AI-linked activity is helping Britain's economy grow, but turning new capabilities into dependable services still requires permissions, capacity and human judgement. Today's stories range from a stronger UK GDP reading to an airline drawing a clear boundary around automated purchases.
UK economy grows 0.7% as AI-linked services help lift output
The UK economy expanded by 0.7% in the third quarter, according to new figures reported this morning by Daily Business. Services grew 0.9%, while production rose 0.1% and construction fell 0.7%.
ONS chief economist Grant Fitzner said growth was almost entirely driven by services, with computer programming recording its fastest growth in three years and evidence that AI-linked businesses helped boost the sector. That is evidence of stronger activity in particular industries, not proof that every business buying AI is becoming more productive.
Our take: Separate growth in AI suppliers' sales from gains inside your own business. Choose a repeatable task and measure time saved, error rates and total delivery cost before increasing your budget.
United blocks AI agents from completing flight bookings
United says third-party personal AI assistants may show customers its flights but cannot complete purchases on their behalf. The airline cites privacy, unintended purchases and regulatory requirements, including declarations passengers must make about their luggage.
The Verge reports that United has not yet seen large volumes of agent bookings and may reconsider as protections develop. For UK businesses planning automated travel or purchasing, the immediate lesson is that a capable agent does not automatically have permission to transact with every supplier.
Our take: Design the handover before switching on purchasing automation. Let the assistant research options, but require a person to confirm declarations, payment and any supplier-specific conditions where automated access is not authorised.
Google Tabletop previews shared apps built from everyday instructions
Google Labs has introduced Tabletop, an experimental tool that generates collaborative spaces and interactive tools from prompts. Examples include itineraries, expense splitters, dinner polls and savings trackers, with several people able to use a shared space at once.
The beta is rolling out through a waiting list and is initially restricted to adults in the US. UK teams should therefore treat this as a product direction to watch, not an available replacement for their current project-management software.
Our take: The useful idea is making a small tool around a specific coordination problem. Before introducing anything similar at work, establish who owns the information, who can change it and how you will export it when the experiment ends.
AlphaGenome 2 predicts genetic effects at individual cell level
Google DeepMind unveiled AlphaGenome 2, trained on more than 100 million individual cell profiles across 68 tissues. It predicts how genetic variants affect different cell types and states, rather than treating the same DNA change as having one uniform effect.
DeepMind says it identifies nearly twice as many gene-expression-altering variants as its previous leading model. It is available for non-commercial research through the AlphaGenome API, with a limited private preview for commercial researchers; more detailed predictions also require more computation.
Our take: For UK life-sciences businesses, the opportunity is better prioritisation of laboratory work, not a shortcut around validation. Check commercial access terms and benchmark a relevant research question before assuming this will reduce development costs.
Google's XSE targets cheaper, faster production inference
The Register reports that Google's newly open-sourced XLA Serving Engine, or XSE, handles more than six million queries per second across services including Search, Ads and YouTube Shorts. It compiles model computations for a native runtime rather than running a Python interpreter for each request.
Google reports two to five times the throughput of its best optimised Python server on short requests, using a quarter of the memory. This is aimed particularly at ranking, recommendation, embedding and vision models, not a blanket replacement for large language model serving systems.
Our take: Optimise the workload you actually have. For teams serving high volumes of small predictions, runtime overhead can matter as much as model size; for a small business using hosted chat tools, this is infrastructure news rather than a migration instruction.
New analysis puts electricity connections at the centre of AI capacity risk
Howden analysis reported by The Register finds US data-centre power demand has grown by about 19% a year since 2000, compared with roughly 2% annual growth in grid capacity. Data centres accounted for around 5% of US grid capacity in 2025.
The report estimates that holding that share steady would require up to 58% of data-centre power to come from outside the grid by 2030. This is a scenario about infrastructure constraints, not a prediction that data centres will consume all available electricity or a direct estimate for the UK.
Our take: When an AI supplier promises major new capacity, ask about delivery dates, power availability and fallback regions. A signed building agreement is not the same as usable compute, and UK data-location commitments still need to survive an outage.
Adobe previews more controllable AI editing rather than finished products
Adobe's Max previews include Project Sight Span, which turns a flat image into an explorable scene, and Project Screen Test, which lets creators shape a generated character through conversation. Project Rinse Repeat transfers edits to other subjects while recreating them with native Photoshop tools.
These are experimental demonstrations, not guaranteed product releases. They point towards more interactive control over image, video and audio work, but UK marketing teams should not promise customers a workflow that is not yet commercially available.
Our take: Watch the tools that preserve editable work, not just the most striking demo. Keeping native edits and an approval trail makes corrections and brand review easier than relying on a single generated output you cannot reliably reproduce.