There is no shortage of vendors promising you that an AI agent will fill your pipeline. There is a shortage of honest answers about what actually happens when you plug one into a live sales process.

In this session, we’ll cover:

• The sales tasks agents handle well today and the ones they reliably botch

• What a realistic first deployment looks like, and how long before it pays for itself

• How to keep reps in the loop without turning every agent action into a manual approval

• The data and access an agent needs before it can do anything useful, and the risks that come with it

Agents can book meetings, write follow-ups, update your CRM, and score leads, on paper. This 60-minute session covers where they genuinely help, where they create more problems than they solve, and how to roll them out without wrecking the process your team already runs.

Thursday, September 10, 2026 · 12:00 PM ET / 11:00 AM CT / 9:00 AM PT · Free live webinar with Josh Sullivan, COO at Kiingo AI.
Register for the September 10 webinar →

• Nvidia made two moves into the open-model layer in one week. It is paying Poolside $6 billion to license the software system Poolside uses to build AI models, with job offers to 109 of its engineers and a further $1 billion invested at a $12 billion pre-money valuation. Days later it agreed to buy Hugging Face, the repository where a large share of the world’s open AI models and datasets live, for $12.9 billion. Hugging Face turned down a $500 million investment from Nvidia late last year at a $7 billion valuation, saying it did not want a single dominant investor swaying its decisions. Neither deal is confirmed by the companies: the Poolside terms come from an investor letter first reported by Newcomer, and reports say there is no signed contract on Hugging Face yet. (WSJ, Forbes, The Information, CNBC)

• A federal judge blocked the Pentagon’s blacklisting of Anthropic and ordered the government to rescind every directive against the company. Judge Rita Lin of the Northern District of California found officials had retaliated against Anthropic in violation of the First Amendment and stripped it of liberty interests without adequate notice or a meaningful chance to respond. The Defense Department designated Anthropic a supply chain risk in March, cutting it out of military contracts, after the company refused to allow its models to be used for surveillance or autonomous weapons. “The empty invocation of national security is not a blank check to punish and retaliate against government critics,” Lin wrote in a 59-page opinion. (CNBC, Axios, NBC News)

• Salesforce and Anthropic announced Claudeforce, which puts Salesforce inside Claude as a plugin carrying 37 prebuilt sales skills. Sellers can run meeting prep, deal health reviews, and pipeline reviews against live CRM data and take approved actions without opening Salesforce at all. It is with pilot customers now, open beta is scheduled for September, and skills for other business functions follow later this year. (Salesforce, VentureBeat, AIwire)

• Anthropic gave Claude Cowork its own built-in browser, so Claude can handle web tasks like gathering research or collecting invoices in a sandboxed browser inside the desktop app while you keep working in yours. It never sees your personal tabs, bookmarks, or passwords; logins get imported site by site, with banking, email, and single sign-on excluded unless you add them. It is on by default and rolling out this week for Pro, Max, Team, and Enterprise plans on macOS, Windows, and Linux. (Anthropic, The New Stack)

• ChatGPT’s scheduled tasks can now be triggered by events instead of only by the clock, watching for a new Gmail message from a particular sender, a message in a chosen Slack channel, or activity on a GitHub pull request. Event triggers went live August 25 for Plus, Pro, Business, and Enterprise accounts, and scheduled tasks themselves opened to free accounts the same day, without the triggers. Task setups can now be shared with other people, who get their own copy to run. (OpenAI, Engadget, Dataconomy)

• OpenAI published the first test results for Jalapeño, the inference chip it designed itself, claiming 1.5 to 1.9 times more compute per watt than Nvidia’s Blackwell systems along with lower latency. The tests ran on SemiAnalysis’s public InferenceX benchmark, and against a GB200 system the peak figure was 85,448 versus 44,960 mixed tokens per second per kilowatt. Analysts called the comparison incomplete, since Jalapeño uses newer HBM4 memory and Nvidia’s Rubin platform would be the like-for-like match. OpenAI starts deploying the chip in its own data centers by the end of the year and says it will keep buying third-party accelerators. (CNBC, Seeking Alpha)

• Google replaced the Gemini side panel in Google Chat with Ask Gemini starting August 26, a single command line inside Chat that searches across Gmail, Drive, and Calendar, summarizes conversations, pulls out action items, and drafts updates without leaving the thread. Usage limits are raised through October 1 before standard limits apply. History from the old side panel does not carry over, and accounts not set to English keep the old panel for now. (Google Workspace, Neowin)

• Alibaba released Wan3.0, which generates 30 seconds of video from documents, spreadsheets, and slide decks as well as from text and images. It takes PDFs, Word files, Excel sheets, slide decks, and web links up to 100 MB. API pricing runs $0.05 per second at 480p, $0.10 at 720p, and $0.20 at 1080p, putting a 30-second 1080p clip at about $6. It arrived the day after Alibaba priced a HK$80 billion share placement, roughly $10.2 billion, with all proceeds earmarked for AI. (TechNode, WinBuzzer)

• Taiwan indicted nine people, including an employee of Nvidia’s Taiwan unit and two from Super Micro’s Taiwan operation, over 130 AI servers allegedly diverted to customers in China. Customs intercepted 56 of them and 74 reached China, 50 of those routed through Indonesia. Both companies said they are cooperating with Taiwanese authorities. No one has been convicted. (PBS, Al Jazeera, Taipei Times)

• XPENG’s robotics unit raised more than $900 million at a $6.3 billion post-money valuation, the largest single private round on record in China’s embodied AI sector, the business of robots that learn physical tasks. IDG Capital led, with Tencent and Alibaba as strategic investors. The money goes toward its IRON humanoid, which XPENG says enters mass production by the end of 2026 with deliveries starting in 2027. (XPENG, Electrek)

• Infosys surveyed more than 2,600 white-collar workers and found enterprise users save about four hours a week with AI, with most turning that time into higher output, either completing new tasks or moving to higher-value work. More than two-thirds now use AI at least once a day. (Infosys Knowledge Institute, July 2026)

• Thomson Reuters surveyed 1,816 professionals across 62 countries and found 66% say AI is meeting or exceeding expectations where their firm has a named AI strategy. Where no active strategy exists, that figure falls to 22%. (Thomson Reuters Institute, Future of Professionals 2026)

Every few weeks something tells you the model got better. A release note, a vendor demo, a headline, someone on your team who is very excited. You have no way to check whether any of it is true for the work you actually do. Public benchmarks measure coding puzzles and exam questions. Five tasks from your own week will tell you more than any leaderboard.

1. Pick five jobs you already know the answer to. Real ones from the last month, where you know what good looked like because you produced it. The client email you were happy with, the summary you would stand behind, the analysis you checked by hand, the memo that survived the meeting.

2. Put the inputs and your answers in one document. Call it House Test Set. The raw input on top, your version underneath. Twenty minutes, once.

3. Run the five whenever something changes. A model upgrade, a new tool, a vendor pitching you their agent. Same inputs every time, pasted cold, with no coaching and no follow-up questions.

4. Read the results against your own answers. Where it landed on target, where it missed, and which of the five moved since the last run. An upgrade that improves three tasks and quietly degrades two is a real finding, and it is invisible from the outside.

5. Keep the file. Add a task when the work changes. Retire one when it stops being representative.

The next time somebody tells you a new model is better, you can find out in ten minutes, on your work, in your words.

We’re in the middle of moving Kiingo onto something we’ve been calling a company brain. One dashboard that securely holds all of our data, with everything else built on top of it: connections between systems, automations, custom views for whoever needs them, and a much easier time talking to each other across departments.

What’s surprised me is how much falls away once the raw data lives in one secure place. Fewer paid apps. A lot fewer tabs. Fewer messages that just say “can you send me that file.” Everyone here is figuratively on the same page because we are all literally on the same page.

I think this is where AI-guided business is headed, and the order matters. Consolidate the raw data first, securely, in one spot. The automations and the AI work sitting on top of it get much simpler once there is one place to point them at.

The dashboard is the visible part. The real asset is the consolidated data sitting underneath it, which is what every AI tool you buy from here on gets pointed at.

Something new claims to be better almost every week now. The companies that get value from any of it know what they are trying to improve before they go looking.

Kiingo makes your company AI native: your data consolidated in one secure place, the automations and internal tools built on top of it, training for the people who will use them, and a way to tell whether it is working.

Let’s talk it through →