.png)
Agentic AI has held board attention for eighteen months. It holds far less of production: among large enterprises, 40% of respondents report scaling AI agents, up from 27% a year earlier. Real progress, and it still leaves six organisations in ten stuck at the pilot. This article answers three questions in the order they actually arrive: what an agent is, how it works, and what separates the projects that reach production from the ones that stop before it. Every figure is sourced, and the examples come from deployments we ran.
Agentic AI describes systems able to act on their environment on their user's behalf. They decide on a sequence of actions, use tools, and pursue a goal without human sign-off at every step.
The wording is not ours. It is France's data protection regulator, the CNIL, which in its 20 July 2026 note describes a technology “now able to act on its environment in place of its user” (original French). The regulator adds two characteristics that commercial definitions usually leave out: persistent memory, and the ability to interact with many different services.
That combination is what changes the nature of the tool. A generative model produces content you review, approve, and then use. An agent produces an outcome: the case is processed, the order is reconciled, the reply is sent.
Generative AI answers a request; agentic AI executes a task. The first produces text, an image, or a snippet of code that you then use. The second chains actions inside your systems until a result is reached.
The difference is not model power. It is access and autonomy: what the system is allowed to touch, and how far it goes without you.
CriterionGenerative AIAgentic AITriggerA promptAn objectiveUnit of workOne answerOne task carried to completionSystem accessNone, beyond copy and pasteTools, APIs, databasesMemoryLimited to the conversationPersistent across runsEnd of runThe answer is writtenAn exit condition is metOversightHuman reviewDefined guardrails and checkpoints
The practical consequence fits in one sentence: a generative AI project is run like a content project, an agentic AI project is run like a software release. Companies applying the first method to the second problem are exactly the ones that never leave the pilot.
An agent combines three things: a model that reasons, tools that act, instructions that bound it. It then runs in a loop until it reaches an exit condition set in advance.
OpenAI describes those three components in its agent design guide as the model, the tools, and the instructions, plus the notion of a run, “typically implemented as a loop that lets agents operate until an exit condition is reached.” Translated into project language: you do not program the steps, you define the goal, the means, and the limits.
A chain of prompts follows a path you wrote. An agent picks its own at each turn, based on what it observes. That freedom is useful where traditional deterministic and rule-based approaches fall short: complex decisions, rules that are hard to maintain, unstructured data. It is useless, and expensive, on a stable process that three conditions describe perfectly.
Hence the guardrails, which the same guide frames as a layered defence mechanism: model-based checks, explicit rules, application controls. Human intervention remains “a critical safeguard,” triggered by a failure threshold being crossed or by a high-impact action. An agent in production with no defined escalation threshold is not an autonomous agent, it is an incident waiting to happen.
Three questions settle it in a vendor meeting. Can the tool use a tool you add later, without redevelopment? What does it do when a step fails, stop or try something else? Where are the escalation threshold and the action log defined?
The question is not rhetorical. Gartner estimates that only about 130 of the thousands of agentic AI vendors are genuine, the rest being what the firm calls “agent washing”: existing assistants, robotic process automation, or chatbots repositioned on the term of the moment.
A product that cannot answer all three is automation. That is not disqualifying in itself, but it does not justify an agent's price.
Scaling is progressing, and it remains a minority. The gap between pilot and production is the real story of the year.
The figures come from the McKinsey survey published on 25 August 2026, covering 1,719 respondents across 97 countries: 40% of respondents at large organisations report scaling AI agents, up from 27% last year, while the share at smaller organisations stayed flat at 22%. On returns, 37% of respondents attribute at least some EBIT impact to AI use, unchanged year on year, and roughly 6% clear the 5% mark.
What separates the two groups is not the tooling. Nearly three quarters of the highest performers report having fundamentally redesigned their workflows, against one quarter of everyone else. Gartner separately forecasts that 33% of enterprise software applications will include agentic AI by 2028, up from less than 1% in 2024. So the question is not whether your processes will be affected, but which ones, and under what control.
Gartner forecasts that over 40% of agentic AI projects will be cancelled by the end of 2027, for three reasons: escalating costs, unclear business value, inadequate risk controls. None of the three is a technical problem.
This is the point most often misread. Projects that stop almost never stop because the agent did not work. They stop because nobody could say what it returned, because cost per run was never calculated, or because the risk function discovered the agent's data access scope three weeks before go-live.
The first is the easiest to handle and the most often skipped: an agent running in a loop consumes tokens at every turn, and that consumption gets projected before deployment, not after. That is the subject of our guide to managing token spend.
Agents that hold up in production handle a repetitive, measurable, bounded task: data reconciliation, high-volume content production, case processing. They are rarely the most spectacular use cases identified at the start.
At Club Med, the programme surfaced 102 candidate agents with 50 ambassadors trained. Eighteen were kept as quick wins, 57 documented as structural projects, for 531 hours saved per week. That filter from 102 down to 18 is the part of the work nobody talks about, and it is the part that decides the rest.
Two further examples, with the figures as they were measured:
What these deployments share is not the technology, which varies case by case. It is that each had a defined metric before the first line of code. Our AI agents in production all start there.
Deploying an agent transfers no legal responsibility. The GDPR continues to apply in full, and the system's autonomy makes attributing responsibility harder, not easier.
The CNIL and France's AI and Digital Council published an exploratory note on the subject on 20 July 2026. It finds that this decentralised operation “complicates the identification of each party's responsibilities”, and that personal data circulating across many services creates “a real risk of losing control over one's personal data”.
In practice, three things have to exist before go-live and not after: a written, approved data access scope, an action log usable in an incident, and a human checkpoint on high-impact actions. That is what governance covers, and it is the part rushed projects handle last. We frame it during the audit and roadmap, before the first agent.
Start with triage, not technology. Identify use cases broadly, then keep only a handful, on two criteria: measurable value and an acceptable level of risk.
The method we apply has four steps. Identify: inventory repetitive tasks with the teams who perform them, not with IT alone. Decide: quantify expected value and cost per run, and drop anything that cannot be measured. Build: a first agent in days or weeks, on a narrow scope, with its metric. Deploy: guardrails, logging, escalation threshold, and production monitoring.
Two notes on the team. An agent designed without the people who do the task today almost always misses an edge case they know by heart, which is why we involve teams before deployment. And internal skill-building is not an extra, it is what determines whether the agent survives the vendor's departure, which is the point of our training programmes.
On the model itself, we are an OpenAI Select Partner and we also build on other models depending on the use case: the vendor is chosen after the need is defined, never before.
Are agentic AI and AI agents the same thing?
Agentic AI is the field, an AI agent is its concrete unit. You say agentic AI to describe the approach, and AI agent to describe the specific system handling a given task.
How long does it take to put an agent into production?
A simple agent on a narrow scope takes days to weeks. The real timeline depends less on development than on approving data access and setting up the control framework.
Do you need technical skills in-house?
To build a first agent, no: low-code and no-code formats are enough for business teams. To run a system in production, you need a team able to monitor, fix, and evolve it, in-house or with a partner.
Is agentic AI compatible with the GDPR?
Yes, the European framework applies and does not prohibit these systems, but the CNIL notes that its implementation needs adapting. In practice, the scope of access to personal data must be defined, documented, and limited before go-live.
Want to know which of your processes justify an agent, and which do not? Book a 30-minute call and we will map your first three use cases, with no preparation needed on your side.
