On Friday 30 January 2026, Anthropic published eleven plugins for its desktop tool Claude Cowork. One of them reviewed contracts and triaged NDAs. By the following Tuesday, Thomson Reuters and LegalZoom had lost more than 15 percent of their market value, and RELX, the owner of LexisNexis, was down by double digits.
Investors, in other words, had panicked over a folder of instructions.
That is a fair summary of the first quarter of 2026. AI stopped being something you talk to in a chat window and started doing work directly in your files, your browser and on your desktop.
January: the agent gets its own tab
On 12 January, Anthropic introduced Cowork as a research preview. A tab in the Claude desktop app, next to chat, where the assistant reads and writes in the folders you let it into. The same day, Apple and Google announced a multi-year deal: Gemini will power the next generation of Siri. Apple, a company that has built its identity on doing everything in-house, bought the brain of its voice assistant from a competitor.
The most talked-about agent of the month came from no big company at all. OpenClaw started as a weekend project by the Austrian developer Peter Steinberger in November 2025, an assistant that answers in WhatsApp and can run things on your own computer. It changed its name twice in four days at the end of January, first after a trademark objection from Anthropic and then again on 30 January. By 2 March it had 247,000 stars on GitHub.
The downside showed up just as fast. Security researchers at Cisco demonstrated that third-party skills could exfiltrate data without the user noticing, and in March Chinese authorities barred state agencies and banks from using it. Steinberger himself joined OpenAI on 14 February.
February: the market does the maths again
After the Cowork plugins, analysts started using the phrase "SaaS apocalypse". The idea was simple and unpleasant for software vendors: if an assistant costing a few dozen dollars a month can do what a specialised licence costs thousands for, what is the licence worth? It is not quite that simple, and several analysts pointed out that large organisations rarely switch systems in a week. Still, the question had been asked out loud.
On 5 February Anthropic released Claude Opus 4.6, with a one-million-token context window in beta and unchanged pricing. OpenAI shipped GPT-5.3-Codex the same day. Google followed on 19 February with Gemini 3.1 Pro.
Two other February events matter more to a small Swedish business than any version number. On 9 February OpenAI began showing ads in the free tier of ChatGPT and in the low-cost Go plan, initially only in the US. The free assistant became ad inventory, with everything that implies about whom it really works for.
On 27 February President Trump ordered federal agencies to stop using Anthropic's technology, and Defense Secretary Hegseth designated the company a supply chain risk. The background was Anthropic's refusal to let its models be used for fully autonomous weapons and mass surveillance of Americans. The dispute is still in the courts, and as recently as 25 September a federal appeals court sided with the Pentagon. If your routines rest on a single AI vendor, it is a reminder that the vendor's relationship with a government can change over a weekend.
March: the computer becomes the workplace
OpenAI released GPT-5.4 on 5 March, its first general-purpose model that can operate a computer on its own through screenshots, mouse and keyboard. On the OSWorld-Verified benchmark it completed 75.0 percent of tasks. Humans score 72.4. On 17 March the smaller mini and nano variants followed, and mini was made available to free users.
Brussels produced something that got less attention and matters more to anyone thinking about compliance. On 13 March the Council agreed its position on the so-called Digital Omnibus, and Parliament adopted its negotiating mandate the same month. The package postponed the AI Act's rules for high-risk systems. With the summer behind us we know how it ended: a deal on 7 May, final Council approval on 29 June. Stand-alone high-risk systems must comply from 2 December 2027, and systems embedded in products from 2 August 2028.
Before handing an agent the keys to your company folders, it can pay to start further away from the technology. What jobs do people want done, and what will they pay for them? For a straight answer, see what people nearby need help with. The requests are there in plain language, with price and location.
What the quarter left behind
2025 was the year AI got cheap, as we covered in our look back at last year. The first quarter of 2026 was when AI got hands. An assistant that reads a folder of receipts and fills in a spreadsheet is no longer a vision of the future. It is an app you can install this afternoon.
For a small company that is good news with fine print. An agent with access to your files is also an agent with access to your customers' data, and OpenClaw showed how quickly things go wrong when nobody has thought about permissions. Start with a boring task that is easy to check afterwards. Customer replies are often a good place, and we have worked out what it actually costs to automate them.
And do not put the whole business with a single vendor. February taught us that one for free.
Sources: Anthropic (Claude Opus 4.6, 2026-02-05); ABC News on the selloff after the Cowork plugins (February 2026); CNN and TechCrunch on Apple and Gemini (2026-01-12); OpenAI, Testing ads in ChatGPT (2026-02-09); Axios and TechCrunch on the Pentagon and Anthropic (2026-02-27); CNBC on the appeals court ruling (2026-09-25); European Parliament Legislative Train, Digital Omnibus on AI; Council of the European Union (2026-05-07 and 2026-06-29).