Skip to content

AI News · Edition 74 ·

The AI agent your staff will try before you've decided on it

Anthropic's Claude Duet is aimed at everyday users, so it will reach your office first. Plus: what a failed antivenom test says about AI hype.

8 stories · sourced from 24/7 Wall St., The Verge AI, OpenAI, TheStreet and others

Abstract illustration accompanying The AI agent your staff will try before you've decided on 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 consumers that it says will take work off your plate

24/7 Wall St. ↗

What happened

Anthropic, the company behind the Claude assistant, has launched Claude Duet. Anthropic calls it an AI agent. It is aimed at ordinary consumers, not company IT departments, and it puts Anthropic up against Meta and OpenAI for everyday users. Early coverage describes it as a tool meant to lighten your workload. More than 50 outlets carried the story, so it is getting a lot of attention. That says nothing about how well it works. The coverage is also short on specifics. It isn't clear from what has been published which tasks it handles, what it costs or how it will be offered in India and Canada.

The word that matters is “agent”. A chatbot answers when you ask it something. An agent is meant to do a task for you. It works through several steps, uses other apps and comes back with the job done instead of a paragraph of advice. Every large AI company is moving this way, and Duet is Anthropic's version for individuals.

What it means for your business

The consumers this is aimed at include your staff. A product like this usually enters a small business through someone's phone, not through a purchase you approved. Picture the office administrator at a 600-student school, the front-desk coordinator at a 30-bed clinic or the purchasing clerk at a 40-person workshop. They sign up on their own because it promises to clear their inbox faster. If it saves them an hour a day, that's useful. It becomes a problem if they give it parent contact lists, patient appointment details or supplier price sheets and nobody has decided whether that's allowed.

Where an agent could really help a small business is the dull work. That means drafting replies to routine enquiries, copying figures from invoices into a spreadsheet, sending fee reminders and turning scattered notes into a weekly summary. If Duet does that kind of multi-step chore reliably, the value is real but modest. Expect a few hours saved per person each week, not a reorganised company.

What it does NOT mean

It does not mean you can replace your office manager, receptionist or accounts clerk this year. AI agents in general still make mistakes, and those mistakes are harder to catch because the work looks finished. A wrong figure in a fee ledger or an appointment booked for the wrong day can cost more to fix than the time it saved. Anyone who tells you a consumer agent can “run your back office” is usually selling something, such as a course, a consulting package or a reseller subscription.

Be careful with the investment coverage too. Anthropic's IPO is in the news, and much of the excitement around Duet is really excitement about the stock. That tells you nothing about whether it will get your clinic's bookings right.

What a sensible owner should do this month

  • Ask before you ban. Find out which staff already use ChatGPT, Claude, Meta AI or similar tools on personal accounts for work. In most offices it's more people than the manager thinks.
  • Write one page of rules. List what can go into a personal AI tool, such as public information and generic letter drafts. List what never can, such as student records, patient details, salaries and customer bank or card details. Canada's privacy laws and India's Digital Personal Data Protection Act both make the business responsible for the personal data it collects. Your employee's choice of app doesn't change that.
  • Run one small test. Pick one person and one repetitive task, such as summarising the week's enquiries, and try it for a month. Duet's price isn't clear yet. Paid consumer AI plans have mostly cost around US$20 a month, with cheaper options in India. So the bet is small: one subscription, one task, one month.
  • Check every result during the trial. Have a person review everything before it goes to a parent, patient or customer. If checking takes as long as doing the task yourself, you have your answer.

If you do only one thing, write the one-page rule. An agent nobody approved, holding your customers' data, is a bigger risk than one you chose to test.

Also worth knowing

  1. An unreleased Claude was asked to design a scorpion antivenom. Experts say it didn't come close

    An investment research firm wanted to prove AI could transform drug development. It teamed up with Anthropic to see whether an unreleased version of Claude could make a scorpion antivenom. Experts say the model offered some promising starting points but nothing close to an actual drug. Remember this the next time a vendor says AI 'discovered' something: the honest version is usually that it gave a skilled person a head start.

    The Verge AI ↗

  2. Uber gives OpenAI's ChatGPT Work to more than 30,000 employees

    Uber is rolling out ChatGPT Work to more than 30,000 staff for research, data analysis, presentations and multi-step tasks. The announcement comes from OpenAI itself, so it reads like a success story. One reason large firms pay for company-wide AI accounts instead of tolerating personal ones is control over where business data goes. That reasoning holds just as well for a 20-person office.

    OpenAI ↗

  3. Report: Volkswagen workers may get Microsoft AI tools instead of a pay rise

    TheStreet reports that Volkswagen workers may get Microsoft AI tools in place of a raise. The coverage available is thin, so treat the details as unsettled. Whatever VW decides, every employer can take a lesson from it: staff don't see a software licence as pay, and presenting it that way breeds resentment fast.

    TheStreet ↗

  4. Anthropic and music publishers fight in court over whether AI gets a 'fair use' defence

    Anthropic and music publishers are arguing in court over whether Anthropic can rely on a 'fair use' defence in a copyright case about AI. Nothing here changes how you use AI tools today, but the question of whose creative work these systems can learn from is still open. Until it's settled, be careful about publishing AI-generated jingles, lyrics or marketing material that closely imitates someone else's work.

    Reuters ↗

  5. OpenShell, a tool for keeping AI agents fenced in, draws strong interest from developers

    OpenShell calls itself a safe, private place to run large numbers of AI agents, and it drew a lot of interest from developers on Hacker News. You won't use it directly. What matters is that the people building agents are working hard to limit what those agents can reach. Take that as a guide to how much access you should give one inside your own business.

    Hacker News ↗

  6. OpenAI releases a 'Decisions API' for quick, low-cost automated choices inside apps

    OpenAI has released a Decisions API, pitched as fast, focused, low-cost decisions for any application. The material doesn't say much beyond that, and it's a tool for software developers, not something an owner would use directly. If you pay for software with AI features, it's fair to ask your vendor whether cheaper building blocks like this will lower your price.

    OpenAI ↗

  7. IBM explains how to combine its Granite models for better forecasting

    IBM published a technical guide on combining several of its Granite time-series models to produce forecasts. Forecasting stock, sales or staffing is one of the more practical uses of AI for shops and factories. This guide is written for engineers, though. Take it as a sign that forecasting tools are getting cheaper to build, not as something to act on yourself.

    Hugging Face ↗

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.