AI Works Best When You Understand Your Workflow

AI Breaks When Workflows Break

This study didn’t just look at success rates. It went deeper into how work happens, on screen, in tools, step by step. And that’s where the real gap shows up.

Human workers use interfaces like Figma, Excel, and Google Docs. They click, scroll, drag, and build visually. In contrast, AI agents don’t work visually. They code their way through every task, even visual ones.

Across all jobs, even design, AI agents used programming-based workflows 93.8% of the time. For example, instead of adjusting a PowerPoint slide layout like a person would, the agent wrote a Python script to generate the slide content.

This isn’t a bug. It’s a behavior. But it also creates blind spots. In one example, an agent couldn’t read a company report uploaded by the user. So it searched the web for another report, pulled in unrelated data, and built a presentation based on the wrong file. All without saying anything.

That kind of behavior is hard to catch unless managers know how the AI is completing the task. If you only look at the end result, you might not realize it’s based on false or fabricated data.

This is why more teams are starting to run workflow audits, not just output reviews. It’s not about checking what was done. It’s about understanding how it was done.

Managers don’t need to become prompt engineers. But they do need to become workflow literate, able to track how tasks unfold behind the scenes.

You can’t fix what you can’t follow. If you don’t see the steps, you won’t catch the flaws.

And that’s why workflow literacy is an essential skill. It helps managers ask better questions, set clearer expectations, and catch silent failures before they affect outcomes.

The Secret to Speed Is Smart Delegation

One of the most exciting moments in the study came when the researchers paired human workers and AI agents on the same task.

The humans handled the steps that involved navigation, decision-making, or visual judgment. The agents did the parts that involved cleaning data or writing code.

The result? Tasks were completed 68.7% faster.

This kind of hybrid workflow, where each player does what they do best, is where real ROI lives.

One emerging practice is to map workflows by programmability:

• High: Give to AI (for example, spreadsheet cleanup)
• Medium: AI drafts, human verifies (for example, presentation copy)
• Low: Human-led (for example, interpreting client tone or visual branding)

Delegation is not just about what AI can do. It’s about what humans should keep. Leaders have to decide where judgment, emotion, and context belong.

Co-intelligence isn’t about handing over tasks. It’s about sharing momentum. Humans lead with judgment. Agents bring acceleration.

This approach also mirrors what ​Gartner is now urging executive leaders to consider: leaders must prepare for not one, but four modes of human-AI collaboration.​ These include:

• AI is filling the gaps where humans are still needed after automation
• AI is running autonomously in parts of the business with little to no human input
• Humans are using AI to get more done in workflows that stay mostly the same
• Humans and AI are transforming work together, unlocking new frontiers of creativity and knowledge

The solution is to break work into steps, figure out which ones are clearly programmable, and delegate accordingly.

Not every task should be automated. But almost every workflow can be co-piloted.

And that only works if your team understands the structure of work, not just the end goal.

The Bottom Line: Lead Through Workflow

AI success is not about the model you use. It’s about how you design the work.

Here’s what we teach in our ​Lead with AI Boot Camp​ when working with teams on adoption:

1. Map the workflow: Take real tasks and break them down into steps. Label each one by how clear, repeatable, or programmable it is. This shows you where AI fits and where it doesn’t.
2. Train for augmentation, not automation: Encourage your team to use AI tools as assistants for parts of the job. Don’t start by replacing the whole task. Start by sharing the load.
3. Delegate by step, not by role: Rethink the handoff. Let the AI agent handle the structured or code-heavy parts. Let people focus on quality, creativity, and judgment.
4. Make workflows modular: Restructure work so that it can be picked up, passed on, or recombined. This makes it easier to experiment with human-AI pairing without breaking the process.
5. Teach managers how to read workflows: Invest in training managers to understand how work flows through AI systems. Help them spot fabrication, verify quality, and create structure.

You can start by running ​a workflow-thinking workshop for your leadership team​.

The future of work won’t be fully human or fully AI. It will be workflow-aware teams powered by the right kind of support.

‍

‍

AI in Organizations Roundup 🗞️

I track how AI is reshaping organizations, bringing you the news and updates that matter most for scaling AI successfully.  

This week:

ORGANIZATIONAL AI BLOCKERS

​HBR: Why AI Fails Inside Organizations​

• People Problems Stall AI: 45% of executives say AI ROI falls short, often due to employee fear of replacement, status loss, and lack of training or incentive alignment.
• Workflow Redesign Required: AI needs to be embedded at all levels, from individual tasks to cross-functional systems. Treating it as a plug-in leads to negligible gains.
• Politics Undermine Progress: Resistance often comes from middle managers protecting headcount or data silos, with AI threatening both hierarchy and resource control.
• Trust Beats Tech: Strategies like transparent incentive systems, human-in-the-loop oversight, and AI dignity frameworks (like DBS’s PURE) help build internal confidence.
• Transformation Is Multi-Level: One firm raised productivity 22% and sales 20% by linking AI to restructured incentives, flatter hierarchy, and end-to-end process change.

🚀 Prompt: Which part of your org resists AI quietly, through delay, hoarding, or gatekeeping, and what signal you are sending about who gains power from change.

EMPLOYEE EXPERIENCE DESIGN

​Canva: Why Your Company Needs a Vibe Team​

• Team with a Mission: Canva’s Vibe team is a multidisciplinary group focused solely on crafting employee experiences, backed by its own strategy and budget, separate from facilities.
• Culture in Action: From global-local celebrations like the Canva World Tour to “sacred lunch hours” and pop-ups, the team uses rituals and joy to drive belonging and well-being.
• Human in the Loop: As AI reshapes how work gets done, Canva shows the value of designing how work feels, reminding leaders that emotional connection and community can’t be automated.

🚀 Prompt: How your team currently experiences joy and belonging at work and who’s accountable for making those moments happen, especially as AI shifts the way work is delivered.

AI BUBBLE DYNAMICS

​Gary Marcus Calls Out AI Hype​

• Unsustainable AI Economics: AI’s trillion-dollar bets rest on shaky foundations. Most companies lose money, hardware depreciates fast, and circular funding hides real risk. Only Nvidia and music licensors show consistent profits. A future “too big to fail” moment may force taxpayer bailouts.
• Psychological Cliff Edge: The industry is past the tipping point—“like the Roadrunner off the cliff.” A stumble from OpenAI or Nvidia could trigger market-wide panic, hitting pensions, banks, and possibly requiring government bailouts.
• Neuro-Symbolic Alternative: Current LLMs lack reasoning and factual grounding. A more reliable future may depend on hybrid AI that combines neural networks with symbolic logic and structured world models.

🚀 Prompt: We talk a lot about AI hype and the looming AI bubble. But what if it’s not just noise? What signs do you see, either in your organization or the market, that suggest we’re headed for a correction? And how are you preparing for it?

💨 Quick Read:

• ​GPT‑5.1 Boosts Speed, Warmth, and Control:​ OpenAI’s new GPT‑5.1 update makes ChatGPT smarter and more enjoyable to use. The Instant model is more conversational and adaptive, while the Thinking model delivers clearer answers and dynamic reasoning time. New tone presets like Quirky and Professional let users personalize response style instantly across all chats.
• ​Spatial Intelligence Is AI’s Next Leap:​ Dr. Fei-Fei Li argues that spatial intelligence, the ability to understand and interact with the physical world, is the missing link in AI’s evolution. From World Labs’ Marble platform to future robotics and scientific discovery, spatially intelligent “world models” promise breakthroughs far beyond current LLMs. This next frontier blends geometry, perception, and action to enable AI that truly sees, reasons, and builds.
• ​Claude-Powered Team Fetches Faster:​ In Anthropic’s robotics experiment, AI-assisted coders completed tasks 2× faster, wrote 9× more code, and nearly achieved full robodog autonomy, showing how Claude boosts human performance in real-world hardware challenges.

‍