Five AI developments sharpen a practical question for marketing and advertising teams: when a platform offers more access, how much control do you actually gain? A persistent shopping cart, spending controls, shared creative specifications, downloadable model weights and outside evaluators each move a boundary. The useful work is understanding where that boundary now sits.
1. Google keeps the shopping decision inside a persistent cart
Google launched Universal Cart in the U.S. on October 15 across Search and Gemini. Shoppers can keep products from different retailers together, receive price-drop and restock notifications, and find promotions. Checkout can happen through Google Pay with participating merchants or by moving items to the merchant's cart.
Google says merchants remain the seller of record and provide confirmations and support. This is a shopper-controlled tool; the announcement does not establish independent purchasing authority for an AI agent.
The big picture
For retailers, the strategic issue is where consideration happens. A customer may keep the unfinished shopping journey inside Google even when the merchant handles the sale. Treat this as a distribution opportunity to measure, not an automatic gain in customer ownership or incremental revenue.
Read more: Google – Universal Cart U.S. launch
2. Google Ads adds steering controls, not guaranteed spending ceilings
Google's channel-prioritization controls are rolling out progressively across Search, Shopping, Performance Max and Demand Gen. Advertisers can rank supported channels high or low. Budget allocation adds a desired daily percentage for eligible low-priority channels, but Demand Gen does not support that allocation feature.
The help documentation makes an important distinction: nonzero allocations operate on a best-effort basis and can be exceeded. A low-priority allocation of 0% prevents spending on that channel. The percentage is calculated against the average daily budget, not actual campaign spending.
The big picture
Media teams gain a more explicit way to express preferences to automation. They still need to reconcile settings with delivered spend. A target percentage should not be described to a client as a hard cap when the platform itself says it can be exceeded.
Read more: Google Ads Help – Channel prioritization and budget allocation
3. Figma brings the brief onto the working canvas
Figma introduced Specs on October 15, with open beta rolling out to paid plans. Specs brings written context beside designs, diagrams and code references. Through Figma's MCP server, coding agents can read specifications and the context they reference; Figma also describes using MCP to update a specification as implementation progresses.
The big picture
Creative and product teams often lose decisions between the brief, the design file and implementation. Shared context could reduce that friction. But if an agent can also change the specification, someone still needs to own approval of the brief. Faster handoffs are useful only when the delivered work follows the intended direction.
Read more: Figma – Introducing Specs
4. Meta makes Muse Spark Open weights available
Meta released Muse Spark Open weights on October 15. It describes a multimodal model with 816 billion total parameters, 42 billion active parameters and a one-million-token context window. Meta says it can run using two M5 Ultra Mac Studios and is also available through hosted inference providers.
The release uses the Muse Model License with limited commercial terms. Downloadable weights should not be confused with unrestricted commercial rights. Meta's deployment and capability descriptions are vendor claims, not independent proof of better results for a particular business.
The big picture
For an agency evaluating private AI workflows, deployment choice matters. It can create alternatives for where sensitive work runs. The decision still requires a workload test, infrastructure costing and license review; a model being downloadable does not answer those questions.
Read more: Meta – Introducing Muse Spark Open
5. Anthropic details embedded outside evaluation
Anthropic said October 15 that six METR employees have been onboarded with hardware, office space and access broadly comparable to its internal risk-assessment teams. Customer data and privileged legal information are excluded. Model-architecture information is normally restricted, with relevant exceptions requiring approval.
Anthropic says it cannot veto METR's publications, although it can require narrow redactions. METR can publicly characterize those redactions and say whether it considers them justified. Other evaluator partnerships have more limited rights.
The big picture
An outside evaluator's name is less useful than knowing what it can inspect and disclose. Enterprise buyers should look for the resulting findings and limitations. Access is an improvement in the evidence-gathering process, not a certification that a model is safe for every marketing use case.
Read more: Anthropic – An update on independent evaluators
THE THROUGH LINE
The common shift is toward tools that expose more of the operating model: where shopping persists, how spend is steered, what an agent reads, where a model runs and who can inspect it. For marketing leaders, the next question is concrete: which decisions can the team actually control, and which still depend on the platform? Before expanding a pilot, write down the boundary, the evidence needed to trust it and the person responsible when it fails. That turns a promising feature into a decision the business can evaluate.
