Aug 10 – Sep 9, 2026 · 5 items
The AI week, for an accountant
A real edition, written for: Accountant
The one thing
OpenAI's newest model works inside spreadsheets and browsers, which is where your manual bookkeeping hours sit — but check every line it touches.
OpenAI's new GPT-6 Astra can drive spreadsheets and browsers itself
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 pay for ChatGPT, and this model works inside spreadsheets and web forms. That is where your bookkeeping hours actually go.
- Enterprise Daybreak customers go first, with paid ChatGPT tiers following in coming days
- Client data in a browser session deserves the same care as your accounting file
- Treat anything it touches in a client ledger as a draft you check line by line
Try this
Next month-end, have ChatGPT reconcile one client's small bank export against the ledger and list mismatches.
Paste this into your AI tool
Here is a bank statement export: [paste rows]. Here is the ledger for the same period: [paste rows]. Match each bank line to a ledger entry. Output three lists: matched, in the bank but not the ledger, in the ledger but not the bank. For unmatched items, suggest the likely reason. Do not guess amounts you cannot see.
Microsoft's new transcription model drops audio to about a dime an hour
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 calls and year-end meetings turn into notes you write up by hand. Cheap transcription with speaker labels removes that step.
- Covers 60 languages and handles mid-sentence switching, useful for German, French and English client calls
- Verbatim mode keeps filler words, which matters when notes support a compliance file
- Access runs through Microsoft Foundry, so this is a tool to trial rather than switch on today
Try this
Record your next client year-end call, then draft the follow-up note from the transcript instead of your memory.
Paste this into your AI tool
Here is a transcript of a client meeting: [paste transcript]. Write a follow-up note with three sections: decisions made, information the client still owes me, and actions on me with dates. Quote the transcript where a figure or deadline is mentioned. List anything ambiguous separately rather than resolving it.
A quarter of firms delayed or killed AI projects over surprise bills
A survey by AI cost management firm Mavvrik, with metrics firm Benchmarkit, found that unexpected AI bills are altering business decisions at most organizations, prompting spending freezes, board escalations and canceled projects. Usage-based charges for model processing were the most cited surprise cost. Forecasting accuracy has worsened since last year, and developer AI tools often sit outside cost reporting.
Why it matters for you
Usage-based AI billing is what caught these firms out. At your size a single month's overage lands on your own P&L.
- Metered token charges were the most cited surprise cost, not subscriptions
- Forecasting accuracy got worse year on year, so early estimates are not trustworthy
- Set a monthly cap for AI spend per client engagement before you scale any workflow
Try this
Write down what you spent on AI tools last month and set a hard monthly ceiling for yourself.
Sources
Gartner: only a quarter of AI customer-service uses show a return
Gartner reviewed 432 uses of AI in customer service and found only a quarter produce a measurable financial return, while a quarter lose money and most of the rest cannot be judged either way. Support teams are nonetheless expanding spending, and more than half of leaders expect their bonuses tied to AI results in 2026. Headcount is rising as often as it is falling, undercutting the savings case.
Why it matters for you
You are choosing where to point AI across bookkeeping, VAT and payroll. This is evidence to spend that time on measurable work.
- A quarter of the cases reviewed actually lost money, and most could not be judged either way
- Repetitive, checkable tasks like reconciliation and data entry are easier to measure than client-facing chat
- Pick one client and time the task before and after, or you will never know
Try this
Time yourself on one client's VAT return prep this month, before you change anything.
EU AI Act disclosure duties are now enforceable
The European Union's AI Act moved into enforcement on 2 August 2026, with disclosure duties now applying to chatbots and to content produced by AI. The EU's AI Office can request information from covered companies and ask for access to their models, though it has not yet pursued anyone for misconduct. Anthropic, Google, Meta, OpenAI and Microsoft have each described compliance steps, including watermarking generated text. Rules for high-risk uses such as education, biometrics and migration arrive only in December 2027 and August 2028.
Why it matters for you
You are in Switzerland, so the Act does not bind you directly. It does bind clients selling into the EU.
- Disclosure now applies to chatbots and AI-generated content shown to EU users
- The EU AI Office can request information and model access from covered companies
- High-risk rules only arrive in December 2027 and August 2028, so there is planning time
Try this
Flag which of your clients sell into the EU and use a chatbot or AI-generated content.
Sources
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.
One line makes a model name the weak point in the answer it just gave, which turns a confident reply into one you can actually check.
Five minutes to set up
Name the single assumption your answer depends on most. If it is wrong, what changes? Answer in three sentences, no restating your answer.
- Ask any question you would normally accept the first answer to, then paste the line.
- Read the assumption it names and decide whether it is true for your situation.
- When the assumption is wrong, tell it so and ask for the answer again on the corrected basis.
- Keep the line somewhere you can paste it, because it works on any answer from any chat.
Yours arrives Thursday.
This one was written for a accountant. Tell us what you do and the next one is written for you — same news, your job, once a week.