AI Pulse Daily Brief | 2026-08-20
Reading time ~5 mins
Lloyds has put 100 million pounds of AI value on the public record for the first half of 2026, with no methodology behind it. OpenAI is previewing a way for a customer to keep a no-retention contract while the vendor still monitors for misuse, with a September rollout. Amazon has made central control over which websites an agent may read enforceable on its own servers, and available from Ireland. The US standards body is organising AI security around three control areas and naming agent identity and supplier evidence as the open gaps.
Industry & competition
Lloyds says AI produced 100 million pounds of value in six months, without showing the workings. Media
American Banker reported on 19 August that Lloyds Banking Group has attributed 50 million pounds of value to AI in 2025 and 100 million pounds in the first half of 2026. The same article states that Lloyds has released no supporting detail behind either figure. Its recommended alternative is to define the outcome of a single workflow, trace the observed result back to the AI that produced it, and add only evidenced results into an enterprise total. A peer number disclosed without an audit trail still becomes the reference point the bank's own AI value is read against in the next planning round, before either figure can be checked.
A US grocer reports bigger baskets from an AI assistant that completes the task, not one that answers questions. Media
The payments trade publication PYMNTS reported on 17 August that Albertsons, a large US supermarket chain, saw average order value rise 26% among early users of a new AI assistant. The assistant builds recipes and shopping lists around a customer's dietary preferences, and customers using only conversational search grew average order value 10%. No denominator, control group or experimental design was published, and early users self-select toward higher engagement, so 26% is a ceiling rather than a planning baseline. The transferable number is the gap between the two designs on one customer base: completing a task earns more than answering a question.
An ABN AMRO governance specialist publicly questions how closely a frontier AI agent is watched during testing. Corporate
Andrew Harrison, who works on agentic and generative AI governance at ABN AMRO, posted on 19 August about OpenAI's stated approach to supervising its most advanced AI agents in testing. He described OpenAI as spending a fifth of the computing power that runs the model on monitoring it, and asked whether that is enough. He also questioned a process in which a human is paged within 30 minutes, then given a further 30 minutes to declare a false alarm or pause the test. He wrote as an individual and made no claim about his employer's policy. Neither the monitoring share nor the human response time appears in a standard model-risk questionnaire, and a Dutch peer's governance staff are now asking for both in public.
Andrew Harrison on LinkedIn (LinkedIn; original source not verified)
Innovation
OpenAI previews a way to keep customer data out of its hands while still watching for misuse. Vendor
OpenAI said on 19 August that it is testing Private Safety Processing with a set of early customers. The feature is meant to let a customer keep a zero data retention agreement, the contract term that stops a vendor storing prompts and responses, while OpenAI still looks for harmful patterns across related requests. Customer content stays on the customer's own infrastructure or is encrypted with the customer's own keys, and OpenAI staff cannot read it, according to the company. A technical paper and a wider rollout are planned for September, so nothing a reviewer could test has been published yet. That retention term is what currently keeps longer-running frontier-model work out of scope in a regulated bank, and it is the constraint this preview is aimed at.
Amazon lets a central administrator fix which websites an AI agent may read, and offers it from Ireland. Vendor
Amazon Web Services said on 19 August that its enterprise AI agent platform can now filter web searches by website and by publication date, applied on Amazon's own servers rather than by the agent. An administrator sets the allow and deny rules centrally, and the workflow calling the service can narrow them further but cannot widen them. Amazon has also made the capability available from its Ireland region and says the search traffic stays inside its own network. The precedence rule is the part worth copying: a limit a workflow owner cannot loosen is one that survives an audit, and setting it centrally costs far less than retrofitting it later.
A US card issuer is moving its financing and rewards inside an AI shopping checkout. Corporate
Synchrony, a large US consumer-credit and store-card lender, announced on 17 August a collaboration with OpenAI to place financing, rewards and loyalty inside AI-driven shopping and checkout journeys. It also said it will deploy OpenAI's latest models across the company through several enterprise tools. The announcement reports no live volumes, no production outcomes and no description of the controls around a purchase an agent completes, so it is stated intent. Enterprise model rollouts are now ordinary at large financial firms; supplying the credit at the moment an agent completes a purchase is a claim on the point of sale. That is the position from which a card issuer is usually displaced rather than out-competed.
Security
The US standards body is organising AI security around three control areas. Authority
The National Institute of Standards and Technology, whose security frameworks are widely used inside banks, has published a workshop report on its draft profile for AI cybersecurity. It splits the field into three parts: securing AI systems, using AI in cyber defence, and defending against attackers who use AI. The report names the control gaps participants want closed, among them a verified identity and bounded access for each AI agent, and supplier evidence of what sits inside a model. It also records that no guardrail resists every attempt to trick a model, so testing has to be continuous rather than a one-off acceptance check. The profile carries no adoption date, and agent identity and supplier evidence are the two items with no settled owner between security, procurement and model risk.
National Institute of Standards and Technology (publication date unverified)