Le Tableau de Bord de Délégation d'AI Agent : Décidez quelles tâches confier et lesquelles garder
Pourquoi ce prompt est important
Teams that pick the wrong first task to delegate -- either something too risky, like client-facing communications with legal exposure, or something too trivial to demonstrate real value -- sour stakeholders on the entire initiative within one bad quarter. A poorly chosen pilot can stall agent adoption for a year; a well-chosen one builds the case for expanding scope.
À quoi nous l'utilisons
You're a marketing operations manager who just got budget approval for an AI agent tool, and leadership wants a rollout plan by Friday -- but your team has 40 different recurring tasks and no clear way to decide which 5 to hand to an agent first.
Prompt
Act as an AI workflow consultant who specializes in helping knowledge workers and small teams decide which of their recurring tasks are good candidates for AI agent delegation, and which should stay manual. CONTEXT: - My role: [YOUR JOB TITLE / TEAM FUNCTION] - Recurring tasks I'm evaluating: [LIST 5-15 RECURRING TASKS, ONE PER LINE, WITH ROUGH TIME SPENT PER WEEK] - AI agent tools currently available to me: [LIST TOOLS, e.g. "Claude with file access, a Zapier AI agent, ChatGPT with browsing"] - My risk tolerance for this rollout: [LOW - client-facing/financial/legal exposure / MEDIUM - internal-facing but visible to leadership / HIGH - drafts and personal productivity only] TASK: For each task in my list, score it on two axes (1-5 scale): 1. COMPLEXITY: how much judgment, ambiguous context, or exception-handling the task requires (1 = fully rule-based, 5 = requires deep contextual judgment) 2. RISK: the cost of a wrong or low-quality output going uncaught (1 = low stakes, easily reversible, 5 = high stakes, hard to reverse or externally visible) Then recommend one of three actions for each task: - AUTOMATE FULLY: agent handles it end-to-end, spot-checked periodically - AUTOMATE WITH REVIEW: agent produces a draft/output, a human approves before it goes live - KEEP MANUAL: not a good current fit for agent delegation CONSTRAINTS: - Never recommend "Automate Fully" for any task with a risk score of 4 or 5, regardless of complexity score - Flag any task where the underlying data or context needed to do it well is not clearly documented anywhere -- that's a blocker independent of the complexity/risk score - Be realistic about current AI agent capability in [CURRENT MONTH AND YEAR] -- do not recommend automating tasks that require live judgment calls involving other people's feelings, legal interpretation, or irreversible financial commitments - If two tasks have similar scores, prioritize whichever is higher time-cost per week for the pilot recommendation OUTPUT FORMAT: 1. A table: Task | Complexity (1-5) | Risk (1-5) | Recommendation | One-line reasoning 2. A ranked list of the top 3 tasks to pilot first, with a one-sentence explanation of why each is a good starting point 3. One task from the list that should explicitly NOT be automated yet, with a clear explanation of what would need to change (better documentation, lower stakes, proven agent track record) before revisiting it
Résultat
Delegation Scorecard for: Marketing Operations Manager
| Task | Complexity | Risk | Recommendation | Reasoning |
|---|---|---|---|---|
| Weekly social media scheduling from content calendar | 2 | 2 | Automate Fully | Rule-based, low stakes, easily corrected if wrong |
| First-draft blog posts from briefs | 3 | 2 | Automate with Review | Good first drafts possible, but brand voice needs a human pass |
| Client campaign performance reports | 2 | 3 | Automate with Review | Data pull is mechanical, but client-facing framing needs a sanity check |
| Responding to customer complaints on social media | 4 | 5 | Keep Manual | High visibility, reputational risk, requires emotional judgment |
| Competitor pricing page monitoring | 1 | 1 | Automate Fully | Simple scraping and diffing task, no judgment required |
| Ad spend budget reallocation between channels | 4 | 5 | Keep Manual | Financial commitment risk exceeds agent trust threshold |
| Internal weekly team status digest | 2 | 1 | Automate Fully | Low stakes, internal only, easy to spot-check |
Top 3 to Pilot First
- Competitor pricing page monitoring -- zero risk, immediate time savings (est. 3 hrs/week), builds team trust in agent output with no downside if imperfect.
- Weekly social media scheduling -- high time cost (5 hrs/week), fully rule-based once the content calendar exists, and errors are cheap to catch before a post goes live.
- Internal weekly team status digest -- internal-only audience means mistakes are low-cost, and it's a visible early win to show leadership the pilot is working.
Do Not Automate Yet
Responding to customer complaints on social media. This task combines the highest complexity and highest risk scores on the list: it requires reading emotional subtext, applying judgment about when to escalate versus de-escalate publicly, and any visible misstep becomes a reputational incident that's hard to walk back. Before revisiting, the team would need: (1) a documented decision tree for at least the 10 most common complaint types, (2) a mandatory human-approval gate before any agent-drafted response goes live, and (3) at least 90 days of agent-drafted-human-approved responses to build a track record before considering unsupervised operation on even the lowest-risk complaint category.
Chaque équipe qui déploie des AI agents en 2026 fait face au même mode d'échec silencieux : non pas choisir un mauvais agent, mais choisir la première tâche erronée à lui déléguer. Confiez à un agent quelque chose de trop trivial, et la direction n'y voit aucune valeur. Confiez-lui quelque chose de trop risqué – communications clients, engagements financiers, tout ce qui comporte une exposition juridique – et un seul mauvais résultat devient l'histoire dont les gens se souviennent, tuant l'élan de tout le déploiement.
Ce prompt corrige cela en imposant une comparaison structurée au lieu d'une sélection basée sur l'intuition. Il note chaque tâche récurrente sur deux axes indépendants : la complexité (combien de jugement et de contexte ambigu la tâche nécessite) et le risque (le coût d'un mauvais résultat qui passerait inaperçu). Ces deux axes comptent séparément – une tâche peut être simple mais à haut risque (approuver un changement important dans les dépenses publicitaires), ou complexe mais à faible risque (rédiger une mise à jour interne que n'importe qui peut corriger rapidement si elle est erronée).
La contrainte dure intégrée dans ce prompt – ne jamais recommander l'automatisation complète pour quoi que ce soit qui note 4 ou 5 sur le risque, quelle que soit sa simplicité – est ce qui empêche de simplement approuver ce qui semble pratique. Cela force également une réponse explicite sur ce qu'il NE faut PAS automatiser encore, et ce qui doit spécifiquement changer avant de reconsidérer cette décision. C'est la partie que la plupart des équipes sautent, et c'est la partie qui évite une erreur publique douloureuse six semaines après le début du déploiement.