AI agents are starting to run whole jobs. Five ideas worth watching.

Most weeks the AI news does not mean much if you run a small business. This week was different.

Anthropic, Meta, OpenAI and Microsoft all moved in the same direction at once. AI is shifting from doing one task when you ask it, to running a whole job while you are off doing something else. That is a bigger change than another chatbot.

It is also further along than most people realise. An SBE Council survey found 58% of small employers already using AI agents regularly or occasionally, with another 14% weighing it up. If that number feels high, remember how many of these tools are now folded into software people already pay for. Plenty of owners are using agents without ever calling them that.

Here are the five ideas we think are worth watching, and what we would actually do about them.

1. Agents that work while you get on with your day

Anthropic has released an early version of Claude Duet, which works away on its own machine in the background instead of sitting there waiting for you to type. It joins Meta's Muse and OpenAI's Dots.

The restriction matters more than the feature. Duet cannot send a message, move money or delete anything until you say yes.

One caveat before you get your hopes up. This is a research preview on a waitlist, open to Claude Pro and Max subscribers in the US. So for a business in Dorset the useful part today is the shape of it rather than the tool itself. Work gets prepared overnight. You approve it in the morning over a coffee, and nothing leaves the building until you do.

2. One agent running the whole job, not one step of it

Most owners still use AI to write one email or summarise one document. Useful, but nowhere near the whole job.

Some have gone much further. Scott Bell, a bankruptcy lawyer in the US, told Entrepreneur that his agents now take new client enquiries, answer the routine ones, get financial paperwork into order, pick out the urgent legal issues and put together first-draft case summaries. They have dealt with clients directly. He describes them getting through "hours of administrative work in minutes".

The interesting part is not the time saved. It is what happened to his own job. He stopped doing the work and started supervising it, signing off decisions made by a system he could not fully see inside. That is a real trade, and worth thinking about before you make it rather than after.

3. One agent plugged into all your apps

Meta's Muse for Small Business joins up tools plenty of small firms already pay for: Shopify, Stripe, QuickBooks, Canva, Slack and Instagram.

Once connected it can draw a year's growth plan out of your sales and advertising figures, work through the inbox when it gets away from you, look over the accounts for expenses that do not belong, and have your next campaign standing by.

The same rule applies as with Duet. Nothing gets posted, sent or paid for without the owner agreeing to it first.

4. AI handling the first reply, every time

The same Entrepreneur piece describes owners putting agents on the front line, fielding customer enquiries, handling refunds and passing the awkward cases to a person. Others point them at marketing instead, drafting campaigns and queueing posts up to go out.

The owner only picks up the conversations that genuinely need a human. If you have ever lost an evening to the same five questions you answered last week, start here.

5. The more freedom agents get, the more the rules matter

This is the one we would read before any of the others.

Microsoft's new Guardian Agents watch other AI agents while they work and narrow what they are permitted to do. An agent might be left able to read your files but not alter them, or required to get a person's approval before anything consequential. Microsoft's own framing is striking. It is telling companies to treat their agents as potential "insider threats".

Banks are stuck on the same problem. Morgan Stanley and Bank of America are both working on how to handle outside agents reaching into customer accounts, where the genuinely hard bit is knowing whether an instruction came from the customer or from something acting on their behalf.

Here is why that lands on your desk too. One AI price index has dropped 54% since its May peak, to under $1 per million tokens. Running these things has stopped being expensive, which puts them well within reach of a two-person office. The rules stopped being a big-company problem at the same moment.

Three things we would sort out first

None of the above needs a big budget. It needs an hour, and a bit of thought about the order you do things in.

Pick a job you already do badly because you never get to it. Chasing missing paperwork. Answering the same five questions. Tidying the inbox on a Sunday night. Agents earn their keep on the work you resent, not the work you enjoy, and starting there means a poor result costs you nothing you were not already losing.

Write down what it is never allowed to do, before you connect it to anything. No sending, no spending, no deleting, not until you have watched it work for a fortnight. Microsoft has built an entire product on that principle. You can get most of the same protection from one sentence and the discipline to stick to it.

Keep the approval step after it starts to annoy you. It will. The morning it feels like a pointless extra click is precisely the morning it is doing something worth checking.

Where to learn this properly

If one thread runs through all five, it is approval. Every serious tool released this week assumes a person signs things off. That is the right instinct, and it is the first thing people drop.

The rest is ordinary work done faster, by tools that have genuinely got good. The difficulty is less about the technology than about knowing which claims to believe. You can learn how to do all of this on our upcoming AI for Business workshops. Bring the job that eats your week, and your most awkward questions.