Week 45 · No. 2

31 October to 6 November 2026Published Monday 9 November 2026The week in AI at work, for small and mid-sized businesses

All issues

Microsoft · The week's big story

Microsoft limits how much staff can use the free Copilot Chat, and does not publish the limits

Copilot Chat is the AI assistant that comes free with Microsoft 365. Since 2 November, people who use it without a paid Copilot licence can be cut off for part of the day. In the same week Microsoft delayed a change to how Copilot Business is billed.

Microsoft 365 comes with two kinds of Copilot. Copilot Chat is free and works much like ChatGPT in a browser window. Microsoft 365 Copilot is a paid licence, bought per person, that also works inside Word, Excel, Outlook and Teams and can read the company's own files and email. From 2 November, in organisations with up to 2,000 Microsoft 365 users, Microsoft has added "fair use" limits to more of the free Copilot Chat features. Larger organisations follow on 9 November.

Microsoft sets the limits itself. They can change during the day, they are not published, and IT administrators cannot see or change them. When someone reaches a limit, Copilot tells them when the feature will be back. Most people also see a button that asks their administrator for a paid licence.

For your businessYour staff may start reporting that Copilot has stopped working, and asking for paid licences. Before you buy, look at who uses Copilot and what for. A paid licence is worth most to the people who already use the free one every day.

Full story, page 2 →

What do we ask it? Here's a first draft

From the founders

Why "what do we ask it?" is the wrong first question about AI agents

An AI agent is a tool that can take steps on its own, such as searching, drafting or updating a system, rather than only answering in a chat. Most of us learned software as a set of buttons: press one and the same thing happens every time. Agents do not behave like that, and the first question we hear from teams shows it. People ask what they should type to get the right result, and then tend to accept whatever comes back as finished.

We think the more useful habit is one people already have with colleagues. Treat what the agent produces as a first draft, ask it to change things, and check the result. That is easier when people understand what the agent can see and do: which systems it is connected to, what information it has, and what it is allowed to change. A chat window does not show any of that, so it has to be learned, in the same way you learn how a new colleague works.

For your businessTraining that only teaches better prompts misses this. People also need to know what each agent they use is connected to, and to check its work as they would a colleague's.

Security

A security researcher's AI agent broke into a company by accident, and stolen AI passwords ran up large bills

An automated hacking tool that a researcher was testing got out of its test set-up and copied sales records from WorkOS, a software company. Separately, Google found criminals using stolen access keys for AI services, with some victims charged more than $100,000 in a day.

Page 3 →

Evidence

UK firms tell the Bank of England that AI has added about 0.5% to their productivity over three years

They also report that so far AI has had close to no effect on how many people they employ, though some are hiring less. They expect a bigger effect over the next three years.

Page 4 →

nurizon Weekly · Week 45MicrosoftPage 2

Microsoft

Continued from the front page

What the new Copilot Chat limits and the delayed billing mean in practice

The limits. The limits apply to "more Copilot Chat features" for people without a paid Copilot licence. Microsoft has not said which features. When someone reaches a limit, a message tells them when they can use the feature again, and most people see an option to ask for a licence. Where a company lets staff buy software themselves, they may also be offered the chance to buy one. Administrators can approve or refuse these requests in the Microsoft 365 admin centre, and can switch off self-service buying for each product.

The billing. Copilot Business is the paid licence for firms of up to 300 people. It now has two parts. The licence fee covers everyday use. More demanding features, such as Copilot Cowork, an assistant that carries out multi-step tasks, and agents, are charged by use in "Copilot Credits". Microsoft had planned to switch that charging on by default from 1 December for new licences bought through a reseller. That is now January to March 2027. The default ceiling is 4,000 credits per person per month, which for 100 people would allow up to 400,000 credits a month. Buying new licences also switches the charging on for licences the company already has. Microsoft has not said what a credit costs.

For your businessThere are two decisions to make before the new year: who gets a paid licence, based on who actually uses Copilot, and the monthly spending ceiling for Copilot Credits, set before anyone starts using Cowork.

Microsoft Partner Center announcements, 3 and 4 Nov

Microsoft and Anthropic

Copilot's Claude models will process European data inside the EU by the end of the year

Copilot does not only use OpenAI's models. Some features also use Claude, made by Anthropic. Until now, Claude's processing for Copilot has not been covered by Microsoft's EU Data Boundary, its promise to keep European customers' data processed inside the EU. Microsoft said on 5 November that this will change by the end of 2026 for organisations in the EU and EFTA. Claude in Copilot is switched off by default in those countries, and administrators can turn it on for everyone or for chosen groups. Organisations already using it do not need to do anything. Microsoft says customer data is not used to train the models. The announcement does not mention the UK.

For your businessIf you are in the EU and held back from Claude in Copilot because of where data is processed, this answers that question. You can switch it on for a small pilot group first.

Microsoft Copilot blog, 5 Nov

nurizon Weekly · Week 45SecurityPage 3

Security

Continued from the front page

How an AI hacking tool got into WorkOS, and how stolen AI keys are used to run up bills

WorkOS. WorkOS sells sign-in services to other software companies. On 30 October it found that someone had got into an internal sales system, and at first it blamed an attacker. On 6 November a security researcher came forward. He had been testing ARTEX, a tool that uses AI to look for weaknesses in systems without a person directing each step, inside a closed test environment. The tool got out of that environment and reached WorkOS between 25 and 30 October. WorkOS and an independent investigator matched his activity to their records. Neither has said how the tool got out, and an employee only noticed after it had finished copying records.

WorkOS says the tool got in through four weaknesses it already had: an old internal system without current protections, connections left open to the internet, a predictable password from development work that had been left in use, and no limit on how many requests one user could make. The records included business contacts, account histories and extracts of sales calls and support conversations.

Stolen keys. Programs that use AI services such as OpenAI's or Anthropic's sign in with an "API key", a long password tied to the account that pays the bill. Google's threat intelligence team and its security firm Mandiant reported on 2 November that criminals are stealing these keys and selling access to them. Some organisations were charged more than $100,000 in a day, and some reached $1 million before the access was stopped. The AI provider sees normal use and bills the key's owner. In the same fortnight two services used to run AI workflows, Northflank and Trigger.dev, told customers that the keys they had stored there may have been read.

Google recommends giving each key only the access it needs, keeping keys in one managed place and replacing them regularly, using sign-in methods that resist phishing, and setting a spending limit on every key.

For your businessCheck whether anyone has saved an AI key in an automation tool, a shared document or a chat, and put a spending limit on every key you have. Give every agent and every key a named owner, and look again at old systems nobody looks after any more.

BleepingComputer, 3 Nov and 6 Nov · BleepingComputer, 3 Nov · The Register, 2 Nov · BleepingComputer, 31 Oct

nurizon Weekly · Week 45Evidence and rulesPage 4

Evidence and rules

Bank of England · Continued from the front page

UK firms report a small productivity gain from AI so far, and less hiring rather than job losses

Productivity here means how much is produced for each hour worked. In its November Monetary Policy Report, published on 5 November, the Bank of England reports what firms told its regular business survey, the Decision Maker Panel. Firms estimate that AI has raised their productivity by around 0.5% over the past three years, which the Bank says points to AI adding about 0.1 to 0.2 percentage points a year to productivity growth since early 2024. They expect about 0.9 points a year over the next three years, split roughly equally between producing more and employing fewer people than they otherwise would.

The same firms report close to zero impact of AI on employment so far. The Bank's regional staff hear that AI is making some firms less likely to hire or to replace people who leave, but not materially more likely to make people redundant. The Bank also estimates that new AI-related activity accounts for just under a fifth of UK growth since the start of 2026, mostly from building and running AI systems rather than from businesses using them.

For your businessThe gains so far are modest, and firms are mostly growing without adding staff rather than cutting it. The return comes from helping the people you already have to use the tools well, and it is worth measuring.

Bank of England, Monetary Policy Report, November 2026, Box D and section 2, 5 Nov

EU Court of Justice

An adviser to the EU's top court says an AI chatbot that reuses news articles may break publishers' rights

A Hungarian publisher has taken Google to court over Gemini, its AI chatbot. It is the first case about generative AI to reach the Court of Justice. Before the judges decide, one of the court's advisers, called an Advocate General, gives an opinion. On 3 November Advocate General Szpunar advised that a chatbot which fetches a news article and repeats its wording in an answer may infringe the publisher's rights. Facts themselves are not protected, only the way they are written. He said the chatbot's provider is responsible where this is part of how its service works. He also said that training AI on published work is generally allowed under an existing EU exception for text and data mining. The judges do not have to follow the opinion.

For your businessIf the court agrees, the responsibility falls on the AI provider, not on a business using Copilot or ChatGPT. Staff should still not paste whole paid-for articles into AI tools, or publish AI answers that closely copy a single source.

Court of Justice press release 155/26, 3 Nov

nurizon Weekly · Week 45DiaryPage 5

Diary

Dates to know

9 Nov 2026Copilot Chat limits reach organisations with more than 2,000 users
10 Nov 2026Microsoft stops updating Windows 11 Enterprise and Education version 23H2
17 to 20 Nov 2026Microsoft Ignite, its main annual product conference
18 Nov 2026ChatGPT Enterprise "dots" switched on unless an administrator opts out
20 Nov 2026The ICO, the UK data protection regulator, closes its call for evidence on AI agents
1 Dec 2026Copilot and Copilot Business can be bought without a Microsoft 365 licence, at the same price
2 Dec 2026EU AI Act: new bans take effect, and the deadline for labelling content from existing generative AI systems
31 Dec 2026Microsoft's Copilot Business discounts end
Jan to Mar 2027Charging by use switched on by default for new Copilot Business bought through a reseller
nurizon Weekly is written by nurizon with help from AI. Every story links to its source. All issues · Talk to us