Microsoft Puts Its AI Agents to Work on Their Own: What Limits to Set Before Doing the Same in Your SME
On September 25, 2026, Microsoft unveiled a new version of Copilot that brings chat, natural-language tool building and — most notably — agents that work on their own into a single app. According to Infobae and Fortune, the app is organized into three sections: Home, Code and Autopilot. For an SME, the news is not Microsoft's product itself but the direction it confirms: AI stops waiting for instructions and starts acting.
What Microsoft announced
Home brings together chat and Cowork, with recent activity and the ability to pick up documents and conversations. Code lets people with no technical background describe apps, dashboards and workflows in natural language; Jared Spataro, Microsoft's AI at Work marketing lead, described them as "small solutions designed for a specific purpose that anyone can build." And Autopilot works as a digital teammate with its own identity, memory and capabilities: after an initial instruction it can monitor channels, follow up on projects, watch deadlines, remind colleagues of commitments and keep working in the cloud even when the user is not present.
Annie Pearl, Copilot's corporate vice president, summed it up to Fortune: "We do want this to become an all-in-one productivity suite."
When it arrives and how it is billed
It is not available to everyone. According to Fortune, most features are rolling out through an early test program with no general-availability date; Code will reach Microsoft 365 Premium and Pro subscribers later in 2026, and Autopilot remains in restricted testing, with a broader rollout depending on tester feedback.
The biggest budget change is in billing. On top of the per-user license, which covers chat and integration with Office, Outlook and Teams, Microsoft adds usage-based billing for Code, Autopilot, long-running tasks and the most advanced models. To keep it in check, Infobae describes admin tools with spending limits, group policies, approval flows and the ability to decide which models each team uses.
What it means for an Ecuadorian SME
An agent that works on its own is useful precisely because you do not have to watch it. That same autonomy forces three decisions that a chat never required:
- What it can do alone and what needs approval. Reminding someone of a deadline or drafting a reply is not the same as writing to a customer, accepting an order or moving money.
- Which data it can access. An agent that reads emails, chats and documents handles customer and employee information. In Ecuador that processing falls under the LOPDP and, since February, under the Personal Data Protection Superintendency's rule on the use of AI systems, according to LexLatin.
- How much it can spend. With usage-based billing, cost is no longer fixed. A poorly configured agent running all day can consume more than planned; the spending cap is set before switching it on, not after the invoice arrives.
Before letting an agent loose: a short checklist
- Start with a single process that is repetitive and measurable: quote follow-ups, appointment reminders, answers to frequent questions.
- Write its limits in one sentence: "it can do X; for Y it must ask Z for approval."
- Give it access only to what it needs, not the whole shared folder or the entire mailbox.
- Assign a human owner who reviews what it does, at least during the first week.
- Set the spending cap and review actual consumption after a month.
How We Approach It at SimCodec
At SimCodec we build voice, WhatsApp and Telegram AI agents that answer, schedule and quote on their own, always with those limits defined from day one: what they resolve alone, when they hand the conversation to a person, which data they access and what they cost to run. We connect them to your business's real systems — inventory, ERP, calendar, invoicing — and they do not depend on any particular suite.
If you are considering putting an AI agent to work in your company, let's talk or call our AI assistant Cyntia at +593 99 726 6838. We help you choose the first process and define its rules before switching it on.
