Research Period: 14 September to 21 September 2026
1. Claude merges Chat and Cowork into one unified interface
What changed
Anthropic has merged Claude Cowork and ordinary chat into a single interface. The rollout begins with Pro and Max subscribers over the coming weeks across web, desktop and mobile, with no geographic exclusions. Beta creation tools for documents, slides and design now sit alongside the conversation, producing outputs you can edit and share.
Why it matters
If you use Claude, you no longer have to decide where to start a piece of work. Chat, background execution and your existing projects live in one sidebar, which cuts the context-switching that quietly eats time across a working day. There is no setup, no new subscription and nothing to configure: the change arrives automatically for existing subscribers. For anyone who has drifted between surfaces and lost track of where a piece of work lives, that is a real, if unglamorous, improvement. API users see no change; this is a consolidation of the end-user experience.
Use cases
- Open claude.ai or the desktop app and check the new sidebar, no settings to change
- Start a conversation, then switch to a project, both now sit in one view
- Draft a document in the beta Document tool and share the editable output directly
- Build a short slide deck from a conversation without exporting to another tool
- Move background Cowork tasks and quick questions into one working thread
Sources: Anthropic Blog
2. Mollick: AI’s ‘Overhang’ and the four things humans still do better
Editorial inclusion. This is a strategic perspective piece rather than a product announcement, included because it addresses a question at the heart of what the Guild is for: where do humans fit as AI gets more capable?
What changed
Ethan Mollick argues we are living through an “overhang”: AI capability has expanded faster than most organisations have adapted to use it, leaving the typical user working at roughly 5 to 10 per cent of what the tools can already do. He names GPT-6 Astra and Fable 5.1 as models capable of weeks’ worth of human work when properly guided.
Why it matters
The useful part is his list of four areas where humans still hold an edge: deep specialised knowledge, meaning knowing which questions matter; wide contextual knowledge, meaning connecting domains that do not obviously touch; taste, meaning judging quality and fit; and agency, meaning making things actually happen. That is a practical sorting rule. Work that falls outside those four is where you hand more over. Work inside them is where your attention is still the scarce ingredient. Same author as our week 36 piece, but a different argument. Roughly ten minutes to read, no paywall.
Use cases
- Read the essay and mark which of the four advantages best describes your own contribution
- Audit one recurring task and decide whether it sits inside or outside the four
- Hand routine drafting and research to AI where none of the four apply
- Reserve your own hours for judgement calls on quality, fit and priority
- Use the framing in a team conversation about where people add value now
Sources: One Useful Thing (Ethan Mollick)
3. TypeSafe launches Jev, an AI that returns decisions instead of paragraphs
What changed
TypeSafe has released Jev, the first of what it calls System One models. Where a normal model writes text you then read or parse, Jev takes your data and a typed question and hands back an answer with a probability attached. TypeSafe’s own framing is clearest: ordinary models are “designed to produce text for humans to read”, while Jev is “built to make fast, structured decisions that software can use directly”.
Why it matters
Most of the automation a small business actually wants is not writing, it is deciding. Which inbox does this go to, is this lead worth a call, should a person look at this one. Today you either do that by hand or ask a chat model and hope the answer comes back in a form your spreadsheet can use. Because every Jev answer carries a number saying how sure it is, you can set a rule like “act on anything above 0.85, send the rest to me”. Five things it does, with the answer it gives back:
- Pick an action from your list of options, with a probability on each, so a near-tie is visible rather than hidden
- Classify: a support message returns billing 0.99, technical 0.01, account 0.00, so you see how clear-cut it was
- Score and rank: TypeSafe’s own benchmark puts one judgment at $0.000081 in 0.114 seconds, against $0.013880 and 8.566 seconds for an LLM doing the same job
- Say yes or no: “is someone trying to trick this chatbot?” comes back as 0.99, and several such checks run in a single call, so one request screens for several kinds of problem at once
- Read your JSON as it is: rows, tables, ticket histories, with no flattening into a prompt first
The shape is what matters here. This is a component you drop into a workflow, not another window to talk to.
Use cases
- Route incoming enquiries, with anything under your threshold landing in your own inbox
- Triage a ticket backlog by urgency before anyone reads it
- Rank applicants, leads or suppliers against criteria you describe in plain English
- Put a yes or no check in front of an existing automation so odd cases stop rather than proceed
- Score records once, then change the weightings without paying to run it again
Sources: TypeSafe | TypeSafe Documentation | Examples and Uses
Additional Noteworthy
OpenAI and Shopify launch Sponsored Agents: Merchants can run labelled conversational ads inside ChatGPT, synced to Shopify catalogues, the first paid advertising format in the product. US merchants only at launch. | The Neuron
Perplexity adds MCP connectors with OAuth: Paid subscribers can register tools such as Gmail and Slack once, then search across live business data without supplying credentials per request. | Perplexity Changelog
Salesforce launches Koa and opens Claudeforce: Koa is Salesforce’s own CRM reasoning model, and the Anthropic-powered Claudeforce add-on is now open to pilot customers, putting reasoning inside sales workflows. | The Rundown AI
Cohere and Aleph Alpha announce merger: The Canadian and German firms are combining at a targeted $20B valuation, creating a more credible non-US option for data-sovereignty cases. | The Neuron
Antitrust class action filed against four AI labs: A US District Court suit accuses Anthropic, OpenAI, xAI and Google of agreeing to slow AI development. Nothing to do; worth watching for pricing and market effects. | CBS News
Strategic Awareness
California Governor Newsom has signed SB 1050, which requires disclosure when video or audio advertisements use AI-generated performers, and tightens protections around the digital likeness of real people. The source does not give an effective date, so treat the timing as open.
This is California law, so for most Guild members it is a direction of travel rather than a compliance job. Disclosure rules for synthetic performers have moved quickly from proposal to statute, and UK and European regulators are watching the same problem. If you run advertising in the US that features AI-generated images of people, or synthetic voices, review it now. Everyone else should assume that within a couple of years, saying so when a face or voice is generated will simply be part of publishing an advert.
Sources: The Neuron
Looking Ahead
A quieter week for product launches, but a strong one for calibration. Watch whether the unified Claude interface reaches all tiers, and whether the disclosure rules signalled by SB 1050 start appearing in UK and European guidance.
This briefing focuses on AI developments relevant to SME operations, emphasising practical applications in productivity, cost efficiency, and business capability. For questions or to suggest sources reach out on the group WhatsApp.