What this year's AI releases change, sector by sector.

Updated 9 November 2026. Each brief pairs what shipped with what it changes for the business, the regulatory frame, and where we would start. Facts come from vendors, the companies involved and regulators, linked under each brief. Vendor performance figures are the vendors' own claims. These are market briefings on public announcements: the companies named are not unus clients.

BankingOct 2026

Agentic KYC and research agents are in production. Governance is now the gate.

What shipped
  • Sep 2026. Deutsche Bank Private Bank put agentic AI into source-of-wealth KYC. Agents research and document, then prepare the assessment for human review. Live first in Singapore and Hong Kong, with a wider rollout planned.
  • Aug 2026. Google Cloud launched Gemini Enterprise for Financial Services in preview, with a financial research agent and 50+ finance skills. Deutsche Bank is design partner for its German MidCorp teams.
  • Oct 2026. UBS reported 608 AI use cases in production and more than 12,700 AI agents run by employees in its Q3 results.
What it changes

Agent adoption is now a disclosed metric at DACH banks, and hyperscalers ship banking-specific agent stacks. The build-or-buy question moves to auditability, data residency and third-party risk.

Regulatory frame

Creditworthiness assessment and credit scoring are Annex III high-risk, with obligations from 2 Dec 2027 after the AI Omnibus. BaFin is Germany's AI Act authority for finance and already supervises transparency, prohibited practices and AI literacy. EBA, EIOPA and ESMA brought frontier-model ICT risk into DORA supervision in July 2026.

Where we would start

One documentation-heavy process, such as KYC refresh or credit-memo drafting. Design the human sign-off first, and route every model call through a gateway that logs it for DORA.

InsuranceAug 2026

Claims agents are moving into the core platform. Life and health pricing stays high-risk.

What shipped
  • Aug 2026. Guidewire's Qusar release added an agentic framework with access to policy, claims and billing data, plus prebuilt agents for claim summaries, policy changes and voice first notice of loss. Several features are early access.
  • Aug 2026. adesso acquired omni:us, whose agents handle simple claims from first notice to payout and act as copilot on complex ones. Allianz and UNIQA are named as users.
What it changes

Claims automation becomes a configuration of the core system rather than a separate project. The work shifts to governing agents over policy and claims data, and to deciding which claims a person must still see.

Regulatory frame

Risk assessment and pricing in life and health insurance is Annex III high-risk from 2 Dec 2027, and BaFin names individual surcharges as in scope. EIOPA proposed in April 2026 to exempt classic actuarial models such as GLMs; that carve-out is not in the adopted Omnibus, so plan as if they are covered.

Where we would start

First-notice-of-loss intake across voice and e-mail, with a structured hand-off to adjusters. In parallel, an inventory of pricing models against Annex III point 5(c).

TelecomOct 2026

Network operations go multi-agent. AI moves into the radio and into the call.

What shipped
  • Oct 2026. Deutsche Telekom, Nokia and NVIDIA reported typical uplink throughput gains of 10–15% from AI-based channel estimation in AI-RAN lab tests. A field trial is next.
  • Feb 2026. Telekom's RAN Guardian agent, built on Gemini, is live in Germany and cut major-event handling from hours to about a minute. MINDR extends multi-agent operations across RAN, transport and core, with production planned for 2026.
  • Mar 2026. The Magenta AI Call Assistant, built with ElevenLabs, adds network-based live translation and call summaries. It is opt-in and notifies every participant.
What it changes

Autonomy moves from single network domains to end-to-end service operations. Voice AI becomes a network product an operator can sell, with consent built into the design.

Regulatory frame

AI used as a safety component of critical digital infrastructure is Annex III high-risk from 2 Dec 2027; whether a given operations agent qualifies is still a matter of interpretation. In Germany the Bundesnetzagentur is the default AI Act market surveillance authority. BEREC is consulting on generative AI and internet openness.

Where we would start

An advisory operations agent over tickets, alarms and runbooks, with read-only tools, before any autonomous action on the network.

Energy & utilitiesNov 2026

Grid capacity is the constraint. Approval-gated agents and robots arrive in 2027.

What shipped
  • Nov 2026. E.ON, the enviaM group and NEURA Robotics announced humanoid robots for substation inspection, monitoring and site security. MITNETZ STROM pilots, with first grid experience expected in 2027. The utilities keep their data and the skills the robots learn.
  • Sep 2026. envelio announced IGP Agents that run load-flow simulations and check grid models in plain language, each action only after the user approves it, with no data sent to commercial AI providers. Launch is planned for Q1 2027.
  • Jun 2026. The Commission published its roadmap for digitalisation and AI in energy, and 48 organisations signed the Fraunhofer-led AI.Grids initiative for open-source grid forecasting AI.
What it changes

For AI data centres and large loads, a grid connection is now harder to get than land or chips: Germany's transmission operators have allocated connections by project maturity since April 2026. Utilities' own AI is moving to approval-gated agents and physical AI.

Regulatory frame

AI used as a safety component in the supply of electricity, gas, heating or water is Annex III high-risk from 2 Dec 2027. Germany's NIS2 registration grace period ended on 31 Jul 2026.

Where we would start

Approval-gated agents on planning and maintenance documentation, with data-sovereignty rules written down before the pilot starts.

ManufacturingOct 2026

From copilots that suggest to agents that act on the line.

What shipped
  • Oct 2026. Siemens showed factory agents at its Erlangen electronics plant, built on NVIDIA's Factory Operations Blueprint. They recommend production changes that planners test in a digital twin while the line keeps running. NVIDIA also released Cosmos 3.5 world models.
  • Apr 2026. The Siemens Eigen Engineering Agent became generally available inside TIA Portal, handling PLC programming, HMI visualisation and device configuration. It was piloted with more than 100 customers.
  • Mar 2026. BMW announced a humanoid robot pilot at Leipzig for high-voltage battery assembly, its first physical AI in European production.
What it changes

Automation engineering, a long-standing Mittelstand bottleneck, becomes an agent workload on tools plants already run. Simulation becomes the safety net for any agent that touches live production.

Regulatory frame

Connected products placed on the market since 12 Sep 2026 must give users access to their data by design (Data Act). The Cyber Resilience Act's 24-hour vulnerability reporting has applied since 11 Sep 2026. The Machinery Regulation applies from 20 Jan 2027, and AI safety requirements for machinery follow by delegated act by 2 Aug 2028.

Where we would start

One machine-data access layer that serves both Data Act requests and maintenance agents, so the compliance work and the AI work share a foundation.

Cross-industryOct 2026

Frontier agents got cheaper, EU-hosted options got serious, and MCP changed shape.

What shipped
  • Sep 2026. Anthropic's Claude Opus 5.5 scored 81.8% on the OSWorld 2.1 computer-use benchmark, with list prices cut to $4 input and $20 output per million tokens. It runs on AWS, Google Cloud and Azure as well as Anthropic's platform.
  • Sep 2026. OpenAI released GPT-6.1 Sol at $2 input and $10 output per million tokens. Its Agents API added computer use and multi-agent runs, and OpenAI agents can now run entirely inside AWS through Bedrock Managed Agents.
  • Oct 2026. Mistral released Large 4 as open weights: a mixture-of-experts model of about 1 trillion parameters with 49 billion active and a 1M-token context, plus a European deployment Mistral operates end to end. Aleph Alpha released Kolibri for German and English on customer-controlled infrastructure.
  • Jul 2026. The Model Context Protocol specification revision 2026-07-28 made MCP stateless, removed the session handshake and deprecated several features. Existing MCP servers and clients need a migration plan.
What it changes

The cost of an agentic task keeps falling, and EU-resident options now reach frontier scale. The advantage moves from picking a model to routing between models, governing them, and keeping integrations current as the protocols change underneath.

Regulatory frame

The AI Omnibus, Regulation (EU) 2026/1744, entered into force on 27 Jul 2026. Annex III high-risk obligations now apply from 2 Dec 2027, and product-embedded high-risk AI from 2 Aug 2028. Article 50 transparency duties have applied since 2 Aug 2026, with a grace period to 2 Dec 2026 for marking output of generative systems already on the market. Switzerland has no horizontal AI law; the Federal Council plans a consultation on sector-specific rules in spring 2027, and the Data Protection Act already applies to AI processing.

Where we would start

A model gateway with EU routing rules and cost limits, and an inventory of MCP servers with a plan to move them to the stateless specification.

Which of these is your next quarter?