Illustration: Brendan Lynch/Axios. Stock: Getty Images

In 30 seconds

• Muse could bring smaller stores into the picture more often. A lot depends on how the agent makes its choices.

• Voice gets interesting when it makes whole tasks easier and leaves less work unfinished afterwards.

• Running their own AI infrastructure makes sense for some companies. Smaller teams should also have a workable alternative to their current provider.

1. Meta Muse is getting involved in shopping and small businesses

Illustration: https://about.fb.com/news/2026/09/introducing-muse-small-business/amp/

Meta is expanding Muse quite quickly. At the end of September, it added features for small businesses, including connections to Shopify, Slack and accounting software. Muse is initially available in the US and Canada. According to Meta, publishing, sending messages and spending money require the user’s approval.

For shopping, Muse can already handle payments in the US through Stripe’s Link after the user approves the amount. Amazon has blocked Muse’s access.

This could get interesting for smaller stores. Many people order from Amazon because they know how it works. Their address is saved, delivery is usually fast, and returns are straightforward. You probably do not want to click through five unfamiliar stores for a charging cable.

If Muse handles that search and finds a suitable product at a smaller retailer, some of that convenience advantage disappears. For a purchase like this, we would probably care very little about which store sells the cable, as long as the quality, delivery and returns are fine.

Price would probably carry more weight. When offers are similar, an agent might look for the cheaper one more consistently than we would for a few euros. Someone who needs the cable tomorrow will still pay for fast delivery. Amazon’s delivery network remains an advantage, even when the search starts somewhere else.

A small store would need to present its products, delivery times and return conditions clearly. That does not sound particularly exciting. We would still spend time on it before trying the next trick that supposedly gets a business into every AI answer.

Muse for Small Business adds another side to this. Someone running a store often also handles advertising, customer questions, invoices and the website. You can be good at your actual business and still spend a lot of time on things you never wanted to do.

An agent that summarises sales figures, prepares a campaign or sorts open questions could help. We can easily imagine more people trying to run a small business because of this. A side project might also become possible when there simply was not enough time before. Finding customers and developing an offer that someone wants to buy is still plenty of work.

We would keep an eye on how much of that ends up with Meta, though. It feels a little strange when the company you buy advertising from also prepares your campaigns and selects products on the buyer’s side. We would want to know what role paid placements and partnerships play in those suggestions.

OpenAI has now joined in with dots. The agents introduced at DevDay can work on longer-term tasks across connected applications. Grok Bot also uses Stripe’s payment infrastructure for shopping. These companies are trying to become a regular part of everyday life.

Meta’s audio and camera glasses fit into that. Muse is due to come to its AI glasses in the coming months. Asking an agent to quickly take care of something will probably become a habit more easily if you do not have to take out your phone every time.

Sources: Meta: Muse for Small Business · Stripe: Muse payments with Link · Axios: Amazon blocks Muse · OpenAI: Introducing dots · Stripe: Muse payments with Link · Meta: Muse and AI glasses at Connect 2026

2. Voice lets you work through a conversation

Illustration: OpenAI

ChatGPT Voice can start tasks, check their progress and pass on further instructions. What exactly works depends on connected tools, permissions and availability in the account.

With a suitable email connection, you could have important messages read out on your way to a client. One needs some information, so you briefly explain what the reply should say and have a draft prepared. You do not have to search your inbox first and copy the whole conversation into a prompt.

The same idea could work in software development. Ask for tests to be run, find out why one failed and have a change prepared. You can review the code and its release on screen afterwards.

We think this could also help people who have not found much use for AI so far. You can simply start talking, add something or explain that you meant it differently. That way of describing a task is more familiar to many people than an empty text box waiting for the right instruction.

It would be a shame, though, if we mainly used this to fit more work into the journey to a client or an afternoon walk. We would get more out of it if there were actually fewer unfinished tasks afterwards, leaving us free to do other things.

DevDay also brought Space and Living Pages. Teams can work together there and have documents updated from connected sources.

That would be quite useful for a project overview. A deadline moves, a task is finished, a decision from the client is still missing. That information often already exists somewhere. Someone still has to collect it for the next update. This would be one of the first tasks we would try with it. It is fairly easy to check afterwards whether something is missing or has been misunderstood.

Sources: OpenAI: DevDay 2026 – Space and Living Pages · OpenAI: ChatGPT Voice features

3. Running your own AI is also about dependence

Illustration: https://www.nvidia.com/en-sg/data-center/dgx-b200/

Latham & Watkins runs its own Nvidia servers and adapts open models. The law firm also continues to use external AI services. It can use different systems depending on the task.

That is a lot for a small business to take on. Buying hardware, finding people to run it, installing updates and managing security. A large law firm finding this worthwhile does not mean a team of twenty needs to take the same route.

The question of dependence still applies. If a provider raises its prices or changes an important feature, you want at least one workable alternative. That becomes difficult when important information and workflows gradually come to depend on its system.

OpenAI introduced additional options for confidential processing through Private Intelligence. Private Inference is initially a preview. We would look at offers like these before buying our own hardware. What matters is whether they actually meet the requirements for data and access.

In a project, this could mean keeping documents in a usable format outside the AI tool and recording how important workflows operate. When choosing a provider, you could also try exporting your data to see how much of it you actually get back.

For most smaller teams, a mix seems sensible to us. Cloud services for tasks they handle well, and separate solutions where there are specific reasons for them. We would consider running our own infrastructure when sensitive data, particular requirements or actual usage justify the extra work.

Sources: Finance Gazette: Latham & Watkins’ in-house AI · OpenAI: DevDay 2026 – Private Intelligence

One question for you

Have you already let an AI agent buy something for you? How did it go?