The AI agent revolution underway   

Today we are witnessing the first wave of deployment of AI agents in companies. Agents are characterized by their ability to "reason", trigger actions, and improve over interactions. This is a major break with the logic of co-piloting and classic automation: the AI agent must be considered as a resource (Persona) of the organization. A perfect integration of AI agents into end-to-end processes appears to be a condition for success in bringing all the expected value to the customer and the company. Our generative AI observatory, based on more than 2700 respondents (CxO, CTO, CMO), reveals a key element: for 22% of respondents, the first success factor is the integration of the agent into business processes and the daily life of teams.   

This requires an in-depth reflection on the roles and responsibilities of agents:  What is the expected value? What exposure to customers and internals? What increase in power and human control in the system? These questions certainly structure the use cases and the trajectory of adoption programs during the agentic revolution.  

4 trends observed in the more than 30 Proofs of Concept carried out in the network in the first quarter of 2025  

Deloitte Digital has supported more than 30 Proofs of Concept of Agentforce, the chatbot developed on the Salesforce platform. Here are some of the lessons learned through these projects: 

• More than 15 use cases identified underline the range of possibilities with agents and above all the necessary adequacy with the specific context of each company. See illustration: 2 cases are predominant: request management (representing 15%) and customer file management (with 25%), the other cases being around 5%. 
• More than 70% of the agents are service-oriented, marking a strong emphasis on companies on this perimeter with levers of efficiency and additional value to human activities. 
• 35% of agents are exposed to customers for activities with a low-risk scope, marking a form of precaution and learning necessary to move to a more advanced stage. This is the reason why a human loop is sometimes preserved. 
• Initiatives are present in all industries in fairly equal ways: Automotive, Consumer Consumer, Energy, Finance, Life Science, Retail, etc. 
  

Want to know more about these Proof of Concept? Listen to Arnaud Le Pestipon talk about it in the Experiences Podcast (in French.)  

Two concrete examples :  

Use cases are multiplying in all sectors. Here are two examples to illustrate them:   

Bank – Canada  

A credit player offered its audience (e.g. newcomers to the territory, students, etc.) a web page to select a credit card, with the constraint for the customer to click a lot and to laboriously scan the general terms and conditions. To provide interaction in a more natural language, an agent was positioned to handle customer requests, recommend solutions, and initiate the card request. If the customer is interested, the request is considered and in a 1st version of the system, a call agent to validate the file.    

Retail – United States  

A distributor faced with a high volume of emails concerning orders (tracking, delays, delivery) set up an agent to automate the response. 6,000 emails per month were thus handled with, in a first batch, a human validation loop kept.    

First challenge identified: to establish a vision of the agentic   

Two major findings emerge from the feedback from the field: 

1. Agentic vision is still blurred in many general directorates. This hinders the identification of priority use cases. 
2. The need for change management is greater than traditional digital transformations. The AI agent profoundly modifies the logic of work and responsibility.  

To succeed, it is essential to structure an exploratory trajectory in three stages: 

• Define a vision and use cases that are outdated and offer a concrete and measurable benefit 
• Seeking a balance between short-term ROI and long-term structuring 
• Work in co-construction with the business lines from the ideation stage to integrate the agent in an end-to-end logic 
  

Second challenge: adapting project-based methods  

An observation from the projects carried out: the challenge of precision implies rethinking the working model. The issue is not in the more than 70% accuracy, the agent will take you there quite quickly, but in the fact of hitting the more than 90%. This has concrete implications: we are moving from a TESTING cycle  to a TUNNING cycle. The time for classic test phases with an agent is over. The steering ratios are split equally 50/50 between design and tunning.   

To make this adaptation a success, some concrete adaptations are necessary:  
 

• Emphasizing operating data:  
  The agent delivers content according to the way it is fed. It's the great return of the knowledge workshops. You have to structure the content and rely on KCS (Knowledge Centered Service) methods.   
• Structure the fundamentals of implementation: 
  • Precisely define the scope of the agent 
  • Carefully configure and check agent permissions 
  • Define the management of escalations... and perhaps favor human control at startup to secure a step towards complete automation for the customer  
• Professionalize the Tunning phase: 
  • Move forward by successive refinement (the funnel model no longer works) 
  • To provide more expert time on phasing (where before we were trying to limit the load of interns) 
  • Organize a short loop between developers and testers 

Three imperatives to secure the value of the AI agent  

1. Regulatory compliance: it is imperative to manage strict compliance with the European Artificial Intelligence Regulation (EU-AI Act) since February 2025, any feature deemed "unacceptable" (exploitation of vulnerabilities, emotional inference, facial recognition, etc.) Is strictly prohibited under penalty of penalties of up to 7% of turnover. 
2. Business case: it is necessary to clarify the business case, the profitability of an agent must be assessed at the dawn of the volume and the value created, especially if it is exposed on the customer side.  The leverage effect on profitability linked to a poorly appreciated use case can prove to be a blocker to continue the agent's activity. 
3. Control of content: you must put safeguards in the agent himself in what he can and cannot do. It is necessary to define the management of escalations... and prioritize human control at startup to secure a step towards complete automation for the customer.