The next wave of AI buying is about control, not just capability. Today brings agent security deals, clearer business definitions and a reminder that customer-service automation changes the work left for people.
CrowdStrike plans $495 million AI security acquisitions and an agent kill switch
CrowdStrike announced plans to acquire Fiddler AI and Dreadnode, with a combined price reported at $495 million. The companies bring agent monitoring, risk testing and simulated attacks on AI systems and their connected tools.
CrowdStrike also announced an AI Kill Switch for Falcon Guardian, intended to let administrators stop risky agents, remove network access and prevent relaunch. UK buyers should distinguish announced capabilities from protection already deployed in their own environment.
Our take: Before giving an agent access to customer records or production systems, decide who can stop it, how access is revoked and what evidence is retained. Buying another security product is not a substitute for those decisions.
Snowflake adds shared business definitions and tighter agent controls
Snowflake announced Business Ontology in private preview, allowing organisations to establish approved definitions for the business concepts their agents use. It also expanded Cortex AI Gateway with abuse detection, tracing, cost governance and model routing.
Cortex Training entered public preview for tailoring open-weight models. For UK organisations, the immediate issue is whether different agents interpret measures such as revenue, active customers and risk consistently, not simply whether they can access the same database.
Our take: Agree the meaning of your important business terms before automating decisions around them. Two agents can read accurate data and still give incompatible answers if the definitions differ.
NICE announces about 500 job cuts during its AI transition
Customer-service software company NICE announced plans to cut around 500 roles worldwide, approximately 5% of its workforce, as it reorganises investment around enterprise customer-experience AI. Globes reports that its latest annual report listed 9,626 employees at the end of 2025.
The company is responding to a market being reshaped by agents that can carry out service tasks. The announcement is evidence of a supplier restructuring, not proof that an equivalent proportion of any UK contact centre can safely be automated.
Our take: Measure successful resolution, repeat contacts and customer outcomes before making workforce assumptions. Automating easy conversations can leave human teams handling a harder workload that needs more training, not less.
Google clarifies school AI controls after Search backlash
Google said Gemini, Gemini Notebook and Search would be switched off by default for schools, with administrators able to choose which products to enable. The clarification followed criticism of reports that schools rejecting AI would lose access to Search.
CNBC notes that Google did not explicitly resolve whether Search access requires AI features, and an education executive said Search and AI are deeply integrated. This is a US controversy, but UK education and workplace buyers should check the actual settings and contractual choices available to them.
Our take: Ask suppliers to demonstrate what happens when you turn AI off. A separate product toggle does not necessarily mean every AI feature inside the remaining service can be disabled.
Traversal raises $120 million for AI-assisted incident response
Traversal raised $120 million at a reported $1 billion valuation for technology designed to investigate and help resolve production-system incidents. Its system analyses logs, metrics and traces to identify likely causes and propose remedies for human engineers.
The funding underlines growing demand for support around the operation of rapidly changing software. For a UK business buying AI-assisted development, the relevant question is who diagnoses failures and authorises recovery after the new code goes live.
Our take: Treat diagnosis and remediation as different permissions. Start with read-only investigation and recommended fixes, then test approval and rollback paths before permitting changes to live systems.
AMD reports $13.38 billion revenue as datacentre demand grows
AMD reported third-quarter revenue of $13.38 billion, up 45% year on year, according to SiliconANGLE. Its datacentre business accounted for 64% of total revenue, compared with 47% a year earlier.
The results show demand extending across the infrastructure needed to operate AI workloads. They do not establish a return on investment for the businesses buying those services, or guarantee that cloud prices will fall.
Our take: Compare the cost of completing a real workflow, including retries, tool calls and supervision. A cheaper model or faster chip matters only if the full job becomes more reliable or economical.
Deep Principle releases Hydrogen-1.0 for scientific research assistance
Deep Principle has published Hydrogen-1.0, a scientific research assistant based on GLM-5.3, with MIT-licensed weights. Its model card lists a one-million-token context and a maximum output of 128,000 tokens, subject to deployment resources and configuration.
The release includes variants for Nvidia and AMD hardware and examples involving scientific computing, experimental planning and structure analysis. The publisher explicitly says outputs require scientific review and must not be treated as verified findings.
Our take: Open weights can give research teams more deployment control, but they do not remove infrastructure costs or validation work. Test a narrow scientific task with expert review before assuming general research competence.