We are spending 60 minutes on the operational work agents handle reliably today, including scheduling, data entry, status updates, and exception handling, plus the work that still needs a person.

• Which back-office tasks agents handle reliably and which still require human involvement

• How agents handle exceptions, including what gets escalated and where the process can break

• Where agents reduce operational work and where they quietly create extra work

• What your data and processes need to look like before an agent can be useful

Thursday, September 17, 2026 · 11:30 AM ET / 10:30 AM CT / 8:30 AM PT · Free live webinar with Josh Sullivan, COO at Kiingo AI.
Save your spot →

• Salesforce and Nvidia launched Koa, a CRM reasoning model built on top of Nvidia's Nemotron 3 Super and trained further on synthetic data modeled on nearly three decades of CRM deployments. Salesforce says it matches or exceeds leading general-purpose models on CRM tasks with three times fewer errors, measured on Salesforce's own benchmark. Pilot customers have it now, general availability is winter 2026. (Salesforce, SiliconANGLE)

• OpenAI launched ChatGPT for Financial Services, running on its new GPT-6 Astra model with built-in data from Daloopa, LSEG, PitchBook, S&P Capital IQ, FactSet, and about a dozen other providers for research, financial models, and customized client materials. Morgan Stanley and Evercore were design partners, giving the product a distinctly Wall Street-shaped starting point. (OpenAI, VentureBeat)

• Anthropic expanded Claude for Small Business to 43 workflows and 27 new integrations, including Shopify, Square, Stripe, Xero, Gusto, and Zapier, and reported more than 900,000 installations since the May launch. The workflows still require a human approval before anything sends, posts, or pays, which is the right default and worth copying wherever you point an agent at live systems. Free workshops run in 10 US cities, with more than 750 more led by approved trainers in their own communities. (Claude, Forbes)

• DaVinci Resolve 21.1 added direct access for Claude, Claude Code, and ChatGPT Codex, letting the assistants analyze projects, organize media, adjust settings, remove unwanted clips, and batch render. OpenAI developer experience lead Brent Schooley used Codex to cut one of the company's own launch videos in Resolve, and kept the editable project at the end instead of a finished file he could no longer change. (VP Land, Brent Schooley)

• Oracle's remaining performance obligations reached $664 billion after it signed more than $30 billion in additional AI cloud contracts during the quarter. Quarterly revenue rose 30% to $19.3 billion, while cloud infrastructure revenue rose 121% to $7.4 billion. Backlog counts contracted future work, so read it as a demand signal more than a revenue number, and as the clearest available measure of how much compute enterprises expect to need. (Oracle, CNBC)

• Temporal raised $550 million at a $12.55 billion valuation to expand the infrastructure it builds for keeping software workflows and AI agents running through failures. The valuation more than doubled in seven months, a useful read on how much investors now value the plumbing beneath agents. (Temporal, GeekWire)

• Cloudflare expanded its AI crawler defaults on September 15, blocking training and agent crawlers on ad-bearing pages for new domains, new customers, and existing free-plan customers who had never set their own preference. Search crawling remains allowed, although multipurpose crawlers such as Googlebot can also be blocked when a site blocks the training category. If your marketing site sits behind Cloudflare, check the setting before you assume AI assistants can still read you. (Cloudflare, Search Engine Journal)

• California created oversight rules for companies that audit AI systems through SB 813 and AB 1405. The laws establish a framework for independent verification organizations and require an AI auditor registry plus independence, transparency, and integrity standards for registered auditors. Expect AI audit language to start showing up in enterprise vendor questionnaires. (Office of Governor Gavin Newsom, Bloomberg Law)

• Microsoft AI published a draft Humanist AI Code of Conduct for the development and release of its in-house MAI models. The 38-page draft commits to human control, transparency, and models that stay interruptible and correctable, and Microsoft opened it to six weeks of public feedback. (Microsoft AI, Fox Business)

• Seven Chinese AI labs ran roughly 190 million Claude exchanges to train competing models, according to Anthropic's September threat intelligence report, led by more than 151 million from Alibaba, 23 million from Moonshot AI, and 12.1 million from DeepSeek. Anthropic says the labs used those outputs for distillation, the cheap way to copy a frontier model's behavior without paying to build one. (Anthropic, CNBC)

• A proposal to slow the pace of frontier AI picked up backing from rival labs and from Brussels in the same week. Anthropic CEO Dario Amodei called for giving independent evaluators employee-level access and slowing the rate of capability gains so alignment work can keep up. Sam Altman said OpenAI would match the evaluator commitment, and Elon Musk posted that Dario is right. Ursula von der Leyen then told the European Parliament that CEOs of the most advanced companies say it is time to slow down on self-recursive models, meaning systems that improve themselves, and pledged to convene the leading labs. No new legislation, since the AI Act already covers it. (Dario Amodei, TechCrunch, Euronews)

• Visa, Mastercard, and Ant International began work on a shared Know-Your-Agent framework for identifying and verifying AI agents across payment networks and wallets. The work runs through the Monetary Authority of Singapore's BuildFin.ai platform and builds on its Safeguards for Agentic Finance at Runtime framework, keeping existing payment risk controls in the loop. While the labs argue about pacing, the payment networks are quietly building the identity layer agents will need before they can buy anything. (Ant International, TNW)

• Deloitte surveyed 1,434 finance leaders in 26 countries and found 43% are prioritizing AI and advanced technology to automate operations. The same study found 95% are comfortable with agentic workflows in finance activities, while 14% support full autonomy for critical decisions. (Deloitte Finance Trends 2027, Sept. 9)

• Gartner surveyed 161 chief audit executives and found 93% report some level of AI use, mainly for engagement preplanning and for drafting audit issues, ratings, and reports. Thirty-eight percent have a formal AI strategy in place and another 39% are developing one. (Gartner, Sept. 10)

• McKinsey surveyed 1,719 respondents across 97 nations and found 80% of AI users report improved individual productivity, and 50% report better decision-making. Thirty-seven percent say AI has contributed positively to their organization's EBIT, roughly level with last year. (McKinsey State of AI, Aug. 25)

An operations board can look busy long after the work has stopped moving. The useful signal is evidence of forward motion: a changed status, a completed dependency, or a dated owner update. A queue-aging sweep turns last-updated dates, blockers, and owners into a clean intervention list.

Paste a current work queue, backlog, or ticket export into your preferred AI assistant, then use this prompt:

"Use the available fields for item, owner, status, due date, last update, dependency, and blocker. Identify items with no recent evidence of movement. Separate them into ACTIVE, WAITING ON EXTERNAL, STALLED, and NEEDS A MANAGEMENT DECISION. Do not invent missing dates or owners; mark them UNKNOWN. For every STALLED item, write the smallest next action and a one-sentence message to the owner. Return as: Item | Evidence of movement | State | Next action | Owner message | Decision needed."

Run it against one live queue, then spot-check three classifications against the source data. The UNKNOWN entries are useful too: they show where the operating record cannot yet support a clean handoff.

What keeps getting easier is seeing the shape of a project before the work starts. Which combination of tools handles which component, and in what order. A coding assistant for one piece, an in-browser tool for another, a chat window for the rest.

Those combinations turn out to be surprisingly repeatable. They behave like recipes: a known set of ingredients in a known order, producing the same result every time you run them. Once you can see one, it's worth writing down as an actual step by step, including how each tool's output becomes the next one's input.

Once a tool combination is written down, it pays off twice. Inside the project, it's a playbook an agent can run end to end, or at minimum it tells each tool what its role is and what the others are handling. Across the week, it means recognizing a combination you've already tested and reaching for it rather than assembling an approach from scratch.

Order is the part worth guarding. The tools are only as good as the handoffs between them, so map the handoffs first, then decide whether to automate the sequence or just run it better by hand.

This week's releases keep moving intelligence closer to the systems where finance, sales, and operations already work. The practical advantage comes from giving that intelligence clean context, clear authority, and an operating record a person can inspect.

An AI-native company keeps its data in one secure company brain, sets governance before tools take action, builds what its workflows actually require, and trains people at the level where they are. Kiingo AI guides that work as one sequenced roadmap, so capability compounds inside the company.

Talk through your next 90 days with us →