Skip to content

AI News · Edition 73 ·

Claude Duet joins the AI agent race. Test it before you trust it

Anthropic's new agent targets small business owners. What it changes, what it doesn't — plus a PHP security flaw worth a call to your web vendor.

8 stories · sourced from The Verge AI, Hacker News, OpenAI, Hugging Face and others

Abstract illustration accompanying Claude Duet joins the AI agent race. Test it before you trust it
Abstract illustration, generated with AI. It represents the idea, not the event.

The one that matters

Anthropic launches Claude Duet, an AI agent for small business owners — but it's an early test version

The Verge AI ↗

What happened

Anthropic, the company behind the Claude AI assistant, has announced Claude Duet. It is an AI “agent” that Anthropic says is designed for professionals and small business owners. It is launching as a “research preview”, which is the maker's own way of saying it is not finished. You get it as a standalone app on iOS and Android and through Claude. Anthropic is pitching it directly against Meta's Muse and OpenAI's Dots, which are agents from two other big AI companies. More than two dozen outlets covered the launch. One headline warns that most people probably can't try it yet, so expect limited access at first.

In plain terms, an agent is AI that is meant to do a task for you instead of just answering a question. It works through several steps, keeps track of where it is and hands back something finished. The reports so far don't say which tasks Duet handles well, which of your other tools it connects to, or what it will cost. Be wary of confident claims about any of those until you have seen it yourself.

What it means for your business

The news is less about one app than about who it is aimed at. Three of the best-known AI companies are now competing with agent products, and this one is aimed squarely at small business owners, not corporations with IT departments. Over the coming months, expect more sales pitches aimed at you and frequent changes in price and features. Expect more AI that runs on your phone, not in a system someone has to install for you.

Agents are most likely to earn their keep on the dull, repetitive, multi-step work that eats a manager's week. At a 600-student school, that could be the office turning five teachers' notes into a weekly parent update. At a 30-bed clinic, it could be the front desk turning a messy shift roster into a clean schedule. At a 40-person workshop, it could be the owner comparing three suppliers' quotations line by line. If a coordinator spends eight hours a week on that kind of work, that is about a fifth of their time. That comes to roughly ₹6,000 a month on a ₹30,000 salary, or about CAD $750 a month on a $45,000-a-year one. Even partial help with that work is worth real money. Nothing grander is likely to pay off in the first year.

What it does not mean

  • It is not a new employee. Expect consultants and resellers to start selling “AI staff” on the back of launches like this. A research preview is still being tested. Before anything it produces goes to a parent, patient, customer or bank, a person needs to check it.
  • It is not a race you need to bet on. The coverage frames this as Anthropic against Meta and OpenAI. That contest matters to them. For you, it mostly means the products will keep changing for a while. That is a reason to stick to short trials and avoid long contracts.
  • It is not cleared for sensitive data. Nothing reported so far says whether it is suitable for student records, patient files or staff details. Under India's DPDP Act, and Canada's PIPEDA and provincial health-privacy laws, that decision is yours, not the vendor's.

Anyone telling you this launch means you can hand your back office to software this quarter is getting ahead of the facts. Even Anthropic is calling it a preview.

What a sensible owner should do this month

  • Write down three tasks. List the three repetitive jobs that take up the most staff hours, with a rough weekly count for each. This list will be useful whichever agent you end up using, whether that's Duet, Muse, Dots or a feature added to software you already have.
  • If you can get access, run one small test. Give one curious staff member four weeks, with a limit of two hours a week. Pick a low-stakes task that involves no personal data, such as drafting the parent newsletter, summarising supplier quotes or turning meeting notes into action points.
  • Keep an honest record. Write down the minutes saved and the mistakes you had to fix. If after four weeks it isn't saving more time than it takes to check its work, stop. That is a perfectly good result, and it cost you eight hours.
  • Sign nothing long. If a vendor pitches a service built on agents, ask for monthly billing. Ask them to say plainly where your data is stored, and how you can export all of it if you leave.

If you can't get access yet, waiting costs you nothing. The list of tasks is what matters, and you can write it today with a pen.

Also worth knowing

  1. AI helped uncover a critical flaw in PHP, the language behind many small business websites

    AI reportedly helped find a flaw nicknamed PHP2Shell, described as a critical “remote code execution” bug in PHP. A bug like that can let an outsider run their own commands on a web server. PHP runs a large share of small business websites, including every WordPress site, so if you have a website, booking form or online shop, ask whoever maintains it this week whether you are affected and when it will be patched. Little is public yet, so don't panic, but do make the call.

    Hacker News ↗

  2. ChatGPT ads expand to 26 more countries, with over 100,000 advertisers so far

    OpenAI says ads inside ChatGPT are now available in 26 more countries. It says more than 100,000 advertisers, from local businesses to global brands, have used them. The pitch is that you reach people while they are comparing options and deciding what to buy. Check whether your country is on the list before an agency pitches it to you, and treat it as a small, measured experiment rather than a replacement for search or social ads you already know work.

    OpenAI ↗

  3. Granite Speech 5.0 promises faster, more accurate transcription and translation

    Granite Speech 5.0 is a new speech model that is said to transcribe, translate and attribute speech more accurately and faster. Transcription is one of the most practical uses of AI, such as a clinic doctor's dictation or the minutes of a school staff meeting. The release doesn't say which languages it handles well, which matters most for multilingual teams in India and Canada. If a vendor sells you transcription, ask which model it uses and test it on your own staff's accents before paying.

    Hugging Face ↗

  4. Elite brings its law-firm financial data into Claude and Microsoft Copilot

    Elite is extending its “Elite Intelligence” features into Claude and Microsoft Copilot, so law firms can use their financial information inside AI tools they already have. If you run a law practice, ask your vendor about it. If you don't, notice the pattern: software you already pay for is starting to connect to AI assistants, and that is often a simpler route than buying a separate AI product.

    Yahoo Finance ↗

  5. Music labels clash with Anthropic's Claude over copyrighted song lyrics

    Courthouse News, which covers legal disputes, reports that music labels are clashing with Anthropic over Claude and copyrighted song lyrics. Few details are available, but the lesson for any agency, school or shop producing content is simple: AI output can contain other people's copyrighted work, and if you publish it, the risk is yours. Don't ask AI to reproduce lyrics, long quotations or anything else you don't have the rights to.

    Courthouse News ↗

  6. OpenAI's models are coming to Google Cloud

    OpenAI and Google Cloud say OpenAI's most advanced models and its Codex coding tools will be available through Gemini's Model Garden and the Google Cloud Marketplace. This only matters if your business or your software vendor already runs on Google Cloud, because it may make OpenAI models easier to use under the billing and contracts you already have. Otherwise, there is nothing to do.

    OpenAI ↗

  7. Mistral Large 4 is out, and most businesses can safely ignore it for now

    The European AI company Mistral has released Mistral Large 4, and it drew a lot of attention from developers. Little has been said about what it does better, and nothing here changes how a school, clinic or shop should work. Its relevance to you is indirect: more serious competitors tend to keep AI prices in check.

    Hacker News ↗

How this briefing is put together

Every morning we read the day’s AI announcements and reporting from the companies themselves and from the technology press, then pick the handful that actually change something for a working business. The analysis is ours and it is written for owners and managers, not engineers. Every story links to its original source — read them, and disagree with us where we’ve got it wrong.