What Every Leader Needs to Know About AI Right Now (September 2026)

They're saving time, yet almost none of it is going back to them. That’s the finding in Deloitte's inaugural GenAI Workforce Survey, a study of 25,000 UK workers published this month.

• Workers report saving around 70 minutes a week using GenAI, but they reinvest that time into more work for the same employer.
• Usage is still shallow. The top reported tasks are searching for information (43%), drafting emails (43%), and creating summaries (31%). About half of employees using GenAI say they've received no formal training on how to do it safely or well.

What should worry every leader reading this is the trust gap.

Nearly a quarter of workers think there's stigma attached to using GenAI at work, and 64% of weekly users worry their manager thinks AI can now do their job.

This is why we push people to set a North Star for their saved time, and why we measure Impact per Hour in place of hours saved.

According to a new McKinsey piece, Cutting the "coordination tax": How agentic AI can reshape workflows, the value lives between the steps.

McKinsey calls it the "coordination interface tax": the aligning, checking, reconciling, approving, and waiting that happens as work moves between people, teams, and systems.

Across knowledge-intensive organizations, this coordination consumes 35% to 60% of total work time.

Look at the table above from a Fortune 500 industrial manufacturer. Nine to eighteen days of total latency in a single demand-to-production workflow. This time isn't for the work itself, but for waiting between steps.

Agentic AI can act at those handoffs, but only under human supervision. Three jobs agents can help with include: verifying whether one step's output is ready for the next, reconciling conflicting constraints between teams, and routing work to the next step or flagging the right employees for intervention when human judgment is needed.

This is the gap our OVER framework exists to close: you have to actually know your own workflow well enough to see where the handoffs are- or Workflow Literacy.

Most companies still apply AI inside individual steps rather than redesigning the workflow around it.

That is why AI makes individuals faster without adding value to the entire workflow.

Why Don't People Trust AI Agents Yet?

Mostly because they don't understand what the agent is actually doing, or what it might do without asking first. That's from a study published in the Harvard Business Review of more than 3,000 people evaluating an AI financial adviser.

• Being explicit about an AI's limitations increases trust. Across two experiments, clearly telling users where the AI was likely to fail made it seem 14.6% more transparent and improved how effectively people worked with it by up to 7.2%.
• Agents are also trusted more when they explain their reasoning rather than just handing over a conclusion. People were 54% more likely to accept an AI recommendation when the system showed it understood their broader goal and context.

The headline finding: the sweet spot is bounded autonomy, not full autonomy.

Let agents handle the routine work, but always wait for a human’s confirmation for anything consequential.

In practice, that means controls like "write this email and share a draft, but don't send it," or "recommend the inventory move, but don't execute until I approve this."

Adoption concerns in the study stemmed from privacy, fear of unwanted actions, and not understanding what the AI was doing. Bounded autonomy, where you can be upfront about limits imposed to AI, directly addresses this.

Every platform had major announcements this month. Here's what you should pay attention to:

ChatGPT

OpenAI launched GPT-6 Astra, its new flagship built to run software on your behalf: updating CRM records, producing documents and slides in your own templates, and handling multistep work end-to-end.

It asks for your sign-off on consequential decisions and asks questions when clarifications are crucial to an accurate output.

It's rolling out to Plus, Pro, Business, and Enterprise plans (Enterprise admins have to switch it on).

ChatGPT now also lives directly inside Microsoft Word, joining Excel and PowerPoint through the same add-in, on every ChatGPT plan, including Free. You draft from notes, revise selected text, and adjust formatting without leaving the document.

Claude

It was a busy month on Claude’s side too.

Anthropic released Claude Opus 5.5, its new leading model for complex knowledge work and long-running tasks.

It performs close to Claude Fable 5.1's level on most work while costing 40% less to run than Opus 5.

Anthropic has merged Claude Cowork and normal chat into one interface, starting with Pro and Max plans. You no longer choose between "chat" and "cowork" before starting a task.

This means conversations can move from a quick question into working with files, using connectors, and producing a finished document without you switching modes. Cowork tasks now run in the cloud by default, so they aren't tied to one laptop or one open session.

Gemini

Google launched a dedicated Gemini app for Windows, letting you press Alt+Space from anywhere on the PC to bring Gemini up over whatever you're working in, with a separate workspace for handing longer tasks to Gemini Spark.

Gemini can also now pull context directly from Gmail and Drive. Google added voice-led drafting in Docs and conversational search in Gmail, alongside a more capable Gemini 3.8 Live for real-time, multi-step voice work.

Copilot

Microsoft's Copilot Cowork picked up GPT-6 Astra for its hardest tasks and Claude Fable 5.1 for the most ambitious work.

Copilot Cowork also introduced an App skill that lets anyone describe a lightweight internal app in plain language and have it built and shared. For now, that's restricted to Microsoft's Frontier preview program.

All of this brings us back to the capability overhang: there is far more possible with AI right now than almost anyone at your company is using.

Closing it takes two things: understanding what AI can do now, and understanding your own workflow well enough to see where that capability fits.

As agents become more integrated into the workflow, human judgment becomes more valuable than ever.

Key Takeaways

• Plan for AI safety now. The labs building this technology are calling for a slowdown, but training is still ongoing. Build your own governance and review practices rather than waiting for the industry to settle the question.
• Train people to judge AI output well. AI fluency means evaluating and improving AI output against professional standards.
• Decide how you use the time saved. Deloitte's 70-minutes-a-week finding shows leaders need to actively redirect that time. Don’t let this time silently turn into more meetings and emails.
• Redesigning the workflow is a top priority. Coordination between teams and systems eats 35% to 60% of work time. That is where agentic AI creates value.
• Build trust by naming the limits. Being explicit about where an AI can fail increases both trust and effectiveness. Additionally, use bounded autonomy: let agents handle routine work, and require sign-off on anything consequential.

This article is based on the September 2026 Lead with AI PRO Executive Briefing with Daan van Rossum and Phil Kirschner.

PRO members receive monthly executive briefings, masterclasses, and a peer community of leaders applying AI across industries. Learn more about Lead with AI PRO, start your AI fluency journey with AI Leader Advanced.