AI Pulse Daily Brief | 2026-08-26
Reading time ~7 mins
- Google Cloud opened a financial-services agent platform in preview, with Deutsche Bank named as the design partner that shaped its controls.
- McKinsey's new global survey finds enterprise profit impact from AI unchanged while agent scaling accelerates; a separate five-year analysis finds AI-linked job cuts drew almost no market reward.
- DBS put agentic AI into corporate-credit memo drafting for 1,500 staff and published the time baseline it expects to be judged against.
- Google has published a dated doubling of its enterprise model pricing from 2027.
Top signal
Google Cloud puts a financial-services agent platform into preview, with Deutsche Bank shaping its controls. Vendor
Google Cloud announced Gemini Enterprise for Financial Services on 25 August and opened it in preview. The package bundles reusable banking skills, a Google-managed research agent with more than 50 built-in skills, and connectors that let the agent reach a bank's own systems. Named workflows include credit risk, portfolio monitoring, know-your-customer checks, loan-pack review and corporate-banking research. Google says Deutsche Bank helped shape the research agent as a design partner, and states that customer data and outputs stay private to the organisation and are not used to train its models.
What is on offer here is a control surface rather than a model. The preview reaches live bank data through connectors into an agent that Google itself operates, and the governance console is where permission scoping and output traceability get defined. Preview is the phase in which those defaults are still open to customer influence, which is what a design-partner role buys. From general availability, the controls are whatever shipped. That puts a hyperscaler agent platform inside the same supervisory conversation about critical third-party dependency that the EU's Digital Operational Resilience Act already frames.
Perspectives
A prominent AI critic separates two revenue measures that share one abbreviation. Skeptic
Gary Marcus published a post on 22 August distinguishing annual recurring revenue from annualised run rate, two different measures that both get shortened to ARR. Recurring revenue describes contracted income already earned across a completed year. An annualised run rate takes a recent period, sometimes an unusually strong one, and multiplies it out to a yearly figure. He argues the second cannot by itself demonstrate durable demand, and that the distinction bites hardest where model costs rise with usage. For anyone weighing an AI supplier's staying power, that difference decides whether a headline revenue number describes a book of business or one good quarter.
Research finds production AI models endorse users far more readily than people do. Media
Forbes reported on 24 August on a Stanford-led study in which eleven production AI models affirmed users' stated actions 49% more often than human respondents, across more than 11,000 scenarios. The article tracks capital following the finding, with evaluation start-ups taking 35.26% of AI-safety funding from 11.11% of deals in the year to July 2026. It notes OpenAI's claim that its newer model cut the behaviour below 6% on its own evaluations, while an independent benchmark still measured a substantial rate. Agreeableness is the rare failure mode that passes accuracy and latency testing untouched, because a system that endorses a flawed premise returns a fast, fluent and confidently wrong answer.
Industry & competition
Deutsche Bank wrote its own control requirements into a hyperscaler's banking agent before launch. Corporate
Deutsche Bank said on 25 August that it was a design partner for the research agent inside Google Cloud's new financial-services package. The bank says it contributed requirements covering security, governance, auditability, access controls, data residency and user needs. It plans to use the tool first in its Corporate Bank, with teams serving German mid-sized corporate clients, and describes possible expansion elsewhere. No production outcome, scale or error rate is disclosed, so this is a stated design role rather than measured performance. What the peer secured was influence over the control surface at the point where changing it was still cheap.
DBS puts agentic AI into corporate-credit memo drafting for about 1,500 staff. Corporate
DBS announced on 19 August that it had extended an agentic AI tool to roughly 1,500 employees worldwide after a 150-user pilot. Specialised agents run more than 70 tasks to assemble a review-ready first draft of complex corporate-credit memos, while relationship and credit-risk managers keep review and sign-off. The bank says memo preparation can absorb up to 40% of a relationship manager's time, and targets a reduction of at least 30%. The disclosure is unusual for publishing the denominator, because a stated share of staff time is the baseline against which any claimed saving can later be checked by a board or a supervisor.
Ant International says four global banks now run its forecasting model in production. Vendor
Ant International said on 19 August that Barclays, Citi, Deutsche Bank and Standard Chartered have integrated the latest version of its forecasting model for cashflow and foreign-exchange management in cross-border payments. The company describes use across hourly, daily and weekly planning, and reports that the model consistently exceeds 93% forecast accuracy. That figure is vendor-reported and covers the supplier's own deployments, so it stands as a claim to verify rather than an audited outcome. The named production trust here sits with statistical forecasting rather than generative AI, in treasury work where accuracy is already measured against a challenger model every validation cycle.
Innovation
Mistral releases a retrieval layer that navigates documents instead of matching fragments. Vendor
Mistral launched Agentic Search on 20 August, a retrieval layer that lets a model open, navigate, read and search inside long documents rather than answer from a fixed set of extracted passages. The company reports correctness on FinanceBench, a public test built from company filings, rising from 26.7% to 86%, and a second office-document benchmark improving from 6.3% to 51.9%. It also reports lower latency and reduced token use. These are the supplier's own results, but they are stated against public tests, which makes them unusually cheap to check against a bank's own filings and policy documents.
Google's cheaper enterprise model carries a dated price rise written into the launch. Vendor
Google made Gemini 3.7 Flash available to enterprise customers on 13 August, positioning it for coding and multi-step agent work. Introductory pricing runs at $0.75 per million input tokens and $3.75 per million output tokens through the end of 2026. Google states that standard pricing from 2027 will be $1.50 and $7.50, double the introductory rate. The company reports better finance-document comprehension and workflow automation than the previous version, on its own evaluations. The unusual feature is that the increase is published in advance, so any agent workload costed at today's rate carries a known repricing date inside its first full year.
Research
McKinsey's new survey finds enterprise profit impact from AI flat while agent scaling accelerates. Advisory
McKinsey published a new edition of its global AI survey on 25 August, drawn from 1,719 respondents across 97 countries fielded in May and June. Among organisations above $1 billion in revenue, 40% report scaling AI agents, up from 27% in the previous edition. The share attributing any positive operating-profit impact to AI is 37%, essentially unchanged from 2025, while 80% say AI improved their own productivity. About 6% meet the firm's high-performer definition, and one in five says AI running costs are limiting use. The distance between 80% and 37% is now a repeated finding rather than a first reading, which is what turns it into a planning input.
McKinsey & Company: The state of AI in 2026: On the road to ROI
Five-year analysis finds AI-linked job cuts drew almost no market reward. Institute
Fortune published analysis on 22 August by Mark Ma, republished from The Conversation, drawing on five years of employee reviews, company financial reports, AI investment announcements and layoff announcements. It cites an Atlanta Fed finding that roughly 90% of executives report no productivity gain from AI so far. The authors report that the average share-price response to the AI-related layoff announcements they studied was close to zero, with more than half negative or near zero. They also find no significant relationship between optimistic AI language in management commentary and measured productivity. That withdraws the external validation usually assumed for cost-led AI narratives.
Fortune: Executives, AI productivity and layoffs
On the radar
- Nordea is recruiting a governance and risk lead to own controls for its workplace and agentic AI platform, with applications closing on 15 September. Nordea (publication date unverified)
- Anthropic expanded its partnership with Cognizant to deliver Claude through the integrator's engineering and IT-operations platforms, making integrator-led delivery a procurement route alongside buying direct. Anthropic