Since the launch of ChatGPT in November 2022, generative artificial intelligence has spread at a speed few tools have ever known. Today, having access to an LLM (Large Language Model) is no longer a competitive advantage: it is a minimum.
The next wave, the one that touches how work is organised rather than how content is produced, is agentic AI. That is what Édouard de Mézerac, CEO of Artefact, explains in the interview Inside the layers of the agentic revolution given to the TCG Summit. And it is why companies such as Samsung, Google, OLX, Workday and Perplexity, with its Computer tool, are already building agents into their offerings.
Generative AI: a first revolution already outdated?
What generative AI is
Generative AI refers to a family of models dedicated to creating data: text, images, video, audio. It relies on large base models, trained on massive corpora, to produce summaries, answers to questions or any other ready-to-use content.
That power has a structural limit: generative AI acts on instruction. It answers a prompt, proposes a solution, generates a summary. But it is the human who asks, validates and executes. For a procurement department receiving quotes, price schedules, bills of quantities, invoices and framework contracts every week, this conversational model quickly shows its limits: nobody is going to “ask” a model to check every line of every document.
The best-known generative tools, by field
- Text: ChatGPT, Claude, Gemini.
- Image: Midjourney, DALL·E.
- Video: Lumen5, Runway, Sora.
- Audio: Beatoven, ElevenLabs, Suno.
Agentic AI: autonomy in the service of a defined task
What agentic AI is
Agentic AI is the extension of generative AI. It does not just produce content: it pursues an objective, chains the steps of a workflow, adapts to what it reads in its environment, with a high degree of autonomy on standard cases. Where generative AI answers, agentic AI acts.
To do so, it uses generative models as building blocks and adds rules, tools and memory, in order to build a program capable of acting on a real, bounded environment. That autonomy has a frame: it applies to tasks defined in advance. Anything that commits money or a relationship remains a human decision. That is, at any rate, the line we hold at Zylio, described on the Zylio AI page.
The six types of agent
- Goal-based agents. They align every decision with objectives set upstream. Suited to managing projects or processes where precision and efficiency matter.
- Model-based agents. They hold a representation of their environment, which lets them spot anomalies, plan better and reduce risk in partially known or changing contexts.
- Reactive agents. They automate repetitive tasks and reduce human error. The validation of alerts or financial thresholds can be entrusted to them.
- Learning-based agents. They gradually improve a process, adapt to market changes and reduce errors over time.
- Utility-based agents. They arbitrate between competing objectives (cost, lead time, quality) and optimise resource allocation in complex decisions.
- Collaborative agents. They coordinate other agents with each other in interdependent environments, and limit human intervention to the cases that require it.
What is the difference between generative AI and agentic AI?
The main difference is the capacity for automation. It shows in six aspects:
- Content creation. Generative AI produces articles, summaries or answers on demand; agentic AI produces content only if it serves the action.
- Analysis and context. One analyses data when asked; the other correlates information and context to decide what comes next.
- Problem solving. One proposes solutions or best practices; the other diagnoses and handles the cases that fall within its rules.
- Decision-making. One supports human decisions; the other makes operational decisions according to defined rules, and escalates the rest.
- Autonomy. One acts on a prompt, with low autonomy; the other acts independently on the objectives it has been given.
- Personalisation. One relies on immediate context; the other on history and long-term preferences.
What are the benefits of agentic AI in business?
The benefits appear above all when agentic AI is built into internal processes, where repetitive tasks recur every day on significant volumes. It can be integrated into most departments: customer relations, procurement, after-sales, finance, marketing. In each case the gain is the same: tasks that never needed to be manual stop being so, and teams keep control of what requires their judgement.
In which areas is agentic AI strategic?
Agents can in theory be integrated into every function. Some areas stand out for a fast return on investment:
- Customer service, through self-service capabilities that relieve teams.
- Marketing and communications, through automatic adaptation of campaigns to languages, markets and products.
- IT, through supervision of operations, cost tracking and proactive error handling.
- Procurement and finance, through contract analysis, quote comparison and detection of invoicing anomalies.
This last area is particularly strategic, because document volume is high and every undetected variance has a direct cost. Procurement departments that automate variance control and data reliability give teams back 15 to 45% of their time, per FTE, depending on scope and data maturity: time that goes to negotiation, strategy and the supplier relationship. Our AI agents page describes what each agent reads and checks.
How to deploy agentic AI in your organisation: five steps
An agentic AI strategy cannot be improvised. Five steps, not exhaustive, structure a solid implementation.
- Define the problem to solve. Identify the most time-consuming processes and the departments where automation would create the most value. Ask the teams: they are the ones who know the day-to-day irritants.
- Secure data quality. This is the prerequisite. An agent only performs if the data it processes is reliable, structured and usable. Data governance is not a technical constraint; it is what lets the system understand your environment and build a solid history. A single repository linking contracts, orders and invoices is the foundation.
- Choose the tools. Once the problem and the data are framed, decide which agents to integrate. In-house development or a ready-to-use solution? The choice depends on your resources, your technological maturity and your timeframe.
- Integrate the tools. Successful adoption requires human support. Agents do not replace teams, they augment them: train people, communicate on progress, integrate feedback from the field. GDPR compliance and security requirements are anticipated from this phase; our security and compliance page details what we apply.
- Test, monitor, measure. The first months are decisive: track the indicators (time returned, variances detected, cases escalated), adjust the parameters and measure results on a defined scope.
Agentic AI applied to procurement
At Zylio, the agents specialise in checking purchasing documents: Capture extracts the data from each document while keeping the link to where it was read, Compare puts quotes on a common basis, Compliance confronts every invoice with negotiated prices, Matching reconciles invoice, order and receipt, Exception prepares the variances. No agent decides in your place: it shows the variance with its proof, and you decide.
If you want to see these agents work on your own documents rather than on a generic demonstration, book a demo on the scope of your choice.
