Better AI models are arriving, but the business case still needs proving. Today's news connects a UK survey on disappointing returns with new Claude releases, Meta's reported shopping-agent plans and stronger demands for safety evidence.
UK survey finds AI spending often starts before the business problem is agreed
Research commissioned by Air IT Group, reported by PCR, surveyed 504 UK decision makers. It found 83% either could not demonstrate returns from some AI spending or did not measure returns at all. That does not mean 83% of all AI investment has failed.
More revealingly, 41% said an AI tool had been introduced before the business problem was agreed. Some 47% said AI had made a task or process take longer once checking, corrections, setup and management were included. The supplier-commissioned survey is a useful warning about measurement, not a national audit of every AI project.
Our take: Before another licence purchase, choose one recurring task and record its current completion time, error rate and checking cost. Compare the AI-assisted result on the same basis. A faster first draft is not a saving if verification consumes the difference.
Anthropic releases Fable 5.2 and Mythos 5.2 with lower prices
Anthropic announced Claude Fable 5.2 and Claude Mythos 5.2 on Monday. The company says they share an underlying model, with Fable carrying stronger safeguards for broad access and Mythos available through verification programmes for cybersecurity and life-sciences work.
Standard pricing is $8 per million input tokens and $40 per million output tokens. An optional faster mode costs twice as much. Anthropic reports improvements in coding, reasoning and vision, but its own guidance still recommends starting most everyday coding and knowledge work with Opus 5.5 rather than automatically choosing the most capable model.
Our take: Test the new model on the difficult exceptions your current system cannot complete. Keep routine tasks on a cheaper suitable model, and compare the full cost of a checked result rather than the headline token rate. Supplier benchmark gains do not establish your own return.
Meta's Muse shopping agent is reportedly heading to the UK
Business Matters reports that Meta plans to launch its Muse personal AI assistant in the UK within weeks, with retailer partnerships intended to let it make purchases. Meta declined to comment, so treat the timing and partnerships as reporting rather than a confirmed UK launch announcement.
The publication says Muse already operates in North America and can handle tasks including shopping, transport bookings and bills. For UK retailers, the practical question is whether product details, stock information, delivery terms and returns rules remain understandable when a software agent sits between the customer and the shop.
Our take: Retailers can prepare without signing up to an unconfirmed programme. Make prices, availability and buying conditions consistent across the website and product feeds. Keep clear approval boundaries wherever an agent can commit a customer's money.
Bailey calls for stronger AI testing and incident reporting in new speech
In a new speech in Zurich on Monday, Bank of England governor Andrew Bailey called for more extensive testing before and after AI models are released, alongside better incident reporting and follow-up. PA reports that he highlighted cyber risk and the financial stability consequences of an AI asset-price bubble.
This is a separate development from our Saturday coverage of the Bank's productivity estimates. The new emphasis is operational assurance: understanding failures, reporting them and checking whether investment expectations match the time needed to deliver returns.
Our take: A small business does not need a central bank's risk apparatus. It does need an owner for each important AI workflow, a record of incidents and a way to stop or revert changes when the system behaves unexpectedly.
Six governments and major AI developers sign Paris safety declaration
Canada published a joint declaration signed on 9 November by Bahrain, Canada, Finland, France, Norway and the Netherlands, alongside companies including Anthropic, Microsoft, OpenAI, Mistral AI and the UK's Synthesia. The statement calls for comparable evaluations, stronger accountability and information sharing about serious incidents.
Developer commitments include disclosing capabilities and limitations, supporting appropriate independent evaluation and maintaining mechanisms to stop model operations within their control. This is a cooperation declaration, not a new UK legal requirement or a guarantee that every signatory's products meet an identical technical standard.
Our take: Use the direction of travel in procurement. Ask suppliers what they test, what evidence they will share, how serious incidents are communicated and what controls customers retain. A signature on a declaration is a starting point for questions, not a substitute for answers.
Google adds confidential and high-memory TPU systems
Google expanded its eighth-generation TPU portfolio with TPU 8tx and TPU 8i Max. The company distinguishes these new system configurations from entirely new chip designs: 8tx uses the existing 8t silicon, while 8i Max combines two 8i chips in one package.
Google positions 8tx for confidential inference with sensitive data and model weights, and 8i Max for large context windows and memory-intensive agent workloads. These are supplier announcements about infrastructure capabilities. They do not automatically establish UK data residency, regulatory compliance or a lower bill for a particular application.
Our take: For most SMEs this matters through a cloud or software supplier rather than a hardware purchase. Ask whether the service offers the required processing location and isolation, then compare measured workload costs. More memory capacity and confidentiality solve different problems.
Women in technology report stronger AI skills alongside workload pressure
Ensono's 2026 Speak Up survey, reported by PCR on Monday, covered 1,500 women working in technology across the UK, US and India. Seventy per cent rated their generative AI knowledge as advanced or expert, while more than four in ten said AI was bringing expectations to produce more work in less time.
The survey also found 43% felt pressure to keep developing skills to remain competitive. These are cross-country, self-reported findings, not UK-only measures, and they do not prove that AI alone causes burnout. They nevertheless highlight the staffing consequences of turning assumed productivity gains into immediate workload targets.
Our take: Give staff protected learning time and measure quality, rework and sustainable output together. Raising targets before workflows actually improve can hide costs in unpaid checking and exhausted teams. Staff feedback belongs in the same review as licence usage and task completion.