AI is moving deeper into spreadsheets, meetings and the systems behind everyday work. Today's announcements make the practical question sharper for UK businesses: how do you check the output, protect access and reward the people doing the checking?
Claude moves into spreadsheets, meetings and shared whiteboards
Anthropic announced Claude Sheets, Meetings and Whiteboards on 10 November. Sheets is available in beta to paid subscribers and supports collaborative editing, formulas and Excel export. Whiteboards groups ideas and drafts action items.
Meetings initially works with Google Meet for Team and Enterprise customers. Engadget reports that it identifies itself as a participant, posts a notice and requires administrator permission for calls with outside guests. For UK teams, meeting access and spreadsheet review now belong in the same adoption conversation as chatbot use.
Our take: Pilot one low-risk workflow rather than moving a whole department. Check formulas against known answers, agree who can invite the meeting assistant and make somebody responsible for reviewing action items before work is assigned.
Google opens Gemini 4 Carbon to paying developers and Ultra subscribers
Google has made Gemini 4 Carbon available through its API to paying developers and enterprises, alongside a rollout to Google AI Ultra subscribers. Introductory pricing is $2 per million input tokens and $10 per million output tokens.
Google reports leading results on coding, research and automation benchmarks and is previewing coordinated agent teams in Antigravity. These are useful signals for a shortlist, not evidence that a particular business process will become cheaper or more reliable.
Our take: Compare Carbon with your current model on the same real tasks, including failures and review time. A low token price can still produce an expensive completed job if the system needs more reasoning, retries or supervision.
Payscale finds AI is changing jobs faster than pay structures
Payscale's new AI Workforce Impact Report says 61% of organisations are rewriting job descriptions because of AI and 49% say salary structures are struggling to keep pace. Only 16% of employees report higher compensation from AI-related upskilling, while 70% report no change.
The vendor's research also finds that only 19% say time saved is going into more strategic or higher-value work. These findings are not a UK-wide employment estimate, but they raise a concrete management question for UK employers introducing AI into existing roles.
Our take: Do not quietly turn saved time into an unlimited workload increase. Agree how responsibilities, quality checks and progression change when staff learn to use AI, then review whether the promised benefit is reaching employees as well as the business.
Mercor raises $750 million to bring business-specific agent testing to enterprises
Mercor announced $750 million in funding at a $20 billion valuation, with investors including Nvidia, Salesforce and Zoom. SiliconANGLE reports that it plans to expand the expert-led evaluation and training work it has supplied to AI laboratories into enterprise customers.
The company builds custom benchmarks around a customer's own systems, policies and edge cases. Its announcement included an employee tender offer, and the split between that and fresh capital was not disclosed.
Our take: The useful idea does not require a large contract: write a small test set from your own work. Include awkward exceptions, permission boundaries and examples where the correct response is to stop and ask a person.
DeepMind reports machine-checked proofs for key Signal security properties
Google DeepMind says its AlphaProof Code framework formally verified selected security properties of the actual code behind Signal's cryptographic protocol. Human experts reviewed the specifications, while AI agents generated more than six million lines of Lean proofs over about four weeks.
Lean checks the proofs mechanically. The researchers also reintroduced two previously fixed bugs into an isolated version and say the system proved those versions insecure. This concerns specified cryptographic guarantees, not a claim that every part of Signal or every possible attack has been verified.
Our take: This is a stronger assurance pattern than asking the same chatbot whether its own answer looks right. Where practical, combine AI generation with independent checks such as reconciliation, tests, explicit rules or formal verification suited to the work.
Microsoft research shows why model files need supply-chain checks
Microsoft researchers describe Lazarus, a technique that can backdoor an AI agent by modifying fewer than ten dormant neurons in its underlying model, without retraining. The modified model can behave normally in standard evaluations until a trigger appears.
The attack requires write access to model weights; this is not evidence that ordinary prompts can rewrite a hosted model. Microsoft recommends provenance and integrity controls alongside limited permissions and monitoring, particularly when agents can use terminals or other powerful tools.
Our take: Treat downloaded model files as software dependencies, not inert documents. Record the publisher and version, verify integrity against trusted originals where possible, and keep an agent's permissions limited even after its output tests pass.
AMD launches Epyc 6 processors aimed at the work around AI models
AMD introduced ten Epyc 6 server CPUs, also known as Venice, with options for higher core counts or higher clock speeds. SiliconANGLE reports that the range is designed partly for the software tools used by AI agents, not just the neural networks themselves.
The densest chips have 256 cores, while higher-frequency options are aimed at different application needs. AMD's performance comparisons remain vendor claims and should be assessed against the actual workload, software licensing and available cloud instance pricing.
Our take: AI infrastructure costs extend beyond GPUs. For a growing agent workflow, measure database calls, tool execution and CPU time before buying more model capacity; the slow or expensive component may be outside the model entirely.