MONTHLY BRIEFING · MARKETING · SALES · CUSTOMER SERVICE
From prompts to proactive, always-on agents.
The month in AI, washed down to what matters for marketing, sales and customer teams.
Hello, fellow laundry inspectors.
October gave us agents that work in the background, advertising inside AI conversations, a shopping cart that follows customers across stores, and design tools that promise to handle the edits after the first draft.
Keeping up with the names was a task of its own.
But underneath the announcements, a useful pattern emerged: AI providers are competing to take responsibility for larger pieces of work.
For businesses, that raises a practical question. Which of those pieces are ready to hand over—and how will we know the work was done properly?
Here is what changed in October, what remains a promise, and what deserves your attention.
From waiting for a prompt to watching for the next move.
For most people, working with AI has followed a familiar rhythm: write a prompt, get a response or an action, then start again. The major providers are now competing to move beyond that reactive model, towards agents that monitor what changes and initiate work without waiting for the next instruction.
OpenAI introduced Dots on 29 September: always-on agents with their own cloud computers, designed to keep working between conversations. That launch belongs to September, but it provides the starting point for October’s competition.
On 8 October, Google announced its Gemini agent, describing a system that plans work, connects to business tools and brings back completed results. On 21 October, Anthropic introduced Claude Duet, a proactive agent that checks connected applications, prepares work and continues when the user steps away.
Duet’s initial research preview opened through a waitlist for US Pro and Max subscribers. Anthropic says it seeks approval before actions such as sending messages, making payments or deleting information.
The attraction is easy to understand. A change in an account can trigger research. An approaching meeting can trigger preparation. A stalled opportunity can trigger a review of what should happen next. Someone no longer has to remember to start every piece of work.
At AiDU, we have been building proactive systems for more than a year. That hands-on work has taught us where they hold up, what makes them useful in day-to-day operations, and where they still need work.
One lesson keeps coming back: proactive does not mean an agent remembers everything. Starting work on its own does not guarantee that it has the latest customer context, understands a commitment made elsewhere, or knows what another system has already done.
Each connected action needs a harness around it: the context to make a sound decision, clear permissions, checks before and after execution, and a way to recover or involve a person when something goes wrong. The system also needs to record what happened so the next action starts from an accurate picture.
Consider a sales follow-up. An agent may notice that a prospect has gone quiet. Before it sends anything, it needs to check whether a colleague has already spoken to them, whether a service issue is unresolved, and whether the proposed next step still fits the deal. Afterwards, it needs to update the shared record and track the response.
That is the work around the action that helps revenue move. Proactivity gives a system the initiative. Context, coordination and reliable execution make that initiative useful.
Sources: OpenAI Dots, Google’s Gemini agent, Claude Duet
Advertising follows the conversation.
On 28 October, Google announced the expansion of ads in AI Mode to ten additional countries, including Canada, India and Australia. It also announced ads in follow-up queries in the US.
That means a commercial opportunity can appear as someone refines a question, compares alternatives or explains what they need.
For advertisers, this adds another setting in which their offer must make sense. A person asking an exploratory question may need different information from someone who has already chosen a product.
The geographic distinction matters. Belgium was not among the ten additional markets announced, and the UK was listed for a later expansion during the fourth quarter.
Google says its internal research shows similar click-through rates for sponsored and organic product-list items within AI Mode. That is a vendor finding about clicks, not evidence that these placements will produce equivalent sales for every advertiser.
Watch how conversational discovery changes customer intent. Judge the value through qualified enquiries and purchases.
The shopping cart moves across stores.
Google’s Universal Cart began rolling out in the US on 15 October, following its earlier preview.
Customers can collect products from different retailers across Search and Gemini, keep track of price drops and stock changes, and use supported checkout options. They can complete purchases with Google Pay or transfer items to a merchant’s cart. The retailer remains the seller of record.
This brings more of the shopping journey into Google’s environment.
A customer may compare your product, save it and return to buy without following the familiar sequence of landing page, category page and product page.
For retailers, the implication is practical: product information needs to be accurate wherever customers encounter it. Prices, stock, variants and delivery details help determine whether an offer is useful.
Treat the product feed as part of the customer experience. A persuasive website cannot correct information a shopper encounters elsewhere before visiting it.
Sources: Google’s Universal Cart launch
Canva works on what happens after the first image.
Canva introduced its image-generation model, Craft, on 22 October. A week later, it introduced Brush, a model designed to understand and edit designs.
Together, the announcements address two familiar stages of creative production: making something and revising it until it is usable.
Canva says Craft plans composition before rendering an image. Brush can make changes within editable designs, including updates across multiple pages and adaptations for different formats.
Brush initially powers supported design edits for Canva Pro subscribers. Canva lists Brand Kit support among the capabilities still being developed.
For marketing teams, revision is where much of the work lives. Change the offer. Correct the headline. Adapt the layout. Preserve everything that was already approved.
Test these tools on a real revision brief. Check whether they make the requested change while preserving the details that need to stay consistent.
Sources: Canva Craft, Canva Brush
Voice agents get another serious option.
Mistral released Voxtral 3 on 20 October, combining new speech models with a toolkit for building voice agents.
The announcement covers transcription, speech generation and a real-time process connecting listening, reasoning, tool use and spoken responses. Mistral also highlights multilingual capabilities and controls for pronunciation and delivery.
For customer teams, the interesting possibilities include handling enquiries, gathering information and connecting conversations to business systems.
But a natural voice is only one part of a useful service.
Can it capture an unfamiliar surname? Repeat a reference number correctly? Handle an interruption? Recognise when a request needs a person?
Those details determine whether customers experience help or friction. Evaluate voice agents with the awkward calls your team actually receives, including local languages, background noise and incomplete explanations.
Sources: Mistral Voxtral 3
CRM work comes closer to the conversation.
On 8 October, Microsoft announced 30 Dynamics 365 skills spanning sales, customer service and customer insights, alongside CRM experiences in Teams and updates to its sales and service agents.
The capabilities include preparing account plans, identifying at-risk deals, proposing record updates and drafting customer responses.
Availability varies. The skills are generally available in Cowork, while several other experiences remain in public or private preview.
The direction is significant: customer information and actions become accessible in the places where teams already communicate.
That could reduce the effort of switching systems and reconstructing context. It also makes the quality of that context more consequential. An agent working from an outdated opportunity record can move quickly in the wrong direction.
Better access to CRM data increases the value of keeping it accurate—and knowing who owns the next action.
Capability and access moved at different speeds.
October brought several developments worth separating:
- 7 October: Anthropic launched Claude Haiku 5.5, positioning it for high-volume, cost-sensitive work.
- 13 October: Google expanded Gemini 4 Argon access to developers and eligible paid Gemini users.
- 26 October: OpenAI released GPT-6.1 Terra, describing strengths in visual reasoning, computer use and writing.
- 28 October: Google announced Gemini 4 Carbon, with initial access limited to selected cyber defenders and trusted partners.
These releases do not all belong in the same purchasing decision. Some concern efficiency, others demanding tasks, and some are not yet broadly accessible.
Compare models on your own work. Measure correctness, completion time and the amount of human correction required.
Sources: Claude release notes, Argon availability, OpenAI changelog, Gemini 4 Carbon
Europe’s open-model story advances.
Mistral previewed Large 4 early in October and released its weights on 29 October under the name ‘Le Chonk’.
For organisations considering deployment choices, downloadable weights create possibilities for operating and adapting a model outside a provider’s hosted service.
Those possibilities come with work: infrastructure, maintenance, security and evaluation still need owners. Open weights do not automatically make a system inexpensive or straightforward to run.
Deployment control is a business choice. Assess it alongside capability, operating cost and the expertise available to maintain the system.
Sources: Mistral’s open-weight release
Railway supplies the operational reality check.
Railway disclosed a security incident on 21 October involving stored customer credentials.
Its preliminary report describes interrupted builds and deployments, potentially exposed environment variables and credential-rotation measures. Railway says existing customer deployments remained online.
The distinction matters: this was more specific than a blanket claim that every hosted application went down.
It also illustrates the dependencies behind an AI workflow. Models connect to applications through credentials. Applications run on infrastructure. A problem in either can affect the work.
Know where credentials are stored, what access they grant and how the team would restore service after an incident.
Worlds you can steer.
Google announced Genie 4 on 27 October, describing research into interactive generated worlds with native audio.
Users can steer scenes through natural language, while the system generates visuals and sound in real time. Initial access is limited to trusted testers.
Google points towards applications in creative industries and the training of robots and agents. These are prospective uses, rather than proof of production readiness.
It earns a place in this month’s wash because it widens the picture: AI development also includes environments people can explore and interact with.
Sources: Google’s Genie 4 announcement
A quick check of the subscription bill.
Anthropic announced monthly API credits for Max and Team plans on 7 October. OpenAI followed on 28 October with API credits for eligible Pro and Business subscribers.
The amounts and conditions differ. These benefits can make experimentation easier, but the useful business measure remains the cost of completed work—including usage, retries and review.
Sources: Anthropic release notes, OpenAI changelog
What does a good result look like?
October expanded the work AI products are designed to take on: monitoring accounts, preparing campaigns, handling conversations, editing designs and helping customers buy.
The opportunity is becoming more concrete. So is the need to decide what a good result looks like.
For your next experiment, choose a recurring task with a visible outcome. Give the system the context and access it needs. Then check whether it completes the work accurately, within its permissions, and with less effort from the team.
That will tell you more than another impressive demo.
What should go into November’s wash?
Seen a launch worth inspecting, a claim that needs washing, or a question we should investigate? Send it our way for November’s AI Laundry.
Add your question