It allows for:

• tracing decisions

• detecting errors

• improving performance

Without observability, there is no trust.

The single model is disappearing:

• +75% of teams use multiple models

Why?

• arbitrating cost / performance

• reducing risk

• avoiding lock-in

AI is becoming an architecture.

A standard architecture is emerging:

1. Models

1. Orchestration

1. Execution

1. Observability

1. Evaluation

1. Iteration

Agent engineering is becoming a discipline.

3 major transformations:

1. Multi-models → standard

1. Continuous evaluation → critical

1. Reliability > speed

Quality becomes a competitive advantage.

AI agents are no longer experimental. They are becoming:

• structuring

• critical

• industrial

The subject is no longer "testing AI" but "deploying it at scale"