AI agents move from simple dialogue to autonomous action. Their effectiveness is based on a precise framework and clear objectives. Human expertise remains essential to manage and validate their results.
In 2024, thecompany Alfi Technologiesan industrial ETI, used a classic chatbot for its customer support. But customers continued to call the after-sales service in large numbers due to a lack of precise answers. More recently, Alfi deployed Aidy, an AI agent capable of interpreting 20 years of technical documentation and machine IoT data in real time in natural language. It has been validated by senior technicians to guarantee the safety of the procedures. This agent enabled a 30% reduction in incoming after-sales service calls in just 6 months. It reduces equipment downtime or accelerates the onboarding of new operators.
This example illustrates a failover. The classic chatbot responds to a request. He produces text, reformulates, synthesizes, proposes ideas. An AI agent starts with an objective and can chain several steps together to achieve it. It has a self-correcting ability that the chatbot does not have. He benefits from autonomy in planning a task, can query sources, use tools, compare options, produce a deliverable. Sometimes, even trigger an action in business software if it has been connected to the agent. An effective agent also has a “working memory”.
“It is this transition from ‘responding’ to ‘doing’ which profoundly changes business usage,” points out Matthew Chaponvice-president of AI and research at the Heroiks agency. “Agents can execute multi-step processes in the real world.” However, it is recommended to leave the human in the loop. The challenges: define a framework, establish quality criteria, establish control points and document decisions.
Manage agents by a supervised delegation
To properly use AI agents in business, different concrete prompting and delegation techniques can be implemented. According to Mathieu Chapon, a good agentic prompt must contain at least five elements.
- First, thefinal goal. “You should not just say: ‘analyze this page’, but rather: ‘prepare an actionable SEO recommendation to improve the organic visibility of this page on this search intent’.”
- Then, the business context. “You have to explain the sector, the target, the level of requirements, customer constraints, the tools available, the internal rules. An agent without context will optimize in a generic way.”
- Also, the expected steps are to be marked. “You must ask the agent to break down their work: diagnosis, collection of information, analysis, recommendations, prioritization, final deliverable.”
- In addition, the quality criteria are to be added. “This is often the forgotten point. It is necessary to say what a good response is: prioritized, justified, actionable, non-generic recommendations, with a distinction between quick wins and heavier projects.”
- Finally, the autonomy limits are to be noted. “You need to specify what the agent can do alone, what they only have to suggest, and what requires human validation. For example: ‘You can suggest content changes, but you must not publish them without validation.'”
At the delegation level, several differences can be observed between the classic chatbot and the AI agent. For Mathieu Chapon, “the real skill that is emerging is not only prompting. It is the ability to manage AI agents. We must learn to delegate to them as we would to a junior. Namely: clarify the expected result, give the context, break down the mission, set the standards, control the critical stages.” He recommends a simple technique: don’t delegate an entire task at once. It is better to ask the agent to first produce their action plan, then validate it. Then just let it run. “This avoids a lot of errors, especially on sensitive subjects like SEO, data, customer relations or content production.”
Another good practice according to Mathieu Chapon is to use “checkpoints”. For example, you can write: “Before writing the final deliverable, present to me your diagnosis and the hypotheses you have retained.” “This forces the agent to explain its operational reasoning, and it allows the human to correct the trajectory before the tool produces a complete but potentially misdirected result,” analyzes Mathieu Chapon.
For him, AI agents therefore do not eliminate the need for expertise. “On the contrary, they make expertise even more important. Someone who does not know how to evaluate a good SEO, legal, HR or financial deliverable will not know how to properly control an agent either. AI accelerates execution, but it does not replace judgment. For me, the challenge of the coming months in companies will therefore not only be to “deploy agents”, but to create a real culture of augmented delegation: knowing what tasks to entrust, how to frame them, how to control the results, and how to integrate these agents in existing tools and processes.”
Business processes under control
Schematically, we can note different techniques, according to different positions in the company, through specific scenarios.
The framework: example for monitoring and decision-making.
Here we seek to move from simple scanning to the analysis of strategic contradictions. The master prompt:
L'objectif final : "agis comme mon shadow Manager. Ta mission est d'extraire les contradictions stratégiques et les signaux faibles du secteur [X] pour préparer ma prise de décision sur le projet [Y]. Je veux pouvoir trancher en X minutes." Le contexte métier : "nous opérons dans un environnement [Secteur] très concurrentiel. Le niveau d'exigence est celui d'une note de synthèse pour direction générale. Contrainte : ignore les communiqués de presse lisses, concentre-toi sur les analyses d'experts et les rapports financiers." Les étapes attendues : - "Scan : identifie les 10 actualités majeures du jour sur le secteur [X]." - "Filtrage : ne retiens que celles impactant directement les piliers de notre projet [Y]." - "Analyse comparative : identifie si deux sources se contredisent sur une tendance." -"Synthèse : rédige 3 points clés avec sources citées." Les critères de qualité : "une bonne réponse doit mettre en lumière un risque ou une opportunité que je n'aurais pas vue seul. Évite le résumé descriptif ; je veux une analyse d'impact "si [A] arrive, alors notre projet [Y] risque [B]"." Les limites d'autonomie : "tu es libre de sélectionner les sources, mais tu dois me présenter tes 3 points clés sous forme de brouillon. N'envoie aucune alerte à l'équipe sans ma validation."
Tip of Gregory Robinsonfounder of Beyond AI: “an executive does not ask ‘summarize that for me’. He asks for a decision-making deliverable. The magic word is ‘I want to be able to decide in X minutes’. This forces the agent to prioritize.”
The wizard: example for multi-tool workflow
In this case, we try to go from file creation to management of a secure customer onboarding pipeline. The master prompt:
L'objectif final : "orchestre le workflow d'onboarding client dès réception d'une confirmation. L'objectif est d'éliminer toute latence administrative entre la vente et le démarrage opérationnel." Le contexte métier : "nous gérons des clients grands comptes. Les outils à utiliser sont [Gmail, Google Drive, Google Calendar]. La règle d'or est la confidentialité : aucun document ne doit être partagé en dehors du domaine de l'entreprise. Les étapes attendues : - "Détection : dès qu'un e-mail contient une confirmation de signature, extrais le nom du client et le type de forfait." - "Architecture : crée un dossier dédié sur le Drive avec les sous-dossiers [Contrats, Livrables, Factures]." - "Planification : ajoute une "Réunion de Lancement" de 45 min au calendrier de l'équipe sur le premier créneau libre commun." - "Action : prépare l'envoi du formulaire d'onboarding personnalisé." Les critères de qualité : "le workflow doit être exécuté sans erreur de nomenclature (format : CLIENT_PROJET_DATE). L'invitation calendrier doit inclure l'ordre du jour standard en description." Les limites d'autonomie : "tu peux créer les dossiers et les invitations, mais l'envoi du formulaire au client est soumis à ma validation. Notifie-moi sur Slack dès que le pipeline est prêt."
According to Gregory Robinson, “for this profile, we try to stay in draft mode. The agent prepares, the human sends. It’s the only discipline that avoids disasters in B2B: the tone slipping, the confidential information leaking or even the poorly timed meeting.”
Freelancing: example for prospecting and sales.
For this example, we can go from identifying leads to writing personalized approaches based on the portfolio. The master prompt:
L'objectif final : "agis comme mon agent de croissance. Ton but est d'identifier 5 prospects qualifiés et de rédiger des approches de prospection basées sur mon portfolio." Le contexte métier : "je suis freelance en [métier]. Ma cible est [type de client]. Mon avantage concurrentiel est [point fort]. Je dispose de mon portfolio en PDF et de mon profil LinkedIn comme références. Les étapes attendues : - "Recherche : identifie 5 prospects sur LinkedIn correspondant exactement à ma cible." - "Audit : analyse leur activité récente (posts, actualités entreprise) pour trouver un "point de douleur" (pain point)." - "Matching : relie ce besoin à un projet spécifique de mon portfolio." - "Rédaction : rédige un message d'approche de 3 paragraphes maximum." Les critères de qualité : "le message ne doit pas ressembler à un template. Il doit commencer par un élément spécifique à l'actualité du prospect et démontrer une valeur immédiate. Pas de vente agressive, juste de l'aide." Les limites d'autonomie : "tu peux identifier les leads et rédiger les messages. Interdiction d'envoyer les invitations ou les messages. Présente-moi la liste et les textes dans un tableau pour que je puisse les envoyer manuellement."
According to Gregory Robinson, “for a freelancer, the trap is to want an agent who does everything. It’s better to start with an agent on a difficult task, master it, then pile on. Otherwise, you can spend the week debugging the agents.”
The small team: example for the progress of the project
In this illustration, we try to start from simple updating to guarantee consistency between discussion and execution. The master prompt:
L'objectif final : "garantis la "vérité du projet" en synchronisant nos échanges Slack avec notre tableau Trello. Ton but est qu'aucune décision d'équipe ne reste lettre morte."
Le contexte métier : "petite équipe de [Nombre] personnes. Nous utilisons le canal #projets pour décider et Trello pour exécuter. Le niveau de réactivité attendu est élevé : mise à jour en temps réel."
Les étapes attendues :
- "Surveillance : analyse en continu le canal #projets."
- "Diagnostic : détecte quand un message implique une nouvelle tâche ("@nom, tu peux faire X ?") ou une décision actée ("Ok, on part sur X")."
- "Action : crée ou mets à jour la carte Trello correspondante."
- "Livrable : poste un court message de confirmation sur Slack : "Carte Trello créée/mise à jour pour [Action]"."
Les critères de qualité : "chaque carte créée doit impérativement avoir : un titre clair, le nom du responsable assigné, et un lien vers le message Slack d'origine pour le contexte."
Les limites d'autonomie : "tu peux créer et déplacer des cartes. Cependant, si une tâche n'a pas de responsable clair ou d'échéance, ne devine pas. Pose la question directement dans le fil de discussion Slack avant d'agir."
According to Gregory Robinson, “in a small team, it is better to appoint a single supervisor from day 1, with 30 minutes per day blocked in his calendar. Without this, the agent can drift away in 15 days.”
AI integrates into processes through various triggers
Note that with AI agents, the prompt is no longer necessarily limited to a conversational interface, like ChatGPT. It can be placed in the agent interface, in a business tool, in an automated workflow, in a no-code platform like n8n or Make, in a CRM, a support tool, an SEO tool, or directly in an application via API. “The difference is that the prompt often becomes a system instruction or a mission instruction, integrated into a process,” observes Mathieu Chapon. Jeremy Lacostegeneral manager at Eskimoz France, advises to “favor Skills or Gems, that is to say mega prompts which are very well worked and historicized in the interfaces in order to replay the sequences”.
To activate the prompt, there are several cases. Either the user launches it manually, for example by clicking on a button, sending a request or triggering an action in a tool. Either the prompt is activated automatically by an event. This could be receiving an email, adding a support ticket, or even a new line in a file. The agent can also self-activate via a scheduled task or active data monitoring, without a direct external event.
