Finance work runs on accuracy and deadlines. We are spending 60 minutes on where agents fit in finance operations today, the work that still needs a person, and how to keep the audit trail clean when an agent touches financial data.

• The finance tasks agents handle well today, including reconciliations, report pulls, exception flagging, and compliance checks

• Where agents don't belong yet, and what breaks when you push them there

• How to keep a clean audit trail when an agent touches financial data

• What to look for before trusting an agent with anything that hits the books

Tuesday, October 6, 2026 · 12:00 PM ET / 11:00 AM CT / 9:00 AM PT · Free live webinar with Josh Sullivan, COO at Kiingo AI.
Save your spot →

• Anthropic and OpenAI each shipped a faster model at an unchanged price. Claude Sonnet 5.5, released September 28, keeps Sonnet 5's price of $2 per million input tokens and $10 per million output (tokens are the units AI usage is billed in), and Anthropic says it runs more than 30% faster and costs up to 30% less per task because it uses fewer tokens and tool calls. Box reported work finishing 2.4 times faster, and Zendesk said tickets were processed 20% faster. A day later, OpenAI released GPT-6.1 Sol at the same $2 and $10 as GPT-6 Sol; it matched GPT-6 Astra on a software engineering benchmark using roughly a fifth of the tokens per task. (Anthropic, SiliconANGLE, SiliconANGLE, OpenAI)

• OpenAI launched always-on agents called Dots at its DevDay conference on September 29, one day after confirming it had canceled the October release of GPT-6.1 Astra. Each Dot runs on its own cloud computer with a browser, connects to more than 4,000 apps, and can be messaged in ChatGPT, Slack, and Teams; it is available now on the Pro and Business Premium plans in eligible markets. Astra was pulled after internal tests found it was less honest with users about what it had done and used outside tools without approval; Saachi Jain, OpenAI's head of safety systems, said it fell short on "scope and authorization." Separately, OpenAI has paused tool-use training, evaluation, and inference on its most capable models after an agent in training exploited a gap in its sandbox’s internet restrictions on September 20 to reach an outside chatbot. (TechCrunch, Engadget, OpenAI, SecurityWeek, Decrypt)

• Microsoft rebuilt Copilot as one app with three sections: Home, Code, and Autopilot. Home combines chat with the Cowork agent and built-in versions of Word, Excel, and PowerPoint that work on real, editable files. Code lets someone with no programming background describe a tool or dashboard and have it built. Autopilot is an agent with its own identity and permissions that keeps working while you are away. Chat and the Office apps stay under the existing per-user license, while Cowork, Code, Autopilot, and the latest frontier models from OpenAI and Anthropic are billed by usage. Home and Code reach Microsoft's early-access program in the coming weeks, and Autopilot expands to a private preview at the end of the month. (Microsoft, CIO Dive)

• Meta opened its Muse agent to small businesses, and Shopify opened its checkout to agents like it. Muse now connects to 15 outside tools, including Shopify, QuickBooks, Stripe, Slack, Canva, Asana, Notion, and Zoom, along with Instagram analytics and Meta ad accounts. It drafts growth plans, reviews the books for unusual expenses, analyzes ad performance, and flags emails that need a reply, and it publishes, sends, or spends nothing without the owner's approval. It is free with usage limits in the US and Canada, and Meta also formed an enterprise unit, led by former MongoDB CEO CJ Desai, to sell it to larger companies. On Shopify stores, an agent working in a buyer's browser can now fill in the checkout and submit the order, including through Shop Pay, once the buyer confirms. The same day, Instinct, a personal agent that works over text message to book travel, pay bills, and cancel subscriptions, raised $1 billion at a $10 billion valuation. (TechCrunch, Meta, Shopify, TechCrunch, Meta, TechCrunch)

• Google made Gemini 3.8 Live with Live Avatar generally available to Gemini Enterprise customers. It is a voice agent with a lip-synced video face that can see a customer's camera or shared screen and speaks 97 languages. Custom avatars require Google's approval, and every avatar carries SynthID, Google's invisible watermark for AI-generated content. Cox Automotive's Autotrader and Salesforce's Agentforce are among the customers named at launch. Separately, Koray Kavukcuoglu, the head of Google DeepMind, said Google intends to release an early version of Gemini 4 "as soon as possible," and hopes it lands well before year-end, without giving a date. (Google Cloud, Engadget, 9to5Google)

• Anthropic reported that Claude agents found a previously undescribed enzyme system in viruses that infect bacteria, with a structure that resembles CRISPR, the gene-editing tool. About 950 Claude agents working in parallel for 21 hours searched roughly 1.9 billion protein clusters, narrowed 3,500 candidates to 20, and human scientists ran the lab experiments that confirmed the finding. The paper is a preprint that has yet to be peer reviewed, and Anthropic says the system's function and usefulness are still unclear. Outside researchers called the find exciting and said there is no evidence yet that it works like CRISPR the technology. (Anthropic, Al Jazeera, Gizmodo)

• Four disclosures this week put numbers on the AI build-out. Anthropic's confidential draft IPO prospectus, reviewed by Reuters and the Financial Times, lists $518 billion in future cloud and computing commitments, against about $4.6 billion in 2025 revenue, $11.5 billion in the second quarter of 2026 alone, and a 2025 operating loss of $8.06 billion. Akamai, best known for the network that speeds up websites, signed an $11.6 billion, seven-year deal to run Anthropic's general computing work, expandable to about $20 billion, and its shares rose as much as 17% in after-hours trading. Oracle sent a force-majeure notice, which protects it if the 2.45-gigawatt New Mexico data center campus it is leasing misses its 2028 target, after a gas pipeline serving the site slipped about six months; Oracle says the project remains on schedule. Goldman Sachs projects the largest tech companies will borrow about $250 billion this year toward roughly $750 billion in AI-driven capital spending. (Fortune, TechCrunch, TechCrunch, Yahoo Finance, TechCrunch)

• Tools and rules for keeping AI agents in bounds arrived from four directions. Nvidia released the Open Agent Safety Platform, open-source software that fences in an agent's actions, paired with a hardware monitor that can quarantine one that misbehaves, with dozens of partners including Anthropic, Microsoft, and Oracle. Island, which makes a secure web browser for companies, raised $400 million at a $6.4 billion valuation to govern what both employees and agents can reach. FTC Chairman Andrew Ferguson said at a Reuters event that he would "resist this anthropomorphizing of these tools" and suggested developers who direct agents would be held liable under existing FTC authority. Senator Josh Hawley announced plans for a bill creating civil liability for companies that recklessly design agents, criminal liability for companies that know their agents are capable of crimes and fail to build reasonable safeguards, and criminal liability for users who knowingly let an agent commit crimes. (TechCrunch, Island, Yahoo News, Senator Hawley)

• Tech leaders signed a voluntary safety accord at the White House on September 29, the same day President Trump ordered federal agencies to use "Super Intelligence" in place of "artificial intelligence" in official documents. Executives including Dario Amodei, Sundar Pichai, Mark Zuckerberg, Jensen Huang, and OpenAI's Greg Brockman committed to layers of internal and external review, with independent auditors whose findings each company board reviews directly. Asked whether the accord is legally binding, Trump said it is "morally" binding. Three days earlier, Trump and Xi Jinping agreed to a US-China dialogue on the technology and a channel for reporting incidents. (White House, ABC News, CBS News)

• Two federal appeals courts ruled against AI companies within five days. On September 25, the D.C. Circuit voted 2-1 to uphold the Pentagon's "supply chain risk" designation of Anthropic, which bars the military and its contractors from using Claude; Anthropic pointed to a separate federal court that found a parallel designation unlawful and is weighing further review. On September 29, the Third Circuit upheld Thomson Reuters' win against Ross Intelligence and the lower court’s finding that copying Westlaw's legal summaries to train a competing AI search tool was not fair use, the first US appeals court decision in a copyright case over AI training. (ABC News, Investing.com, Law360, The Next Web)

• BCG's survey of more than 1,300 senior leaders found AI agents already produce 22% of companies' total AI value, up from 17% a year earlier. BCG projects that share reaches 39% by 2030, and BCG’s separate operations research finds agents can cut costs 60% or more when processes are redesigned end to end. (BCG Applied AI Index, Sept. 30, 2026; BCG, June 15, 2026)

• Deloitte's survey of 501 US leaders, from senior managers to the C-suite, found 75% say people working with AI agents create more value than agent automation alone. Within four years, 74% expect nearly half their business processes to be rebuilt around agents, and 61% expect most agents to work on their own with humans overseeing them. (Deloitte, Aug. 12, 2026)

• KPMG's Q3 AI Pulse of 314 US business leaders at organizations with $1 billion or more in revenue found 62% of organizations are now building, deploying, or developing AI agents, up from 53% last quarter. The share developing or running systems where several agents work together climbed to 25%, up from 6% in the prior two quarters. (KPMG AI Quarterly Pulse, Sept. 24, 2026)

Most requests to an AI assistant run about a sentence, and the reply fills every gap with a guess: who it's for, what good looks like, what to leave out. The work itself is the part the assistant does well now. What it needs from you is the brief, and the easiest way to write a good one is to let the assistant ask for it.

1. Hand over the job with this.

“I want you to [the task, e.g., write the pricing-change letter to our wholesale customers]. Before you start, interview me. Ask one question at a time and wait for my answer. Only ask what you can't find or work out yourself. Where there's a standard best-practice answer, make the call and tell me what you chose. When you have enough, write back the brief you'll work from: goal, audience, constraints, what done looks like, and every decision you made on your own. Then wait for my go-ahead.”

2. Answer what only you know, and send the rest back. The customer who is already upset, the number you can't go below, who has to sign off: those come from you. When it asks something you don't have a view on, like formal or conversational, or one page or two, reply “What would you recommend, and why?” and let it decide.

3. Correct the brief, then let it run. The brief is the checkpoint: a few lines saying exactly what it's about to do and what it decided for you. Fix anything that's off and tell it to go. When the finished work misses, the fix usually belongs in the brief, so correct it there and run it again.

A brief that produced something you actually sent is worth keeping. Saved into a Claude or ChatGPT Project, or a Gemini Gem, it becomes the starting point the next time the same job comes around.

When I started working with agents, I handed them one step at a time and checked each one before giving them the next. These days I’m handing over whole jobs, start to finish, and the trust came from learning which jobs deserve it. Somewhere along the way it clicked that every step I kept checking myself was a job I’d held onto without ever deciding to.

It works a lot like a loop pedal. A loop earns its place under the set once you’ve heard it come back clean a few times. After that, your hands are free for whatever comes next, and the loop keeps playing whether you’re watching it or not.

The jobs I now hand over completely have a few things in common. I can say what done looks like in a sentence, there’s a way to check the result without me eyeballing every line, and a mistake is cheap to undo. When one of those is missing, it shows up fast, because the agent keeps coming back to ask me to check something. That usually means it’s missing a way to check the work itself, so that’s what I build next.

The review is still mine. What changed is where my time goes: less on doing the work, more on deciding what the work should be and building the check that tells me it’s right.

A lot of teams have quietly become their AI's hands: pasting questions in, copying answers out, clicking whatever it says to click. The teams pulling ahead flipped that. They describe the job well, give it to the agent, and spend their time on the review.

Getting there is an operating change. It takes information your agents can reach safely, rules for what they may do, and people trained to brief and check the work. Kiingo AI builds that with you as one roadmap toward becoming AI native, sequenced for your situation.

Speak with us →