When pilots turn into headcount plans

Highlight

When pilots turn into headcount plans.

On 8 July, Tomas Kunzmann, chief executive of Allianz Partners, agreed with employee representatives to cut between 1,500 and 1,800 jobs across Europe. The reason is stated plainly in the announcement: AI is taking over customer service and claims handling. Of the roughly 22,000 people the unit employs, some 14,000 answer phones or process claims.

Six days earlier, Claudia Buch appeared before the European Parliament's ECON committee with a figure worth remembering. More than 85 per cent of the banks supervised by the European Central Bank (ECB) already use AI tools, and over 40 per cent rate at least one use case as highly relevant to their business. A KPMG survey finds 51 per cent of banks already piloting agents in the narrower sense. Whether AI has arrived in banking is no longer an open question. It has, and it is now showing up in headcount plans.

What supervisors make of that is the interesting part. On 6 July the Financial Conduct Authority (FCA) published the Mills Review, which proposes an agentic supervisory model. Its author is the FCA's own executive director, Sheldon Mills. The regulator, in other words, has been advised from within to rebuild itself. A security finding from the same week shows how pressing that has become. You will find it below under Agentic AI. The agent did not break in. It used the permissions it already had.

LinkedIn Featured

Codex for Bankers: What the Agent Race Between OpenAI and Anthropic Means for Model Strategy

OpenAI has unveiled role-specific Codex plugins for investment banking. Anthropic got there first, with a suite of prebuilt finance agents already on the market. For banks, the interesting question is not who wins the race, but how they steer models, data connectivity and governance in the front and middle office when vendor release cycles keep outrunning procurement.

Read the full article: Codex for Bankers →

Agentic AI

tools, skills & what's trending

OpenAI news: GPT-5.6 arrives as Sol, Terra and Luna, alongside an agent called ChatGPT Work

OpenAI released the GPT-5.6 family on 9 July 2026, tiered into Sol, Terra and Luna, with an emphasis on long-running agentic tasks, coding and science. ChatGPT Work launched in parallel: an agent that acts across apps and files and, according to OpenAI, stays with a project for hours rather than returning an answer and waiting. From the same day, GPT-5.6 also became the preferred model in Microsoft 365 Copilot. If your organisation runs a Copilot fleet, you are largely changing models without having decided to.

Market structure: Meta starts charging for its models

Meta Superintelligence Labs, led by Alexandr Wang, presented the Muse Spark 1.1 reasoning model on 9 July 2026 and opened the Meta Model API, its first paid enterprise route to the company's models. For a firm that has positioned its models as the open counterweight to OpenAI and Anthropic, this breaks with its own story. Muse Spark 1.1 ships closed-weights, available only through the API. Meta charges 1.25 US dollars per million input tokens and 4.25 per million output tokens, roughly a quarter of what Anthropic and OpenAI ask, and it limited the preview to US developers at launch. Meta is buying its way in on price: one more serious vendor to compare, and one fewer open-weights option to rely on over the long run.

Security: GitLost, or when an agent turns its own permissions against you

On 6 July 2026, researchers at Noma Labs disclosed how the AI agent inside GitHub's Agentic Workflows can be induced to post the contents of private repositories as a public comment. The trigger is an innocuous instruction hidden in a public issue belonging to the same organisation. What the attack exploits is the reach of the permissions the agent already holds, with no software vulnerability in the classical sense. This is precisely the risk class that belongs in the prompt-injection chapter of any agent governance framework: give an agent access to core systems, and you must assume that every piece of text it reads may be an instruction.

Banking & Regulation

what really matters now

FCA: the regulator considers making itself agentic

The Financial Conduct Authority (FCA) published the Mills Review on 6 July 2026, led by its own executive director Sheldon Mills. It describes four shifts that artificial intelligence could set off in retail financial services by 2030, from reshaped customer journeys to sharper exposure to financial crime and cyber attacks, and it makes seven recommendations. Among them are an expanded FCA AI Lab and an agentic supervisory model in which the regulator itself works with AI agents. The counterweight sits in the same document: just 23 per cent of consumers trust an AI not to give them misleading information, and only 21 per cent trust it to handle their personal data responsibly. Supervisors are gearing up while customers hold back.

Digital euro: 36 payment service providers selected for the pilot

On 14 July 2026 the European Central Bank (ECB) named 36 payment service providers from across the euro area to take part in the digital euro pilot. The project now leaves the concept stage and becomes an integration problem, with concrete demands on onboarding, wallet connectivity and settlement. For the firms selected, the engineering work starts now. Everyone else should still watch closely, because this pilot phase defines the interfaces that the rest of European payments will eventually have to plug into.

Signal & Noise

what your time is worth

• SWIFT's shared ledger is ready for use, with 17 banks piloting – SWIFT
  Tokenised deposits, payments around the clock, banks across six continents. Thierry Chilosi describes the ledger as the foundation for future agentic commerce use cases. The rails are being laid before the agents run on them.
• Visa goes live with the first agentic purchases across Europe – Visa
  More than 30 European issuers, real merchants, secured through the Trusted Agent Protocol and Payment Passkeys. An agent pays on a customer's behalf while remaining compliant with Strong Customer Authentication. This is the point at which the demo becomes a product.
• Opening Remarks on Sound Practices for Artificial Intelligence – Michelle W. Bowman, Federal Reserve Board
  The first Fed position on the Financial Stability Board (FSB) consultation on responsible AI adoption. Her line: supervisory intensity should follow the materiality of the use case rather than the size of the institution.
• JPMorgan builds AI agents that beat the 60/40 portfolio in backtests – Bloomberg
  An excess return of 0.7 percentage points a year over a 20-year backtest, at lower volatility. A backtest is not a track record. That a bank's research desk publishes this at all still says something about the pace.
• 6 months to live for open models – Nathan Lambert, Interconnects
  Nathan Lambert sees a narrow window in which planned US regulation could squeeze open models out of the market, and reads the push as regulatory capture by a handful of vendors. If your data-residency plan rests on open weights, this scenario deserves a read.

“A central element of managing AI risks is understanding the specific use cases.”