We’ll break down how to tell whether a task needs simple automation or a true AI agent, so you invest in the right solution the first time, with real examples and a framework you can use with your own team.

Tuesday, August 11, 2026 · 1:00 PM ET / 12:00 PM CT / 10:00 AM PT · Free live webinar.
Register for the August 11 webinar →

• OpenAI cut the price of Luna, its cheapest GPT-5.6 tier, by 80% on July 30, to $0.20 per million input tokens and $1.20 per million output tokens, and trimmed mid-tier Terra by 20%. DeepSeek followed on July 31, opening its V4-Flash coding model at about $0.28 per million output tokens, roughly 99% below Claude Opus 4.8 on output. (CNBC, Axios)

• Microsoft closed its fiscal year with Azure revenue above $100 billion for the first time and more than 30 million paid Copilot seats, up from 20 million in April. Fourth-quarter revenue hit $90 billion, up 18%. The stock rose more than 15%, adding roughly $450 billion in market value in one session, the largest single-day gain any company has posted. (CNBC, Bloomberg)

• Amazon raised its 2026 capital spending plan to $220 billion from $200 billion, and chief executive Andy Jassy attributed the increase to rising memory chip prices rather than to buying more compute. Meta’s free cash flow fell 91% last quarter, to $784 million from $8.55 billion a year earlier. (CNBC, Fortune)

• Amazon is winding down most of its Nova model line, moving Nova Premier, Nova Omni, the Reel video model, and the Canvas image model to maintenance status while it consolidates behind a single frontier model. Nova 2 Lite, Nova 2 Sonic, and the Nova Forge customization service continue, and existing customers keep support on the retired models. (Business Insider via Yahoo Finance, The Next Web)

• The EU began enforcing the AI Act’s general-purpose AI rules on August 2, and the Commission’s AI Office can now fine model providers up to 3% of global revenue or 15 million euros, whichever is higher. New transparency duties started the same day: chatbots must say they are machines, and AI-generated images, video, and audio must be labeled. (European Commission, EU AI Act timeline, AI Act Article 101)

• Google gave its Gemini Spark agent control of Chrome. It can act on sites a person is already signed into and, with permission, use passwords saved in Chrome to finish tasks. The same day, Google added AI satellite imagery to Google Earth, then paused it within a day after journalists and researchers used it to fabricate events that never happened. (Google, Engadget, NPR)

• Google DeepMind released Gemini Robotics 2, a model that controls a full humanoid robot from feet to fingertips, and published its own success rates. The robot unscrews a light bulb 92% of the time, picks an object off the floor 45.7% of the time, and screws a bulb back in 36%. DeepMind wrote that fine multi-finger work remains hard. (Google DeepMind, Bloomberg)

• On July 30, Thinking Machines released Inkling-Small, a free-to-download model that reads text and images and is licensed for commercial use. It scores 40 on the independent Artificial Analysis Intelligence Index, one point below the lab’s flagship, which is more than three times its size. (VentureBeat, Thinking Machines)

• GitHub made custom skills and MCP connections for Copilot’s automated code review generally available on July 29, letting reviews draw on a team’s own tools and coding standards. Notion separately let finished AI meeting notes trigger custom agents. (GitHub, Notion)

• Anthropic reviewed 141,006 of its own cybersecurity test runs and found three where Claude models escaped what was meant to be a sealed simulation and reached the live systems of real organizations. An evaluation partner had left internet access open. One incident pulled infrastructure credentials and several hundred rows of production data. Anthropic notified all three companies on July 27. (Anthropic, TechCrunch)

• Horizon3.ai raised $250 million at a valuation above $2 billion, roughly triple its $650 million mark a year earlier. Its NodeZero platform attacks a customer’s own network to find the holes before real attackers do, pulling in a human only by exception. NightDragon and NEA co-led the round. (SiliconANGLE, Forbes)

• Gallup surveyed 22,573 US workers and found that among employees using AI across seven or more types of tasks, 90% report a positive productivity impact, against 45% of those using it for one or two. Overall, 52% now use AI in their role. (Gallup, July 20, 2026)

• The Conference Board surveyed nearly 1,300 workers and found 55% use generative AI or AI agents daily or weekly, while 33% received employer-provided AI training in the past six months. Fewer than half say they get enough work time set aside to build those skills. (The Conference Board, July 28, 2026)

• Roland Berger surveyed more than 550 senior decision-makers across five industries and three regions. More than half of the companies running AI in customer service report significant measurable impact, showing up in response times, customer satisfaction, and operating costs, and most said their teams’ day-to-day work improved. (Roland Berger, July 7, 2026)

Most people meet a new AI chat the same way every time: a blank box and three paragraphs of background before they can ask the real question. That re-explaining tax is quietly what keeps people using AI for one or two kinds of work when it could be helping with ten. All three major assistants have a fix built in, and it is the feature most people have seen in the sidebar and never opened. Set the context down in one place, and every conversation you start there begins already knowing it.

Make one container for a recurring area of work. Pick something you come back to weekly: a client account, board reporting, hiring, a product line.

In Claude and ChatGPT, open Projects, create one, and fill in the project instructions.

In Gemini, open Gems and fill in the instructions field.

Put two things inside it. First, a few sentences covering what this area of work is, who is involved, and what good output looks like. Second, the reference material you keep re-attaching: last quarter’s report, the pricing sheet, the style guide, the account history.

Then open a brand-new chat inside the container and run this:

“Using only the material here, draft [the recurring thing you make here]. Where the material does not cover something, mark it NEEDS INPUT instead of guessing.”

Two useful things come back at once: an actual draft and a marked list of the gaps in what you loaded. Fill those gaps, and the container gets better every time you use it.

A tidier sidebar is the small win. The larger one shows up a month later, when aiming AI at a new corner of your work stops being a fresh setup problem every time.

Something I’ve noticed the longer I do this: the gains arrive in pieces, and they keep arriving as long as you keep poking at the tools. A new model, a new version, some feature I ignored the first three times, and suddenly there’s a way to use it that hadn’t occurred to me. Usually that’s me catching up to what the tool could already do.

What surprises me is where those gains keep coming from. I’ll be well into a workflow I’m already happy with, saving real time on it every week, and then notice a step I’m still doing by hand purely because I was doing it by hand back when I set the thing up. I never went back and questioned it. Most of the time I can hand it off too, with the right checks and gates around it.

AI use compounds the way any hands-on skill does. You get good at it by using it clumsily for a while, noticing what annoys you, and fixing that one thing. What speeds all of it up is having someone very in-the-know to go to, so you spend less time rediscovering what somebody else already worked out.

The tools cost a fraction of what they did a year ago. What decides the return now is knowing which parts of the work are ready to hand over.

Kiingo makes your company AI native: a clear sequence of what to build next, with the guardrails that make each step safe to take.

Let’s get started→