The Agents Have Email Now
An AI agent inside Google Cloud got its own Gmail address this week. Not borrowed, not shared with a human owner — its own account, with its own calendar and Drive to match. That is the quiet pivot around which almost every other story bent.
This week: Google Cloud gave an AI agent its own Gmail, calendar and Drive accounts. Anthropic cut Claude Haiku 5.5 to US$0.10 / US$0.50 per million tokens — a 90% price drop. Meta and Sierra published a draft Personal Agent Protocol backed by Walmart, Stripe and Shopify. In Putrajaya, Budget 2027 cut MSME income tax by one percentage point and handed MDEC RM30 million to onboard 4,000 small businesses to AI. The RM10,000-per-transaction e-invoice rule stays live.
Here's what happened, why it matters, and what your business should do about it.
The Big Three
1. Google's agent moved out and got its own address
At Gemini at Work on 8 October, Google Cloud launched a universal Gemini agent — one provisioned with its own Gmail, its own calendar and its own Drive inside the enterprise directory. Not a chat window pretending to be a colleague. A directory entry that can be invited to meetings, cc'd on threads, filed with, and, crucially, held to account.
The shift here is that an agent now has a mailbox the way a vendor or a part-time contractor does. You can review what it did by looking at its sent items. You can revoke its access the way you offboard a staff member. For the first time, the control vocabulary that already exists in your business — accounts, permissions, reviews — maps directly onto the AI itself.
So what for a business owner: This is how AI becomes manageable. If your team already treats WhatsApp numbers and shared Gmail inboxes as "people," you have a working mental model for the next hire. Expect this pattern to spread fast — to agents from Microsoft, Salesforce and the local service providers who repackage them. Decide now, before any of them land, what your rules are for an agent login: who provisions it, who signs off, who audits it, and what it may never send without a human in the loop.
2. Anthropic cut Haiku 90%, and the price floor fell again
On 7 October, Anthropic released Claude Haiku 5.5 at US$0.10 per million input tokens and US$0.50 per million output tokens — roughly a 90% cut from the previous Haiku price (for prompts under 100,000 tokens). That is one-fifth of what Haiku cost at the start of the year. It also lands exactly level with GPT-6 Luna, meaning two of the three major labs are now competing on the same bottom line.
The context, per Wall Street Journal data this week citing OpenRouter: Anthropic's share of enterprise spending among 120,000 companies fell from about 75% in January to roughly 50% in September. Haiku 5.5 is a direct response. Meanwhile the build-out side of the market tells the opposite story — SpaceX is seeking about US$40 billion for Nvidia chips, Broadcom is arranging over US$50 billion for OpenAI's custom silicon, and SpaceX's five-year credit-default-swap spread hit a record 194 basis points. Retail prices down, debt up. Both are true at once.
So what for a business owner: For small volumes — a receipt reader, a WhatsApp triage bot, a Shopee description writer — AI has effectively stopped mattering as a line item. If a quote still charges you per-call AI fees at last year's rates, that is a markup, not a cost. Ask your vendor what their per-token cost actually is this month; if they refuse to say, that is your answer. Treat the lenders' nerves as background weather — prices do not have to retrace just because the bond market is nervous.
3. Walmart, Stripe, Shopify, and the shopping agents' handshake
Also this week, Meta and Sierra published v0.1 of the Personal Agent Protocol, a draft open standard for how AI shopping agents talk to merchants. Backers at launch include Walmart, Stripe and Shopify. Amazon, pointedly, blocked Meta's own Muse agent the same week. Separately, startups Hark and Tab emerged with personal agents designed to operate the user's own computer and calendar on their behalf, pitching "trust" over raw capability.
The protocol itself is unfinished — the spec is v0.1 and a final release date has not been given. What matters this week is who signed on. When Stripe and Shopify back one standard, the merchants plugged into them — including a long list of Malaysian stores — get the plumbing for free, whether they asked for it or not.
So what for a business owner: If you sell online, a slice of your traffic will arrive as an agent rather than a browser within 12 months. That agent will not see your pretty photo galleries. It will read your structured data: product name, price, stock, SKU, variant, dispatch time. Clean that data this quarter. The merchants who win agent traffic will be the ones whose feeds parse cleanly.
Closer to Home: Malaysia
Budget 2027 was tabled on 9 October. The core SME-relevant measures:
- MSME income tax drops by one percentage point — 14% on the first RM150,000 of taxable income and 16% on RM150,001 to RM600,000, from YA 2027. The government projects around 300,000 MSMEs could save up to RM6,000 each.
- MDEC gets RM30 million to help 4,000 MSMEs adopt AI and automate operations. This is the first time MDEC's grant line has been explicitly earmarked for AI adoption rather than general digitalisation.
- BNM SME fund doubled to RM10 billion for businesses affected by the West Asia conflict; total SME financing and guarantees raised to RM57 billion.
- Accelerated Capital Allowance extended to 31 December 2030 for local plant, machinery, ICT equipment and software — including the kind of hardware and software a small business would buy to run AI work in-house.
- Minimum wage rises to RM2,000/month in June 2027, with MSMEs under RM50 million turnover exempt (duration of the exemption not yet announced).
- E-invoicing: no changes. The Phase 4 (RM1–5M) grace period remains to 31 December 2027, with penalties of RM200 to RM20,000 per invoice from 1 January 2028. By end-August 2026, over 265,000 taxpayers had submitted more than 1.8 billion e-invoices, and the RM10,000 per-transaction rule still applies to every business, exempt or not.
That last line is the one most owners get wrong. The grace period is not an off switch. If a single transaction crosses RM10,000, a Phase 4 business must still issue an individual e-invoice today. LHDN's February 2026 report of 500,000+ non-compliant cases and RM1.4 billion in unreported income has not been retracted — audit risk is live during the grace period, not after it.
Singapore note: No regionally material Singapore SME story broke this week.
What This Means for Your Business
- Pick one AI task to move to Haiku 5.5 or Luna this month. The high-value candidates: product-description generation for Shopee/Lazada/TikTok Shop listings, WhatsApp message triage, invoice line-item extraction, customer-review summarisation. At US$0.10 per million input tokens, a shop processing 2,000 WhatsApp messages a day costs cents per month, not ringgit. If your current provider cannot match this price within 30 days, start switching.
- Write your "agent policy" before an agent asks for a login. One page, no legalese: which systems an agent may read, which it may write to, who approves its account, how often it is reviewed, and what it may never do without a human signing off — refunds above X ringgit, payments to a new vendor, changes to a bank account on file. Draft it this month, while there is no agent in the room to argue with you.
- Audit your product data for agent legibility. If you sell on Shopee, Lazada, TikTok Shop or your own site, export your current listings. Make sure each has a clean name, consistent variants, accurate stock, and dispatch time in a structured field. Fix the top 20 earners first. A cleaner catalogue is also a precondition for anything you want to plug onto MDEC's new RM30 million AI grant line.
The Practical Question
If an AI agent walked into your business tomorrow and asked for a Gmail address, a login, and a job description — what would you give it?
Not a joke. In 2027, this will be an HR conversation, not a tech one. The firms that answer well will be the ones that already know which three tasks they would hand it, which five systems it may touch, and which human signs off on its work. The firms that don't will either hand it everything by accident, or refuse it entirely and lose the productivity gain. The middle path is the one worth planning now, this week, while you can still think about it in peace.
At The Empyrean, we help Malaysian SMEs find the practical, repeatable tasks where AI delivers value without disruption. If you're not sure where to start, we're happy to take a look at your operations and tell you honestly what would make sense.