This website uses cookies

Read our Privacy policy and Terms of use for more information.

THE AI PROFIT WIRE

Issue #27 | November 17, 2026 | Weekly Intelligence Briefing

Facts checked as of November 12, 2026.

Notion confirmed that attackers reached Notion AI chat transcripts used for its Insights feature. The transcripts sat at Fireworks AI, an inference vendor Notion used, and fewer than 1% of Notion workspaces were affected.

Fireworks' own bulletin traces the way in. A third-party cloud provider hosting Fireworks R&D machines had a security incident on October 23, and the attackers used it the next day.

That's the vendor behind the vendor behind the app. Neither company appears on your Notion invoice.

The same week, agents took on more of the business. Natera's lab agent gives 96.2% of support tickets a first answer with no engineer starting it.

AWS and Google both shipped skills: your team's rules, written once for the AI to reuse.

United Airlines went the other way. It's restricting unauthorized third-party AI agents and names shared account credentials as the risk.

The question for your stack this week: once the agent does the work, who holds the data, the logins and the bill?

Below: 5 signals, the Hype Check Spotlight, and the 1 tool worth your Monday.

What happened:

Natera, a cell-free DNA testing company, published how it built SWARM, an agentic support system for its lab instruments, on Amazon Bedrock AgentCore. The post went up November 12, co-written with AWS.

The trigger was a departure. A senior technician who spent years diagnosing a particular instrument family left, and the knowledge left with them.

What the data says:

Average first response fell from 17.4 hours to about 8 minutes, about 130x faster, over a 90-day production evaluation. The team handles more than 65 tickets a day across 3 Jira queues.

Coverage reached 96.2%: that share of tickets gets an automated first response before an engineer has to start the diagnosis.

The accuracy numbers are the honest part.

High-confidence answers were right 74% of the time across 154 cases. Moderate and Low both landed at 65%, and the team calls the 37-case Low sample too small to conclude from.

Adding missing vendor manuals for the weakest instrument family lifted accuracy from 71% to 75% with no model change. "Diagnostic accuracy is a data quality problem first," says Ariel Jirau, who leads Natera's lab engineering team.

Natera hasn't published what SWARM costs to run.

About 8 minutes to a first answer, and right 74% of the time on its surest calls. The accuracy gain came from adding manuals, with the model unchanged.

Business impact:

→ List the 3 ticket types your team answers from 1 person's memory, and write those answers down before that person takes a week off.

→ Feed an agent your manuals and past tickets before you pay for a bigger model. Natera's 4-point gain came from documents.

→ Log a confidence label on each AI answer and check it against the real fix for a month, the way Natera did for 90 days.

Read the full signal.

What happened:

AWS published how agents and skills work in AWS Quick on November 9. A skill is a reusable set of instructions that teaches an agent a task, like your weekly status update built from your template.

Google started rolling out skills in Google Workspace the same week. Gemini applies them in Gmail, Docs, Slides, Drive and Chat when it spots a match, or when you @mention one.

What the data says:

AWS opens its walkthrough on the problem: "If you use AI agents, you've probably explained the same task more than once."

Quick's Free plan is $0 per user a month. Plus is $20 per user a month on yearly billing, and the AWS-account plans start at $3 per user plus a $250 monthly infrastructure fee per account.

Google lists no separate price for skills. They come with Business Starter, Standard and Plus and the Enterprise editions.

Rapid Release domains started November 9, and Scheduled Release domains start November 30, each with up to 15 days for the feature to appear.

The limits: Workspace skills don't work in Sheets or Meet yet, and they take English prompts and instructions only.

A skill is the training binder you haven't written yet, and both vendors now read it for you. Whoever writes it decides how the AI sounds to your customers.

Business impact:

→ Write 1 skill for the document your team rebuilds weekly, such as the status update or the quote follow-up, and test it on last week's real inputs.

→ On Workspace, check whether your domain is on Rapid or Scheduled Release before you plan around it. Scheduled domains start November 30.

→ Give brand-voice skills 1 owner. A skill shared from Workspace Studio arrives as an independent copy, so later edits don't reach the people who already have it.

Read the full signal.

Source: Google

What happened:

Google opened the Gemini Decisions API on November 11, built on a new model, Gemini Decision 1. It scores up to 1,000 yes/no, choice and rubric questions in a single pass across text, images, audio, video and PDFs.

What the data says:

Gemini Decision 1 costs $0.10 per million input tokens and $0 for output, with a no-cost tier in Google AI Studio for prototyping. The shared input bills once per request, however many questions ride on it.

Google reports 150 to 250 milliseconds end to end for inputs up to 16K tokens, almost flat from 1 to 100 questions. It puts the API almost 5x cheaper than Gemini 3.5 Flash-Lite across its own benchmark suite.

Inputs run up to 256K tokens. Each answer comes back as a probability with a confidence score, which your code can use to send unclear cases to a person.

Google's Gemini Notebook uses it to route queries and reports up to 40% lower latency on everyday queries.

A routing call needs a yes, a no or a score. On this endpoint you stop paying for the words around it.

Business impact:

→ Pick the sorting job a chat model does for you now, like ticket routing or lead tagging, and price it at the $0.10 input rate.

→ Send anything under your confidence threshold to a person, and count how often that happens in week 1.

→ Treat the 5x figure as Google's own benchmark mix. Your ticket mix sets your real saving.

Read the full signal.

What happened:

Issue #25 covered Carbon going to cyber defenders first. Google has now opened it to paying developers and enterprises through the API in Google AI Studio and Gemini for business.

Google AI Ultra subscribers get it in the Gemini app, Google Antigravity and AI Mode in Search.

What the data says:

The introductory price matches Gemini 4 Argon: $2 per million input tokens and $10 per million output. Google's footnote sets it at $4 and $20 after November 30, 2026.

Google says Carbon ranks #1 on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench at more than 2X lower cost per task than the next best external model. Google chose those benchmarks.

Carbon uses more tokens than Argon by design, to reason at length and run extended verification loops.

The per-token price is the same. The per-task bill isn't.

Same price per token as Argon, more tokens per task. Run Carbon where a wrong answer costs more than the extra tokens.

Business impact:

→ Run your hardest multi-step job on Carbon before November 30, and compare its total bill against Argon on the same task.

→ Budget at $4 and $20 from December 1. Issue #23 flagged the same reset for Argon.

→ Keep routine summaries and tagging on Argon. Google itself calls Argon the ideal choice for the majority of software engineering and knowledge work.

Read the full signal.

Source: TechCrunch

What happened:

Nectar, founded by former Meta employees Misbah and Farah Uraizee, launched Creator Agent on November 12. It finds creators, checks their content and past partnerships, drafts and sends outreach, and tracks campaign results.

What the data says:

Nectar says the agent searches more than 300 million creator profiles, including smaller creators gaining traction. It reviews video content in context for brand safety instead of searching for keywords.

"Brands can share voice guidelines in any format, and Nectar picks up the rest from their history," CEO Misbah Uraizee told TechCrunch.

U.S. spending on creator advertising is expected to reach $44 billion in 2026, according to the Interactive Advertising Bureau, as TechCrunch reports.

Aspire and GRIN are building the same kind of agent.

The launch came with no independent performance numbers.

An agent that writes in your brand's voice sends messages with your name on them. Read the first 20 before it sends the next 200.

Business impact:

→ Give the agent your written voice rules, then review the first 20 drafts by hand before you let it send.

→ Ask the vendor where creator conversations are stored and which model providers process them. Notion's week is the reason to ask.

→ If you work with fewer than 20 creators, a spreadsheet still does the job.

Read the full signal.

Source: AutoGPT Blog

An AutoGPT Blog roundup updated November 9 covers 9 demo platforms with free plans. The newest layer is an AI agent that talks to your buyers while your team is offline.

Community adoption. Supademo's pricing page claims more than 3,000 paying organizations. That's the vendor's count, and the roundup carries no independent usage figure.

Pricing model. Supademo's free plan gives 1 creator 5 guided interactive demos and 50 video recordings. Navattic's free Starter plan allows 1 seat and 1 demo.

Supademo's AI Demo Agent isn't in the free plan. It runs about $1 per call, and the vendor says a 15-minute session uses about 20 credits.

Benchmark data. The strongest number is Arcade's. Its own report on 14 million demo sessions found AI voiceovers raised completion rates by 14%, which is Arcade measuring Arcade demos.

Expert sentiment. The roundup's author tested an AI-narrated demo and wrote that it "explained each step clearly." That's 1 reviewer's hands-on test.

Release maturity. The click-through tours are a settled format. The buyer-facing agent is new and priced by usage, and Arcade moved to credit-based tiers in late October 2026, per the roundup.

The verdict: the free tiers cover 1 homepage tour. The agent that answers buyers is where the meter starts, at about $1 a call.

Build the free tour this week. Before an agent answers buyers for you, run 10 test calls yourself and read the transcripts.

Read the full signal.

Source: n8n Blog

Nod is a free, open-source personal agent from n8n Labs, released November 10 as a research preview under Apache 2.0. It runs on your Mac, Windows PC or server, in its own local sandbox and browser, and asks before it acts.

The jobs fit a small business desk: sign in to supplier portals, download the invoices and file them for your accountant. n8n reports 93.2% on the LongMemEval memory benchmark, its own published score.

The software costs $0.

You pay your model provider, and Nod checks in once a task passes $5. n8n's own post says some early testers found tasks cost more than they expected.

The weak spots: it stops when your laptop sleeps, each install has 1 owner, and it can't make phone calls. Its permission check is a model too, and n8n says it can make mistakes.

Install it on a machine that stays on, give it 1 portal chore with the $5 check-in, and read the receipt before you hand it a second.

Read the full signal.

The Wire: What Else Made the Cut

Who holds the data, and who holds the login.

Notion AI chats sat at Fireworks AI, and Fireworks was reached through its own cloud host.
Notion says fewer than 1% of workspaces were affected and has disabled Insights. Fireworks lists billing records, plus datasets and incomplete model weights for certain customers. Payment card numbers weren't stored there.
Details from Notion and Fireworks AI.

United Airlines is restricting unauthorized third-party AI agents on its booking channels.
United names shared account credentials as the risk, along with unintended purchases, account changes and MileagePlus spending. It says banking and retail leaders have taken similar steps.
Announcement from United.

Enterprise AI spending reached $177 billion, and agents now generate 80% of OpenRouter tokens.
Menlo Ventures counts $33 billion a year earlier, with growth of 5.4x. On OpenRouter, agent token use now runs about 7x human use.
Reported by Menlo Ventures.

Otter added 27 languages, and a mid-meeting language switch is on the way.
Of the new languages, 13 are live now, for 31 in total once the rollout ends. They run on Otter's own speech models, checked by native-speaking volunteers.
Reported by TechCrunch.

The full week's signals, detailed breakdowns, and action items are on the site. If this issue earned its place in your inbox, forward it to whoever signs off on your AI budget.

This issue goes to subscribers before it hits the web. If you want next week's first, subscribe at metadatamarketer.com/subscribe

Test. Cut. Share.

Moe Sbaiti, The AI Profit Wire https://metadatamarketer.com

Disclosure: some tools referenced in this newsletter are affiliate partners. Full disclosure and analysis at metadatamarketer.com.