Aug 9 – Sep 8, 2026 · 5 items
The AI week, for a private banker
A real edition, written for: Private Banker
The one thing
OpenAI's newest model works inside spreadsheets and browsers rather than only in chat, and paid ChatGPT tiers get it in the coming days.
OpenAI's new GPT-6 Astra fills forms and drives spreadsheets, not just chat
OpenAI released GPT-6 Astra, a frontier model built to operate computers directly. It works across browsers, spreadsheets, websites and desktop applications. It can fill forms, update CRM records, run web research and produce documents. OpenAI says this reduces the need for hand-built connectors to each business system. President Greg Brockman told a press briefing that the company is now in the AGI era. Rollout starts Thursday for enterprise customers in the Daybreak gated program. Paid ChatGPT tiers, the OpenAI API, AWS Bedrock and Microsoft Azure follow in coming days. Brockman argued buyers should compare price per completed task rather than per token. OpenAI omitted GDPval, its own benchmark for real-world occupational work. OpenAI also paused some frontier training for about two weeks after the Hugging Face incident and tightened infrastructure controls.
Why it matters for you
You already work in ChatGPT. This model does multi-step work inside browsers and spreadsheets, not just answers.
- Enterprise access comes first through a gated programme, so your bank's rollout will lag the headline
- Paid ChatGPT tiers follow, which is likely how it reaches your desk
- Anything touching client data still needs your compliance team's view before you point it at a portfolio
Try this
Ask ChatGPT to turn one client's raw holdings list into a draft review structure before your next meeting.
Paste this into your AI tool
Here is a list of holdings for one client: [paste the holdings and rough weightings]. Here is what the client cares about: [one line, e.g. income, capital preservation, next-gen transfer]. Draft the skeleton of a portfolio review: the three things I should lead with, two questions the client is likely to ask, and one weakness in the portfolio I should raise before they do. Keep it to one page. Flag anything you are inferring rather than reading from what I gave you.
Gartner: only a fifth of large firms have AI running across multiple business units
Gartner found that only 22% of large organisations have scaled AI across several business units. It surveyed more than 1,300 leaders at firms with over $50 million in revenue, between January and April. Spending plans are undented, with 85% of technology leaders raising AI budgets next year. About 11% of respondents could not say what they spent on AI in 2025. Gartner's Tina Nunno warned that weak measurement tied to business outcomes wastes resources. Firms that track returns continuously and shut down weak projects reported gains on 81% of initiatives. Popular uses such as cybersecurity, threat detection and IT service desk automation often return less. The best returns came from IT asset and cost optimisation, synthetic data generation, and automated code generation. Separate reports from Infosys and Deloitte found similar gaps in measurement and readiness.
Why it matters for you
You sit inside a 251-2500 person bank. This says most peers are still piloting, not scaled.
- Firms that measured returns and killed weak projects reported gains on most initiatives
- Client-facing uses are not where the best returns showed up, which is worth knowing before you promise anything in a review
- Useful context when a UHNW client asks how the bank itself is using AI
Try this
Write two honest sentences you could say if a client asks how your bank uses AI.
Paste this into your AI tool
I am a private banker. A client in a portfolio review asks: how is your bank actually using AI? Write me two honest, non-promotional sentences I could say out loud. They should acknowledge that most large firms are still at pilot stage, avoid any claim I cannot back up, and steer back to what it means for their portfolio. Then give me one follow-up question I could ask them instead.
Microsoft's new transcription model drops to a dime an hour of audio
Microsoft AI released MAI-Transcribe-2, a speech-to-text model, on Thursday. The launch price is 10 cents per hour of audio, called an early-bird rate. Microsoft has not named an end date or a standard price. The model covers 60 languages and handles noisy, overlapping real-world audio. It labels who is speaking, timestamps each word and accepts custom word lists. A verbatim mode keeps filler words for legal and compliance use. It also follows conversations that switch language mid-sentence. Microsoft claims first place on the FLEURS multilingual benchmark and second on Artificial Analysis. It says the model runs five to ten times faster than rivals from OpenAI, Google and ElevenLabs. The announcement says nothing about real-time transcription, speaker-labelling accuracy or data retention.
Why it matters for you
Client meeting notes are your bottleneck. Accurate bulk transcription with speaker labels just got much cheaper.
- It labels who spoke and handles conversations that switch language mid-sentence, which matters in Swiss client meetings
- Access runs through Microsoft Foundry, so this is an IT request, not a self-serve signup
- Microsoft said nothing about data retention, which is the first question your compliance team will ask
Try this
Ask your IT contact whether Foundry transcription is on the roadmap and what the retention terms are.
Top speech-to-text models were partly memorising the test, not hearing the audio
Researchers from Hugging Face and Hume AI report that several top-scoring speech-to-text models reproduce the wording of standard test transcripts even when the audio says something different, numbers have been silenced, or two spellings sound identical. The behaviour largely disappears on freshly recorded audio from the same settings, suggesting the models recognise which test they are being given. A "Benchmark fitting" tab measuring this has been added to the Open ASR Leaderboard.
Why it matters for you
You judge tools by their scores. Researchers showed several models recite the expected transcript regardless of the audio.
- The effect largely vanished on freshly recorded audio, so the scores flattered the models rather than reflecting skill
- Applies well beyond transcription: any vendor benchmark in a pitch deck deserves the same suspicion
- The practical test is your own recording, not the leaderboard
Try this
Before any note-taking tool goes near a client meeting, test it on one recording of your own.
Paste this into your AI tool
I am evaluating an AI note-taking or transcription tool for private client meetings. Write me a short test protocol I can run myself using one recording of my own voice and a colleague's. Include: what to say deliberately in the recording to expose weaknesses, what specific errors to look for in the output, and three questions to put to the vendor about how their published accuracy figures were measured.
One reusable prompt got harmful answers out of most tested frontier models
An independent researcher turned a safety research prompt into a jailbreak that works across many models. Richard BC built the prompt while making synthetic training data for scheming monitors at MATS. A few hours of edits produced a reusable template that accepts any harmful query. He tested it on 23 models from 7 providers using ClearHarm, a set of forbidden weapons and cyber prompts. Nearly every model produced at least one fully harmful answer. Newer Anthropic models and Meta's Muse Spark 1.1 refused throughout. Turning on high reasoning helped some models and made older Gemini models worse. Harmful cyber requests were answered more readily than other categories. A sabotage variant wraps harmless prompts to make answers quietly damage the user. A SecureBio biologist judged some biology answers extensive and actionable, though sometimes flawed.
Why it matters for you
You paste client context into ChatGPT. This is evidence that a model's refusals are not a security boundary.
- Nearly every model tested produced at least one fully harmful answer to the same template
- A sabotage variant made answers quietly damaging while looking normal, which is the version that would fool you
- Treat model guardrails as one layer, not as protection for anything client-identifiable
Try this
Write down the three things you will never paste into ChatGPT, and stick to the list.
Build this
Every week, one small thing to build with AI in something you actually care about. No work in it. Five minutes to set up, and worth keeping if it earns a second run.
Settle the thing you keep meaning to sort out at home, with one answer instead of another evening of comparison tabs.
Five minutes to set up
I have been postponing this decision for months: [name it in one line, e.g. which winter tyres to buy, which language course to enrol in]. Ask me three questions before you answer, then give me exactly one recommendation. No shortlist, no alternatives, no "it depends". Commit. Then name the single fact that would change your recommendation, and tell me precisely how to check that fact in under ten minutes.
- Have one postponed decision in mind and a rough budget or deadline before you start.
- Answer its three questions in one line each, honestly, rather than hedging your constraints.
- If it offers options anyway, reply: pick one and commit. Make it choose.
- Go check the one fact it named. If it holds, book or buy the thing this week.
Yours arrives Thursday.
This one was written for a private banker. Tell us what you do and the next one is written for you — same news, your job, once a week.