“Prepare for the next sales committee meeting.” For an employee, this sentence opens a discussion. For an AI agent connected to files, a CRM and a messaging system, it can trigger a chain of actions: collecting data, comparing performance, drafting a summary or even sending invitations. The potential gains are considerable. So is the potential for misunderstanding. Looking ahead to September 2026, the real management challenge may be less about knowing how to talk to AI than knowing how to entrust it with a mission without handing over the power to decide on our behalf.
From chatbots that answer to agents that act
The trajectory is already apparent in developments documented in 2023 and 2024. Function calling in model interfaces allows an assistant to invoke external tools. Projects such as AutoGPT have popularized the idea of systems capable of breaking a goal down into steps. In 2024, the unveiling of Devin, an agent designed for software development, illustrated this ambition without, however, demonstrating reliable autonomy in every situation.
The difference from a conventional chatbot is operational. An assistant drafts a proposal; an agent can, depending on its configuration, search for information, run a program, modify a document or trigger an operation. It nevertheless remains dependent on its tools, its permissions and the model’s errors. Autonomy is not a binary quality: it is a scope of action that someone must define.
By September 2026, the spread of these capabilities into office and business management software is a plausible prospect, not an established outcome reported here. The examples that follow are forward-looking scenarios. They extend a documented trend: bringing models closer to workplace tools. Their oversight therefore concerns not only developers, but also sales, HR and administrative managers.
A mission is not a vague instruction
Imagine a procurement manager asking: “Find the best supplier.” Best by what measure? Price, lead time, financial resilience, environmental impact? Without explicit criteria, the system may favor whatever is easy to measure or appears first in the available data. A convincing answer can thus conceal a trade-off that no one actually chose.
The first skill is turning an intention into a mandate. The manager should specify approved suppliers, comparison criteria, acceptable documents and the expected outcome. Above all, she should distinguish analysis from commitment: comparing offers does not authorize the agent to negotiate on the company’s behalf or place an order. Effective delegation starts with what must not be done.
An agent’s minimum brief
- Objective: what problem needs solving, and for whom?
- Scope: which data may be consulted and which tools used?
- Limits: which actions must be prohibited or subject to approval?
- Success: what criteria will determine whether the result is acceptable?
- Escalation: when must the agent stop and request a human decision?
This framework resembles good delegation to people. But one difference changes everything: an agent cannot be assumed to have a stable understanding of implicit rules. “Use common sense” is no substitute for a precise prohibition. And a limit written into an instruction is not equivalent to a technical safeguard: permissions must reflect the mandate.
Granting autonomy in stages
There will be a strong temptation to connect numerous tools to deliver an impressive demonstration quickly. The cautious approach is to start with read-only access, then allow changes to be prepared before they are executed under supervision. Broader autonomy comes only after representative testing, including ambiguous or incomplete cases. A successful demonstration is no guarantee of reliability.
For a sales campaign, for example, the agent can propose segmentation and draft emails. Sending them remains subject to approval. In another workflow, reorganizing recoverable work files may justify greater latitude. The level of oversight must reflect the consequences of an error: reversibility, confidentiality, cost, legal or reputational impact, and the number of people affected.
The manager must also set an action budget: maximum duration, request volume and number of changes. An agent that gets stuck in a loop or pursues the wrong lead can consume resources without generating value. A stop button, a designated person in charge and a rollback procedure are not technical details: they are the conditions for controlled delegation.
Checking evidence, not a polished answer
The most commonplace risk is far from spectacular. It is a neat table, a fluent summary or reasoning that seems to hold up. Models can confidently produce inaccurate information; a sequence of actions can then propagate the error. A piece of data misinterpreted at the outset becomes a ranking, then a recommendation, then an email ready to send.
Oversight therefore requires more than a stylistic review. For a sales analysis, managers should ask for the period covered, the sources consulted, any exclusions and verifiable calculations. For invoice reconciliation, they should examine anomalies and a sample of apparently straightforward cases. The aim is not to obtain an account of the model’s internal reasoning, but observable records: documents, operations performed, tests and results.
Human approval can itself become an illusion if the person responsible mechanically approves a long list of proposals. They must be given the information needed to decide, with uncertainties flagged and sensitive actions singled out. Useful oversight must be designed as a work stage, with time allocated to it, not as a checkbox added to shift responsibility.
The new soft skill: knowing how to challenge the tool
Managing an agent requires a familiar combination: clarity, critical thinking, the ability to weigh trade-offs and a willingness to acknowledge what one does not know. The difficulty lies in the interface. A system that is available, fast and polite encourages users to keep going rather than stop. Managers must retain the instinct to ask: “What assumption is missing?” and “What evidence would invalidate this recommendation?”
This vigilance must be collective. Employees often know the exceptions that procedures describe poorly: a vulnerable client, outdated data or a verbal commitment awaiting confirmation. Organizing their feedback prevents automation from turning these blind spots into routine practice. Authorized instructions must also be distinguished from the content consulted: a web page or document may contain malicious instructions intended to hijack the agent.
What now? Looking ahead to September 2026, the best-prepared organizations may be those that have learned to allocate autonomy judiciously rather than maximize it. A practical first step: choose a limited task, write its mandate, test its failure modes and designate who makes the final call. Managing AI agents does not mean monitoring every click. It means creating the conditions to delegate execution while retaining control over decisions.
