According to the PwC AI Agent Survey of May 2025, 79% of companies already use AI agents in some form, but only 17% have crossed the threshold of large-scale deployment. That figure describes the situation precisely: a technology that has proven its value in well-prepared contexts, but has not yet demonstrated its ability to industrialise in more complex environments like contact centres. So why write about it now? Because the foundations are being laid today. And because a CX leadership team that waits for the technology to be “ready” before starting to understand what it is will find itself several years behind by the time they realise it.
The term "agentic AI" is used with very different levels of autonomy depending on the vendor or the article, which creates genuine confusion. It deserves a clear definition before going further.
A system is agentic when it combines five characteristics simultaneously: it receives an objective, not just a question. It breaks that objective down into steps autonomously, without being told the sequence. It uses tools: browser, API, database, CRM, code. It makes intermediate decisions based on what it observes at each step. And it verifies its results, adapts if something does not work, and continues toward the objective. It is this action-observation-decision loop that fundamentally distinguishes an AI agent from a conversational LLM. As the Armatis CX glossary puts it: "Unlike GenAI which generates, agentic AI acts."
A concrete example to anchor the difference. A customer reports an incorrect invoice. A GenAI chatbot can explain how to contest an invoice. An agentic AI agent can access the customer file, compare the invoice to the contract, identify the anomaly, create a credit note, notify the customer, and close the ticket, without a human adviser intervening. It is that shift from response to resolution that defines agentic AI.
This technology already exists and functions in several domains. In software development, systems break down a ticket, write code, test and correct in an autonomous loop. In research, tools navigate dozens of sources, read and synthesise without intervention at each step. In finance, agents monitor transactions, detect anomalies, and trigger procedures. This is not prospective: it is in production. Why is the contact centre lagging behind these contexts? Because customer relations is structurally more complex to automate agentically: heterogeneous data, often poorly integrated systems, high emotional load, specific regulatory constraints, and a much lower tolerance for error. It is not that the technology is not there: it is that the conditions to deploy it seriously take time to assemble.
To understand why everyone is talking about agentic AI, you need to look at the market projections and the first ground-level data. The agentic AI market is estimated at $7 billion in 2025 and is expected to reach $93 billion by 2032. That growth rate is not that of a niche technology being experimented with: it is that of a technology preparing to restructure entire markets.
On contact centres specifically, the projections are equally striking. Gartner estimates that 80% of routine customer requests will be resolved by agentic AI without human intervention by 2029. Cisco projects that 68% of customer service interactions will be managed end-to-end by agentic AI by 2028. Salesforce indicates that 30% of service cases are already fully resolved by AI today, with a projection of 50% by 2027. These figures are based on observations from real programmes, even if "cases resolved by AI" today mostly covers simple requests.
The potential savings also fuel the enthusiasm. Gartner estimates that conversational and agentic AI could reduce labour costs in contact centres by $80 billion by 2026 globally. And according to a Cornell study cited by PwC, mixed human-AI teams are already 60% more productive than 100% human teams on the same tasks. That last figure points to something important: the gain does not come from replacing the human, but from the combination.
Between five-year projections and zero deployment, there is an intermediate reality that deserves to be described precisely. Agentic AI is not absent from contact centres. It is present in partial forms, within well-defined perimeters, with technical prerequisites that are often underestimated.
The most documented case is Orange with its ClariFibre system, deployed at the end of 2024 to automatically handle fibre connection failures. ClariFibre captures data from around twenty systems, reasons about possible causes and acts automatically: sending an SMS, updating the file, correcting the identified anomaly. Result: ticket analysis time reduced by 75%, technician training accelerated by 50%. This is a mature agentic AI example, on a well-defined perimeter, with structured data and clearly delimited actions.
In the financial sector, a major French credit institution implemented an AI-driven campaign orchestration engine to individualise every customer contact, delivering a 30% marketing revenue increase. This is not agentic AI in the pure sense, but it follows the same logic: a system that acts autonomously, makes contextual decisions, and produces measurable results.
What these examples have in common: a well-defined perimeter, structured and accessible data, a measurable objective, and human governance over exception cases. These are exactly the conditions that are missing in most contact centres when they consider deploying agentic AI at scale.
Honesty about current limits is not pessimism: it is a useful reading for CX leadership teams that want to prepare seriously rather than endure the next wave of vendor demonstrations.
The first barrier is data quality. An agentic AI system is only as intelligent as the data it can access. If it queries a poorly maintained CRM, an outdated knowledge base, or systems that are not connected to each other, it will produce incorrect actions. As one market observer summarised: "Agentic AI does not compensate for knowledge that is vague, contradictory or not kept up to date." Before deploying agentic AI, you need to structure your knowledge. That is often the longest piece of work.
The second barrier is technical integration. Agentic AI needs to access heterogeneous systems in real time, read from them and write to them. CRM, ERP, ticketing tools, billing databases: each integration is a project in itself. Most contact centres are still working with systems that do not communicate fluidly with each other. This is an infrastructure prerequisite that technology alone does not resolve.
The third barrier is governance and compliance. Agentic AI makes decisions and triggers actions. Who is responsible when a decision is incorrect? What guardrails prevent an irreversible action on a sensitive file? In Europe, the EU AI Act came into force in August 2024 with general applicability from August 2026. High-risk systems are subject to transparency, human supervision, and audit obligations. This framework is not an obstacle to deployment, but it requires a rigour of design that projects driven purely by technology pressure tend to neglect.
The question everyone asks implicitly is that of employment. It deserves a direct answer, grounded in available data rather than convictions.
The most realistic projection is not one of massive replacement but of task redistribution. McKinsey puts it precisely: mixed human-AI teams are more productive, not fewer in number on the cases that matter. Interaction volumes do not decrease with automation: they increase, because the barriers to making contact are lowered. 57% of customer care managers anticipate an increase in call volumes over the next one to two years, precisely as automation accelerates.
What changes is the nature of the work. Advisers who today handle simple, repetitive requests will progressively be positioned on the cases AI cannot handle: complex situations, high emotional load, sensitive decisions, high-stakes commercial interactions. This repositioning is an opportunity for teams that prepare for it, and a risk for those that do not.
As one financial services CX director put it: "For simple cases, 100% AI. For complex cases, 100% human, supported by AI tools." That is the balance all serious market players are moving toward. Our article on the transformation of contact centre roles and the omnichannel strategy covers this evolution in depth.
For a CX leadership team that wants to prepare seriously without falling into the trap of the proof-of-concept that never reaches production, the path is clear.
Step 1: map interactions by complexity and volume. What are the most frequent, simplest, best-documented cases? These are the first candidates for agentic automation. A standard refund process, an address update, a delivery tracking request: these interactions have clear logic, available data, and a verifiable outcome. This is where agentic AI first demonstrates its value.
Step 2: audit the quality of the knowledge base and system integrations. Agentic AI amplifies what already exists: a good knowledge base becomes an autonomous resolution engine, an approximate one becomes a source of autonomous errors. This audit is often revealing, and almost always uncomfortable.
Step 3: define governance before deployment. Which cases require human validation, what confidence thresholds trigger an escalation, how are the agent's decisions traced and auditable? These governance questions are not constraints that slow the project: they are what allows it to industrialise without creating operational or regulatory risks.
At Armatis, this logic of controlled progression guides our technology approach. The SquAire suite integrates AI and process orchestration in an architecture designed to combine human performance and intelligent automation: SquAire Interaction for omnichannel centralisation, SquAire Quality for AI-enriched quality monitoring, SquAire Knowledge for knowledge management. Not to replace advisers, but to give them the means to focus on what has genuine value.
What is the difference between a chatbot and agentic AI?
A conversational chatbot responds to questions within a predefined framework: it understands an intent and produces a response. An agentic AI system acts: it breaks down an objective, connects to multiple systems, makes intermediate decisions, executes actions, and verifies its results, without constant human supervision. The distinction is not one of degree: it is one of nature.
Will agentic AI eliminate jobs in contact centres?
Available data points toward redistribution rather than elimination. Interaction volumes increase with automation, because the barriers to making contact are lowered. What changes is the nature of tasks: AI handles simple and repetitive cases, advisers focus on complex, emotional, and high-stakes commercial situations. Mixed human-AI teams are 60% more productive according to McKinsey data, which points toward role evolution rather than headcount reduction.
What are the prerequisites for deploying agentic AI?
Three prerequisites are non-negotiable: a structured and up-to-date knowledge base, functioning system integrations (CRM, ERP, business tools), and defined governance over automated decisions. These prerequisites are often the real work, and typically take longer than the technology deployment itself.
When will agentic AI be truly operational at scale?
First industrialisations on well-defined perimeters exist right now. Generalisation to complex and emotional interactions is still several years away. Gartner projects 80% of routine requests resolved by agentic AI by 2029. What is at stake in 2026 is laying the foundations: data, integrations, governance, team development.
How does the EU AI Act regulate agentic AI deployments?
The AI Act came into force in August 2024, with general applicability from August 2026. Systems classified as high-risk are subject to transparency, human supervision, traceability, and audit obligations. These obligations require rigorous design but do not block deployments.
Armatis is a European specialist in customer relations and business process outsourcing (BPO), operating across multiple continents with thousands of employees serving companies of all sizes and sectors. The company designs and manages end-to-end customer service operations: multichannel contact centres, complaints handling, technical support, back-office and digitised processes. Backed by integrated technology infrastructure and the ability to adapt to any sectoral and regulatory context, Armatis helps its clients combine operational performance, quality of experience and cost control, wherever they need it.
Contact our teams to discuss your challenges and find out how we can support you
Join the leaders who trust our multilingual and technological expertise.