AI Pulse Daily Brief | 2026-08-13
Reading time ~3 mins
Fortune finds technology chiefs rationing AI access by role and watching the bill daily.
ConnectOne Bank reports commercial-lending document searches falling from twenty minutes to thirty seconds.
Amazon publishes a working control pattern for payments that AI agents authorise on their own.
A Berkeley study of 24 organisations finds only two using a control group to prove AI value.
Perspectives
Technology chiefs are replacing blanket AI access with tiered entitlements and daily spend monitoring. Media
Fortune reported on 12 August that executives who first opened AI tools to everyone are now rationing them by how much a role actually needs. The fleet-software firm Samsara tracks its AI bills daily, caps some non-technical staff, and leaves room for research teams to experiment. Docusign narrowed how much of its codebase its coding assistants read by default, and says that nearly halved the AI processing it pays for. The same narrowing that cut the bill also limits how far an assistant reaches into company data. For the bank, AI spend control and agent data-access control turn out to be the same design decision.
Industry & competition
ConnectOne Bank says an AI assistant cut commercial-lending document searches from twenty minutes to thirty seconds. Media
American Banker reported on 14 July that ConnectOne Bank piloted an AI assistant from its lending-software provider nCino across commercial-lending operations. ConnectOne said document search fell from about twenty minutes to about thirty seconds, and a separate agent cut document-update time by 60%. Management presented the result as capacity for relationship managers to serve more clients rather than as a headcount reduction. Both figures measure elapsed time on one administrative step, with no reported link to files handled per banker or to credit turnaround. That is the shape of evidence a lending business case in the bank would be asked to defend at renewal.
Innovation
Amazon publishes a working design for letting AI agents authorise payments with an audit trail. Vendor
Amazon Web Services published on 12 August a reference build in which every agent-initiated payment passes a policy check before it goes through. The component making that decision runs in tamper-sealed hardware, and each transaction leaves a record that can be replayed afterwards. Amazon says the full sequence completes in under four seconds. The example customer is a crypto firm settling on a public blockchain rather than a bank, and no supervisor has accepted the design. Even so, a major cloud provider has now set a public bar for what evidence should exist when software commits money on its own.
Research
Berkeley study finds most organisations cannot show what their AI licences are actually worth. Institute
California Management Review published Beyond AI Licenses on 5 August, based on a questionnaire and interviews with 24 people responsible for AI adoption or measurement. Ten of the 24 said the main barrier to measuring financial impact was that no success metric had been agreed before purchase. Only two used a control group, while sixteen still treated self-reported time savings as their primary proof of value. The sample is small and skewed to smaller US firms, so it establishes a failure pattern rather than its prevalence in banks. For a business line whose AI tools come up for renewal this year, the study names exactly which evidence will not carry the decision.
California Management Review: Beyond AI Licenses: How to Measure Whether Your AI Is Actually Working