AI Recap: The week in which agents log in without a password and Mistral builds voices

1Password for Agents: Agents log in without ever seeing the password

On October 22, 1Password launched the platform 1Password for Agents. Personal AI agents can use it to access accounts without ever seeing the password. For the launch, 1Password names Muse from Meta, Claude Duet from Anthropic, ChatGPT Work and dots from OpenAI, Grok Bot from SpaceXSI, and Townie from Town. Muse and Grok Bot are set to roll out in the coming days, and the other integrations are announced as coming soon.

For you, this means: if you let agents access customer or supplier accounts, you get a clean separation between credentials and model. However, this does not replace rules for what the agent may do after logging in. A password vault does not prevent prompt injection, meaning an agent being redirected by manipulated content. If you already connect agents to your systems, limit their permissions per task. You'll find more in our security guide for SMBs.

Microsoft brings AI into the contact center without replacing the platform

On October 21, Microsoft launched Contact Center Intelligence in public preview. The solution is designed to connect with existing contact centers through Microsoft AI and agents, without replacing everything. It offers voice and digital self-service, conversation analysis, and actions using business data.

This is meant for businesses that already run a contact center platform such as Five9, Cisco Webex Contact Center, or Avaya. A regular office phone system is not what this addresses. The preview is available worldwide except in Austria. Billing is usage-based through Copilot Credits, and the contact center platform provider may charge separately for API or voice access. But a preview means features and prices can still change, so test it in a pilot first and don't switch your entire reception at once. To compare it with standalone phone assistants, see our guide to AI phone assistants.

OpenAI makes the Decisions API generally available, decisions cost only input

OpenAI's Decisions API doesn't return free text. Instead, it returns a choice from fixed answers, a probability, or a score. According to trade media, it has been generally available since October 22. On the GPT-6 Luna model, input costs $0.10 per million tokens, and output is not charged. According to OpenAI, the endpoint responds up to ten times faster than the Responses API.

For you, this suits ticket routing, lead scoring, or input moderation: short decisions, clearly limited answers, no text fees. Still, do the math, because costs depend on the amount of input tokens, and long inputs become correspondingly more expensive. For more on how agent building blocks fit into the picture, see our article AI Agents for SMBs.

Mistral follows with Voxtral 3: voice agents with EU hosting or your own server

On October 20, Mistral AI introduced Voxtral 3, a family of speech models with an SDK for voice agents. It transcribes in more than 50 languages, generates speech in over 35 languages, and calls external tools during a conversation. It can be run through EU-hosted interfaces or on your own infrastructure.

For practices, trade businesses, and hotlines, this means speech technology is getting cheaper and can be self-hosted if conversation data must stay in-house. The effort then lies in integration and testing with your scheduling system. For more, see our AI phone assistant guide.

Hamburg becomes the AI supervisor for public authorities, for you the date is December 2

The Hamburg Data Protection Authority will become the market surveillance authority for AI use in Hamburg's public authorities starting January 1, 2027. The city parliament passed the legal amendment on October 14, 2026, and the authority informed the public on October 21. This initially affects public administration and does not directly concern you.

The date that does concern you is set, however: starting December 2, 2026, a new ban applies to AI systems that generate non-consensual intimate images or AI-generated depictions of child sexual abuse. On the same day, the transition period ends for machine-readable labeling of AI content for systems that were already on the market before. Check now which of your AI tools generate content that goes external. This is not legal advice, but a list by December is no luxury.

Google packages Gemini by industry, the first version is for government agencies

On October 20, Google Public Sector introduced Gemini for government agencies, an agentic package for administrative tasks. It launched initially as a private preview for federal, state, and municipal agencies in the US. Google explicitly describes it as part of a series: packages for legal advice and financial services have already been released.

For you in Germany, this product is not yet relevant. A private preview is not an open test, and Google has not yet announced pricing or a date for general availability. The direction is what matters: providers are no longer selling AI as a general-purpose chat, but as ready-made packages for specific industries. If you work in law firms, tax, or financial offices, pay close attention to such templates when choosing your next provider.

What stays with you from this week in day-to-day business

First: agents need their own access rights. If you let an agent access an account, define approvals per task and document what the agent is allowed to do. A password vault alone does not make an agent secure.

Second: phone and voice are becoming the standard interface. With Voxtral 3 and the Microsoft preview, there are now two ways to test voice agents, either with EU hosting or within an existing contact center platform, though the Microsoft preview is not available in Austria. Compare prices per minute and pilot periods before you sign a contract.

Third: the cost of simple decisions is falling. Classification and routing become cheap through input tokens, but the price depends on the length of your inputs. And two hard dates are on the calendar: the new ban and the end of the labeling transition period take effect on December 2, 2026.