This Week in AI: November 9–16, 2026 — Your Business Moved Into the Chat Window, and the Phone Agent Learned to Remember
This Week in AI: November 9–16, 2026 — Your Business Moved Into the Chat Window, and the Phone Agent Learned to Remember
What changed in AI this week that a small business owner should actually notice? On November 10, OpenAI gave businesses a way to claim their own listing inside ChatGPT and correct the price it quotes. The same day, a Toronto phone company put its voice agent into full production. And on November 12, a survey of 1,731 owners put a hard number on a gap most shops will recognise: two-thirds now use AI, and only one in four has a single written rule for it.
I run NexAIS (https://www.nexais.ca/) in Richmond Hill. We put AVA, our AI receptionist, on small businesses' phone lines, so I read a week like this looking for one thing: where will a customer meet your business next, and who will be answering when they do? This week the answer moved. It is no longer only your phone and your website. It is also the chat window your customer opened instead of searching.
1. OpenAI put businesses inside ChatGPT on November 10 — claim the listing before you worry about the agent
On November 10, OpenAI opened Businesses in ChatGPT in beta. An eligible business can claim its profile in the new Business Hub, verify it, and supply its own product details, including prices and stock. If you sell through Shopify, you can connect the shop from Shopify or from the Hub. Once there is enough activity, the Hub shows you mentions, website clicks, and the subjects people were asking about when ChatGPT pointed them your way.
There is a second piece, and it needs a cooler head. Business Agents — an agent that answers questions about your offering inside ChatGPT — are open only to eligible ecommerce and online service businesses in the United States for now. Claiming and correcting your profile is the wider job, and it is the one to do first.
Here is why. If your website says $49 and ChatGPT tells a customer $39, the customer does not blame the model. They bring you the screenshot. The fix is boring and it works: take your best seller, check the price, check which version is actually in stock, and check the page a stranger lands on. Then try the awkward version — the colour that is sold out, the size that costs more — and see whether a stranger could still tell what they are buying.
Three days later, on November 13, OpenAI added ChatGPT Wallet: a small first rollout, through its dots agents, for US Pro subscribers with eligible US cards. The buyer still reviews the items, the total, and the card, and approves each purchase. Put Wallet in the diary, not on this week's list. Fix the listing first. A checkout that quotes the wrong price faster helps nobody.
2. The phone agent moved into the phone system you already have — November 10 to 12
On November 10, Toronto's Sangoma put Sonella, its voice agent, into production after a pilot earlier this year. Sonella answers calls, books appointments, routes callers to the right department, and opens support tickets. Sangoma is not a startup with a demo page. It sits behind more than 2.7 million business phone seats at over 100,000 customers, and it is the company that looks after Asterisk and FreePBX, the open phone systems a lot of small offices quietly run on. Its named markets will sound familiar: healthcare, hospitality, retail, professional services.
Read that the plain way. Voice AI is no longer a separate gadget you bolt onto a business. It is becoming a feature of the phone system itself. A day later in the UK, Wavenet launched an AI receptionist for GP practices that works through triage questions and writes the answers straight into the practice's own workflow. Different country, same direction.
The test to hold every one of these to — ours included — came from ElevenLabs on November 11. At its summit, ElevenLabs previewed Memory for returning callers and Routines for follow-up work around a call. Previewed is the word to keep. There is no confirmed date when your account gets either one, so do not promise them to a customer yet. But the demo to ask for is now clear. Ring once about a repair. Ring back, change the collection day, and interrupt halfway through the address. Then check the booking. Did it recognise the right customer, drop the old date, and leave one correct appointment? A pleasant voice that books two vans has still booked two vans.
One diary date from the same week: OpenAI said on November 10 that ChatGPT's Standard and Advanced voice modes retire on December 11, as it puts its weight behind GPT-Live. If your team talks through jobs in ChatGPT away from a keyboard, try one real handover before then — a customer name, a product code, a changed collection time — and check the written result. Those three details are the ones somebody else has to act on.
This is the ground AVA works on every day, and the contractors we talk to most feel it first: an HVAC technician on a roof and a plumber under a sink cannot take the call that pays for the next job. Each AVA is built for one business — its services, its booking questions, its rules for what counts as urgent, and the named people urgent calls go straight to — and she remembers returning callers, with their call and message history, instead of meeting a regular as a stranger.
3. Two-thirds of owners use AI. One in four has a rule for it. — November 12
Gusto published its 2026 State of Small Business report on November 12: 1,731 US owners, all running businesses with 24 or fewer people, surveyed in late August and early September.
67% use AI tools. The split matters more than the headline. 19% are experimenting. 28% use AI for some tasks. 15% use it regularly across several parts of the business. Only 5% say it is built into every part where it fits. Professional services run deepest, at 27% using AI across several functions or the whole business, against 16% in personal services and 13% in community services.
Then the gap. Only 25% of businesses with employees have an AI policy at all — up from 15% last year, which tells you how new this is. And owners are far more comfortable letting AI draft marketing with a light check — about 60% of AI users — than letting it near a customer's direct question.
That gap is not hypocrisy. It is good judgement arriving in the wrong order. The tools reached the customer before the rules did. The fix is one page, written before the next tool goes live: what the system may answer on its own, what it must never promise, and who it passes to when a caller is unhappy, confused, or asking about money. On our phone lines that page is the setup itself. Routine questions and bookings run. Refunds and complaints go to a person, every time.
The rest of Gusto's report explains the mood behind the numbers. 61% of owners are optimistic about their own business, the same as last year. Only 18% say that about the wider economy, down from 29%. 52% raised their prices in the past year. Owners are not buying AI because they feel rich. They are buying it because the phone still rings while both hands are busy.
4. Google taught your office tools to remember how you work — November 9 to 12
Three Google items this week belong together, because they all do the same job: they remember.
On November 9, Google started rolling out reusable skills in Workspace. You give Gemini a set of instructions, examples, and reference files once, and you can call that skill up later inside Gmail, Docs, and Slides. Business Starter is among the plans that get it. If your account is on the Scheduled Release track, your start date is November 30. The first skill worth saving is the reply you type most: "Can you send me a quote?" Give it two replies you would happily send again, the details you need before you can quote, and one thing it must leave blank — the price, until you have worked it out. Test it on an enquiry with the quantity missing. If it asks for the quantity, keep it. If it invents an order, you found that out cheaply.
On November 10, Google widened two Gmail features: an AI Overview that sums up a long thread, and Help me write to draft or tidy a reply. It covers Workspace users with smart features turned on, and Google allows up to 15 days for it to reach an account. Use it on the thread where three people agreed to three different Fridays. Read the summary, find the message where the customer actually confirmed, and check who said yes before you send. A summary that swaps your "we could do Friday" for the customer's "Friday works" puts the wrong sentence in the order book.
On November 12, Google launched Gemini 4 Hydrogen, rolling out in Gemini and in Search. It works across documents, spreadsheets, and images you upload. Give it the job most owners put off: two supplier quotes, one priced by the box and one by the unit, delivery explained on page six. Ask for the delivered cost of a fixed quantity, with pack size, minimum order, and delivery charge shown separately, and the page behind each figure. Leave missing figures blank, and ask it to write the question you need to send the supplier. Either you get a comparison you can check in a minute, or you get the one email that finishes it. Both beat a neat-looking guess.
5. Claude got an office, a deadline tonight, and two honest scorecards — November 9 to 12
Anthropic had the busiest week of the three big labs. On November 10 it released Claude Sheets in beta on paid plans — ask for a spreadsheet and Claude builds the tabs, formulas, and drop-downs, then edits the cells with you, with changes marked, and exports to Excel or CSV when you are done. Whiteboards arrived the same day, also in beta on paid plans: talk through a plan and get an editable board with owners, dates, and a timeline. On November 12, Claude Meetings was announced, rolling out in beta to Team and Enterprise accounts: Claude joins a video call as a visible participant, takes notes, and leaves a transcript and the work it made on the meeting page afterwards. And Claude Data became generally available on November 10 for accounts with a connected source such as a CRM — useful only after your team agrees on one sentence: "A sale counts when..." If your figures live in one spreadsheet, that sentence may be the only part of the launch you need.
Above all of it sits the new model. Claude Fable 5.2 launched on November 9. If you are on a Pro plan or a standard Team seat, there is a clock on it: Anthropic says Fable can use up to half your weekly allowance through tonight, November 16, at 11:59 p.m. Pacific. After that, it draws on paid usage credits if you have them turned on. Do not spend the window testing it on a poem. Hand it a finished proposal and the original brief, and ask it to find anything you promised in one and dropped in the other, with the exact lines from each. One missing revision round is a real saving. That is a better reason to keep a model than any launch chart.
Because the scorecards this week were unusually honest, and they point the other way as well. On the Remote Labor Index — 240 real freelance projects, first done by 358 working freelancers — Fable 5.2 produced work a client would accept on 41.3% of projects, up from 2.5% for the best system a year ago. That is a large jump, and it still means most projects in the test did not meet the standard. The test also leaves out any job that needs the client in the room, which is most of a small business's week. A second study, reported November 8, used 273 tasks taken from real coding sessions: the best model met about 70% of the requirements on an average task and met every requirement on only 26.4% of tasks.
Put those two numbers on the same page and you have the owner's rule for the year: hand AI whole, well-defined jobs, and keep a person's check on anything a customer will see. A quote with a wrong price is not 70% right. It is wrong.
One security note, because it affects shops that tried a tool from last week's roundup. On November 13, Cursor published its findings on an October incident: some users may have received an unauthorised copy of the Claude Code extension from its marketplace on October 29. Cursor says it has contacted the affected customers and team admins directly, and that removing the extension alone may not be enough — the notice carries the recovery steps. If a developer built something for you in Cursor, the question to ask them this week is simple: did your account get that notice?
6. Sorting the queue is the quiet job worth having — November 11 and 13
The last two items are unglamorous, and they may save more time than everything above.
On November 11, Google released its Decisions API. A developer gives it a fixed set of choices or scoring questions, runs it across text, images, audio, video, or PDFs, and gets back an answer with a confidence score. Point it at the inbox queue every small business knows: new quote, existing order, refund, or a person needs to read this now. Test it on last month's messages with the right answers hidden, and count where each one lands — above all, the refund requests it files under sales. A high average score can still leave the person waiting for their money in the wrong queue.
On November 13, Firecrawl said its Alexandria company database now covers more than 75 million companies, with fields such as headcount, funding, and the software a company runs. For anyone who sells a service, the useful question is small: is this prospect the right size, and set up the right way, to need what I sell? Look up five firms you would like to work with, check what the database says against the firms themselves, and spend your research time on the two that fit. A database entry is a lead to check, not a fact to open a pitch with. Congratulating someone on software they replaced last Christmas is a poor first impression.
What this week adds up to
Your business now gets quoted, priced, and answered in places you do not sit in front of: a chat window, a search answer, a phone system with an agent built in. The owners who handle that well are not the ones who turned everything on this week. They claimed the listing and fixed the price at the source. They saved the one reply they type every day as a skill, with the price left blank. They wrote the one-page rule before the next tool went live. And they kept the test that matters on the phone: one caller, recognised on the second ring, with one correct booking at the end of it.
That is how AVA is set up. Our price is flat, in Canadian dollars, with no per-minute counting. Starter is $197 a month (https://crm.nexais.ca/pricing/ai-receptionist): the receptionist on its own. Professional is $399 a month and adds the CRM and the app — that is where restaurants start, because a restaurant needs bookings handled, not just messages taken. Grow is $599 a month and adds website and Instagram management. There is no tier smaller than Starter, and the price does not rise in a busy month.
The trial runs your own AVA, built for your business, on your own line for seven days. It is free, your card is taken at signup, nothing is charged during the seven days, and charging starts on day 8 only if you continue. Hear her first the way your customers would: call the demo line, +1 (437) 529-0400, at noon or at 9 p.m., and ask the questions your callers ask you. A missed call can be a missed customer. This week, your business started being answered in one more place — make sure the answer carries your price, your rules, and your name. Questions about your own setup? Our contact page (https://www.nexais.ca/contact) reaches a person on our team.
Sources
- Lilach Bullock, "AI News This Week for Small Business" (November 15, 2026) — Businesses in ChatGPT beta and Business Hub reporting (Nov 10), Business Agents limited to eligible US ecommerce/online services, Gmail AI Overviews and Help me write expansion (Nov 10), Workspace reusable skills rollout (Nov 9; Scheduled Release Nov 30), Gemini 4 Hydrogen launch (Nov 12), Claude Sheets beta (Nov 10), Fable 5.2 trial terms (launched Nov 9; up to half the weekly allowance on Pro and standard Team seats through Nov 16, 11:59 p.m. Pacific), Firefly Quick Actions (Nov 10), Claude Meetings (Nov 12), Firecrawl Alexandria past 75 million companies (Nov 13), Claude Whiteboards beta (Nov 10), ChatGPT Wallet (Nov 13; dots, US Pro, buyer approves each purchase), Claude Data generally available (Nov 10), ChatGPT Voice Standard and Advanced modes retiring December 11, Cursor incident findings (Nov 13; unauthorised Claude Code extension copy, October 29), Gemini Decisions API (Nov 11), ElevenLabs Memory and Routines preview (Nov 11; upcoming capabilities, no confirmed access date) — https://www.lilachbullock.com/ai-news-this-week-for-small-business-15-november-2026/
- Sangoma, via CloudCommunications, "Sangoma Launches Sonella: Bringing Agentic Voice AI to Businesses Communications" (announced November 10, 2026, Toronto) — Sonella in production after a pilot; answers calls, schedules appointments, routes calls, creates support tickets; healthcare, hospitality, retail, education, professional services; Sangoma reports 2.7 million UC seats and more than 100,000 customers — https://www.cloudcommunications.com/news/sangoma-launches-sonella
- Gusto, 2026 State of Small Business report, via PR Newswire / Morningstar (November 12, 2026) — survey of 1,731 US owners of businesses with 24 or fewer employees (Gusto customers, fielded August 17 – September 7, 2026): 67% use AI tools (19% experimenting, 28% some tasks, 15% regularly across multiple functions, 5% built in everywhere it fits); professional services 27% deep use vs personal services 16% and community services 13%; 25% of businesses with employees have an AI policy, up from 15%; about 60% of AI users comfortable with AI drafting marketing with limited review; 61% optimistic about their own business, 18% about the US economy (down from 29%); 52% raised prices in the past year — https://www.morningstar.com/news/pr-newswire/20261112sf94521/small-business-owners-are-more-confident-in-their-own-companies-than-in-the-economy-and-two-thirds-use-ai-gusto-survey-finds
- StreamlineCY, "AI News: Claude Fable 5.2 & AI Agent Limits" (November 2026) — Fable 5.2 and Mythos 5.2 released November 9 (Anthropic reports up to 30% faster than 5.1; developer price $8 / $40 per million tokens; Anthropic advises starting most work on Opus 5.5); Co-Design Bench reported by The Decoder on November 8 (273 tasks from more than 26,000 real sessions, 2,331 criteria; best model averaged 69.9% of criteria per task and fully satisfied every criterion in 26.4% of tasks); CNBC on November 8 on multi-agent research (groups more effective but less ethically aligned; teams failing to match their best member) — https://streamlinecy.com/blog/ai-news-claude-fable-5-2-ai-agents-reality-check-november-2026
- Jaroha, "Can AI Do Freelance Work? What the 41.3% Score Means" (November 2026) — Remote Labor Index (Scale Labs and Center for AI Safety): Claude Fable 5.2 at 41.3% (99 of 240 real freelance projects judged at least as good as the human professional's work), up from a best score of 2.5% when the index launched on October 30, 2025; Opus 5.5 at 21.3%; set drawn from 358 verified Upwork professionals across 23 domains, combined value about $144,000; excludes work needing physical labour, long-term evaluation, or direct client interaction; single benchmark, no confidence interval published — https://jaroha.com/articles/main/ai-agent-clears-41-3-of-curated-freelance-projects-with-big-caveats


Comments